AI CCTV Monitoring Platform
White-label ready
Model guide
Choose, install and use the right model
What a model is, how to install one in a few minutes, and every model that could be added, job by job, in plain words. No machine-learning knowledge needed.
Start here: what a model is
- A model is the platform's eyesA model is a trained program. It looks at each picture from a camera and says what is in it: a person, a helmet, a vehicle, smoke.
- One model does one jobFinding people is one job. Checking their PPE is another. Each job has one model running at a time, and you choose which.
- Ready-made models install with one clickYou do not train anything to begin. The platform's Model library holds ready-made models for finding people, checking PPE, spotting vehicles and more.
- Try it before you trust itThe Test bench checks your own photos and videos with a model, so you see how it does on your site before it judges a live camera.
- Every model has a licenceThe licence says whether the model may be used in a business. Each model in this guide carries that answer in plain words.
- Your site makes it betterWhen your team reviews events, the platform collects pictures to learn from. A model trained on your own cameras fits your site best.
Install a model in five steps
Open Models. It is in the sidebar under Improve. Administrators see it.
Pick a model and a size. In the Model library, find the job you need (people, or PPE) and choose a size. The one marked Recommended is a good start.
Choose Download. The model downloads and installs in the background. Its card shows the progress and then says Installed.
Choose Activate. The first time, the platform asks for a reason, because the model has not been scored on your own pictures yet. Enter one such as "first install".
Check it on the Test bench. Drop a few photos or a stored video from your site and look at what the model finds. Then let it run on your cameras.
Have a model file of your own? On the Models page, Import model takes a
.pt or .onnx file, and Add from URL downloads one from a link.Note. Downloading a ready-made model needs the internet. After that the site can run offline.
Which size to choose
Most ready-made models come in sizes. A bigger model looks more carefully and needs more computing power. Start small and go up only if the Test bench shows you need to.
- NanoChoose it whenYou have many cameras, or no graphics card. The fastest.NeedsRuns on a normal processor (CPU)
- SmallChoose it whenYou are not sure. The balance of speed and care that suits most sites.NeedsA graphics card (GPU) helps
- MediumChoose it whenPeople are far from the camera, or Small misses things.NeedsA GPU is recommended for several cameras
- LargeChoose it whenYou need more care still and have the hardware.NeedsA GPU is recommended
- XLargeChoose it whenFew cameras and the most careful result. The slowest.NeedsA GPU is required
Speed and accuracy figures for any model are its makers' own. Nothing in this guide was measured on your site: the Test bench is where you find out.
How to read a model in this guide
- Start here
- The recommended first choice for the job.
- Licence-safe alternative
- The one to take when the first choice's licence has conditions you cannot meet.
- Avoid
- Listed so that nobody picks it by accident. The reason is beside it.
- Allowed
- Commercial use: the licence lets a business use it.
- With conditions
- Commercial use depends on something, and the list says what: for example a paid licence. Ask your supplier before you rely on it.
- Not allowed
- For research and trials only. Do not use it in a business.
- Yes under Loads today
- The platform can run it now.
- Not yet
- The model exists, and your supplier must add support for it before the platform can run it.
- Not in the platform yet
- The job itself is not part of the platform today. The models are listed for planning.
- Unconfirmed
- Its details could not be confirmed from the makers' own pages.
- Good to know
- A short technical note for whoever installs the model. If it means nothing to you, skip it: the labels above are what matter.
Models the platform can load today
These 40 load in the platform now. The ones for finding people and checking PPE install from the Model library; ask your supplier about the others.
| Model | Job | Commercial use | Good to know |
|---|---|---|---|
| RF-DETR (N / S / M / L detection, plus Seg N to 2XL) | Person detection | Allowed | The person detector in the app's catalogue (N, S, M, L; sha256 checked). Objects365 provenance is unconfirmed. |
| D-FINE (N / S / M / L / X) | Person detection | Allowed | In the app's catalogue (S, M). Take the COCO-only checkpoints; the README says Objects365 ones are not cleared. |
| RT-DETR / RT-DETRv2 (Baidu / PekingU) | Person detection | Allowed | Use the COCO-only checkpoints; first-class Transformers support. |
| Ultralytics YOLO11 / YOLO26 (and YOLOv8, v5) | Person detection | With conditionsEnterprise License or AGPL terms. | Works today; AGPL applies even to models trained from scratch; no public price. |
| RF-DETR N/S/M/L (Apache sizes) fine-tuned on PPE data | PPE detection | With conditionsPretraining data terms not stated. | Licence-clean base to fine-tune on site crops plus CC BY / CC0 sets; no ready PPE weights exist. Training page: Detector, RF-DETR (learning rate 1e-4); the result is an ONNX model. |
| killuminati1/construction-ppe-yolov8 | PPE detection | With conditionsEnterprise License or AGPL terms. | Richest taxonomy among ready weights (19 classes); ear protection and boots are weak. |
| Ultralytics YOLO11 / YOLO26 trained in-house on PPE data | PPE detection | With conditionsEnterprise License. | Fastest way to a working detector today; not for a closed sale without the Enterprise licence. |
| Hexmon/vyra-yolo-ppe-detection | PPE detection | With conditionsEnterprise License or AGPL terms. | 14 classes with "No-" forms; no metrics on the card. |
| thalostech2025/dfine-small-construction-ppe-v1 | PPE detection | With conditionsFine-tune data source unknown. | Only permissive-tagged ready PPE detector with five gear classes; weak outside person, hard hat and vest. The author's own scores, out of 1, are vest 0.48 and hard hat 0.39: check it on your cameras. |
| keremberke/yolov8{n,s,m}-protective-equipment-detection | PPE detection | With conditionsEnterprise License or AGPL terms. | Already in the catalogue; low accuracy, near-silent on site frames. |
| keremberke/yolov8{n,s,m}-hard-hat-detection | PPE detection | With conditionsEnterprise License or AGPL terms. | Helmets only, already in the catalogue; reads caps as helmets on wide CCTV views. |
| Hansung-Cho/yolov8-ppe-detection | PPE detection | With conditionsEnterprise License or AGPL terms. | Already in the catalogue; the MIT tag is wrong for an ultralytics model. |
| mancalazure/construction-ppe-safety | PPE detection | With conditionsEnterprise License or AGPL terms. | Ready ONNX in the stock YOLOv8 layout; demo-grade and unmeasured. |
| MobileSAM and EfficientSAM | Open-vocabulary detection | Allowed | Cheap box-to-mask on CPU; upstream LICENSE not fetched; EfficientSAM not checked. |
| YOLOE (THU-MIG YOLOE and Ultralytics YOLOE-11 / YOLOE-26) | Open-vocabulary detection | Not allowed | Fastest text-prompt detector, for labelling only; bake prompts offline and do not ship it. |
| RTMPose (MMPose) via rtmlib | Pose | With conditionsTraining-set terms not read. | Per-person keypoints on crops; the paired RTMDet detector is trained on Objects365. |
| YOLO-pose (Ultralytics YOLO11 / YOLO26 pose) | Pose | With conditionsEnterprise License or AGPL terms. | Only pose model that loads today, for development. |
| HF community YOLO face models (AdamCodd/YOLOv11n-face-detection, Bingsu/adetailer face_yolov8*, arnabdhar/YOLOv8-Face-Detection) | Face detection | With conditionsEnterprise License or AGPL terms. | Only face detector that loads today, for internal tests; an Apache tag does not remove the AGPL. |
| Ultralytics YOLO11 / YOLO26 COCO weights, class 67 "cell phone" | Phone use | With conditionsEnterprise License or AGPL terms. | In-house spike: no false alarms on 72 negatives, about half the phones found at about 220 px person height. |
| Beehzod/smoking-detection-finetuned | Smoking | With conditionsEnterprise License or AGPL terms. | In the model library as "Cigarette — YOLO11m (experimental)", pinned to commit d81c4396. The smoking check runs it on person crops at conf 0.5, and a phone or food box overlapping the hit vetoes it. Finds about half of site-scale cigarettes. |
| Pose rule (hand-to-mouth) on RTMPose keypoints plus a small own-trained cigarette/vape detector | Smoking | With conditionsBody7 training-set terms not read. | Licence-clean way to build smoking; needs your own labelled clips. |
| Beehzod/smoke_cigarette-detection2-yolo11m and best_smoke_cigarette-detection | Smoking | With conditionsEnterprise License or AGPL terms. | Second cigarette detector to compare; a 0.994 mAP50 on near-duplicate frames says little. |
| rabahdev/fire-smoke-yolov8n | Fire and smoke | With conditionsEnterprise License or AGPL terms. | Works today; use conf 0.5 or more for fire. |
| xiazh0219/yolo26m-UAV-fire-smoke-detection | Fire and smoke | With conditionsEnterprise License or AGPL terms. | Aerial and outdoor smoke; not for ceiling or wall cameras. |
| Pyronear YOLO (yolov8s, yolov11s/n, yolo11s_*) | Fire and smoke | With conditionsEnterprise License or AGPL terms. | Early smoke at long range from fixed towers; indoor steam and dust will trigger it. |
| RTMPose (rtmlib / MMPose) plus fall rules | Falls | With conditionsBody7 training-set terms not read. | Loads now; the person-down pose rule is built on it. |
| Ultralytics YOLO11-pose | Falls | With conditionsEnterprise License or AGPL terms. | Lets you try the fall rules before the Apache route exists. |
| Ultralytics YOLO26-pose | Falls | With conditionsEnterprise License or AGPL terms. | In the model library as "Pose — YOLO26 (official)" (weights from ultralytics assets v8.4.0). Feeds the Person down check, method pose; s is the default. |
| melihuzunoglu/human-fall-detection | Falls | With conditionsEnterprise License or AGPL terms. | Quick fall-by-posture detector; cannot tell lying down from a fall without a time rule. |
| mustafahacioglu/yolo11s-forklift-person-detection | Forklift | With conditionsEnterprise License or AGPL terms. | Works today; its card adds a box-containment rule for operator presence; the map_up licence is unread. |
| yihong1120/Construction-Hazard-Detection | Forklift | With conditionsEnterprise License or AGPL terms. | Works today as an alternative forklift source: it reports forklifts as "machinery", and the catalogue ignores its "vehicle" class because it false-alarms on site. |
| keremberke/yolov8{n,s,m}-forklift-detection (and yolov5 versions) | Forklift | With conditionsEnterprise License or AGPL terms. | Tiny model for a first demo; a Roboflow-derived forklift model once boxed the whole frame. |
| RT-DETR and RT-DETRv2 | Vehicles | Allowed | Easiest DETR to prototype with in Transformers; avoid the AGPL ultralytics port. |
| Ultralytics YOLO11 and YOLO26 | Vehicles | With conditionsUltralytics Enterprise licence. | Only family that loads today; AGPL applies even to internal use or training from scratch. |
| PaddleOCR PP-OCRv5 recognisers (default, latin, arabic, devanagari, cyrillic, ta, te, th, korean, el) | Plate reading | Allowed | Only permissive route to non-Latin scripts; not tuned for plates, so fine-tune on cropped plates. |
| OpenCV classical checks (Laplacian variance, histograms, edge density, ORB or phase-correlation shift, SSIM) | Camera health | Allowed | Blur, dark, blocked, moved and frozen frames. Tune thresholds per camera and for day and night. |
| OpenCV image checks (Laplacian blur, brightness, SSIM, frozen frame) | Camera tamper and picture quality | Allowed | Near-zero cost on every camera every minute. Thresholds per camera. |
| OpenCV ORB or AKAZE matching with a homography | Camera tamper and picture quality | Allowed | Detects that the camera moved and by how much, so zones can be marked invalid. Needs background features. |
| SAM 2.1 Hiera Tiny (zone outline) | Zone outlines and depth hints | Allowed | No pinned ONNX file exists, so the install downloads the commit-pinned checkpoint, checks its SHA-256, and exports the encoder and decoder on the site with the trainer's torch. Small, base+ and large are not offered: too slow on a processor. |
| Depth Anything V2 Small (depth hint) | Zone outlines and depth hints | Allowed | One commit-pinned ONNX file from the onnx-community mirror, checked by SHA-256. Relative depth only, never metres. |
Every model, by job
664 models for 63 jobs, in eight groups. Choose a group, or search. In the app, the same list is on Models → All models, where you can also shortlist one model per job.
No models match. Try fewer words, or clear the filter.
People and PPE
Person detection31 models · 4 load todayStart hereRF-DETR (N / S / M / L detection, plus Seg N to 2XL)Allowed
What it does: Finds every person in the picture, so each one can be followed and checked.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| RF-DETR (N / S / M / L detection, plus Seg N to 2XL)Start here | Allowed | Yes | The person detector in the app's catalogue (N, S, M, L; sha256 checked). Objects365 provenance is unconfirmed. |
| D-FINE (N / S / M / L / X)Licence-safe alternative | Allowed | Yes | In the app's catalogue (S, M). Take the COCO-only checkpoints; the README says Objects365 ones are not cleared. |
| RT-DETR / RT-DETRv2 (Baidu / PekingU) | Allowed | Yes | Use the COCO-only checkpoints; first-class Transformers support. |
| YOLOX (Megvii) | Allowed | Not yetNeeds a loader first. | Cheap CPU fallback; its ONNX head has an objectness score, not the ultralytics layout. |
| Ultralytics YOLO11 / YOLO26 (and YOLOv8, v5) | With conditionsEnterprise License or AGPL terms. | Yes | Works today; AGPL applies even to models trained from scratch; no public price. |
| RT-DETRv4 | With conditionsClear the DINOv3 teacher licence first. | Not yetNeeds a loader first. | Whether a student of a DINOv3-licensed teacher is a derivative work is unsettled. |
| RT-DETRv3 (Baidu, PaddlePaddle)Unconfirmed | With conditionsWeights licence not stated. | Not yetNeeds a loader first. | Gives nothing the PyTorch RT-DETR family does not; Paddle dependency. |
| LW-DETR | Allowed | Not yetNeeds a loader first. | Ancestor of RF-DETR; take RF-DETR instead. |
| DEIM (original, D-FINE lineage) | Allowed | Not yetNeeds a loader first. | The same lab's DEIMv2 is non-commercial; check the exact repo and tag. |
| RTMDet (MMDetection)Unconfirmed | Allowed | Not yetNeeds a loader first. | Weights licence is inferred from the repo; needs MMDeploy. |
| PaddleDetection PP-YOLOE+ / PicoDetUnconfirmed | With conditionsWeights licence not stated. | Not yetNeeds a loader first. | PicoDet suits very small edge boxes; Paddle toolchain. |
| YOLOv9 MIT re-implementation (MultimediaTechLab/YOLO) | Allowed | Not yetNeeds a loader first. | The one YOLO-style CNN under MIT; not ultralytics-compatible; check that it shares no GPL code. |
| LibreYOLO (wrapper library)Unconfirmed | With conditionsCheck each wrapped weight's licence. | Not yetNeeds a loader first. | One API over several detectors; its open-vocabulary DEIM weights are CC-BY-NC-4.0. |
| DAMO-YOLO (Alibaba) | With conditionsWeights licence not stated. | Not yetNeeds a loader first. | The OSS weights URL tried returned 404; last release April 2023. |
| NVIDIA TAO PeopleNet / PeopleNet-TransformerUnconfirmed | With conditionsNVIDIA licence text not read. | Not yetNeeds a loader first. | Tuned for overhead and CCTV views; NVIDIA GPUs only. |
| torchvision Faster R-CNN / RetinaNet / SSDUnconfirmed | With conditionsWeights licence not stated. | Not yetNeeds a loader first. | A stable baseline to compare against; older and slower than DETR and YOLO. |
| Intel Open Model Zoo person-detection-0200 to 0203Unconfirmed | With conditionsWeights licence inferred, data undisclosed. | Not yetNeeds a loader first. | Tiny CPU-only detector for Intel hardware; the zoo is archived. |
| DETR and Deformable / Conditional DETR (original) | Allowed | Not yetNeeds a loader first. | Reference baseline only; slow to converge. |
| Co-DETR, DINO and detrex models (large DETRs)Unconfirmed | With conditionsWeights licences not confirmed. | Not yetNeeds a loader first. | Offline teacher or auto-labeller only; far too heavy for live streams. |
| MediaPipe object detector (EfficientDet-Lite, SSD MobileNet)Unconfirmed | With conditionsWeights terms not read. | Not yetNeeds a loader first. | For phones and tiny CPUs; low accuracy. |
| dfine-person-s | With conditions0.633. | Not yet238.0. | |
| rf-detr-person-s | With conditions0.863. | Not yet300.3. | |
| rf-detr-person-n | With conditions0.856. | Not yet214.2. | |
| D-FINE-S (yardstick) | With conditions0.633. | Not yet201.6. | |
| D-FINE-N (`dfine_n_coco`) | With conditions0.317. | Not yet84.1. | |
| RT-DETRv2-R18 (`rtdetr_v2_r18vd`) | With conditions0.659. | Not yet321.3. |
Avoid 5
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| RF-DETR Plus (Atto / Femto / Pico / XL / 2XL)Avoid | Not rated | Not rated | PML 1.0, With conditions, and only worth it with a Roboflow plan; the PML text was not found. |
| DEIMv2 and EdgeCrafterAvoid | Not rated | Not rated | Non-commercial licence; was Apache-2.0 until 2026-08-24. |
| YOLOv10, YOLOv12, YOLOv13AvoidUnconfirmed | Not rated | Not rated | AGPL-3.0 with no enterprise route. |
| YOLOv9 (WongKinYiu), YOLOv7, YOLOv6, Gold-YOLO, MMYOLOAvoid | Not rated | Not rated | GPL-3.0 copyleft. |
| YOLO-NAS (Deci / NVIDIA)Avoid | Not rated | Not rated | The weights licence bars commercial use. |
PPE detection19 models · 9 load todayStart hereRF-DETR N/S/M/L (Apache sizes) fine-tuned on PPE dataWith conditions
What it does: Checks each person for helmet, vest, gloves, goggles, mask and boots.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| RF-DETR N/S/M/L (Apache sizes) fine-tuned on PPE dataStart here | With conditionsPretraining data terms not stated. | Yes | Licence-clean base to fine-tune on site crops plus CC BY / CC0 sets; no ready PPE weights exist. Training page: Detector, RF-DETR (learning rate 1e-4); the result is an ONNX model. |
| D-FINE / DEIM / RT-DETRv2 (Apache) as fine-tune basesLicence-safe alternative | Allowed | Not yetNeeds a loader first. | Second route to a clean PPE model; use COCO-only checkpoints. |
| killuminati1/construction-ppe-yolov8 | With conditionsEnterprise License or AGPL terms. | Yes | Richest taxonomy among ready weights (19 classes); ear protection and boots are weak. |
| Ultralytics YOLO11 / YOLO26 trained in-house on PPE data | With conditionsEnterprise License. | Yes | Fastest way to a working detector today; not for a closed sale without the Enterprise licence. |
| Hexmon/vyra-yolo-ppe-detection | With conditionsEnterprise License or AGPL terms. | Yes | 14 classes with "No-" forms; no metrics on the card. |
| thalostech2025/dfine-small-construction-ppe-v1 | With conditionsFine-tune data source unknown. | Yes | Only permissive-tagged ready PPE detector with five gear classes; weak outside person, hard hat and vest. The author's own scores, out of 1, are vest 0.48 and hard hat 0.39: check it on your cameras. |
| keremberke/yolov8{n,s,m}-protective-equipment-detection | With conditionsEnterprise License or AGPL terms. | Yes | Already in the catalogue; low accuracy, near-silent on site frames. |
| keremberke/yolov8{n,s,m}-hard-hat-detection | With conditionsEnterprise License or AGPL terms. | Yes | Helmets only, already in the catalogue; reads caps as helmets on wide CCTV views. |
| Hansung-Cho/yolov8-ppe-detection | With conditionsEnterprise License or AGPL terms. | Yes | Already in the catalogue; the MIT tag is wrong for an ultralytics model. |
| mancalazure/construction-ppe-safety | With conditionsEnterprise License or AGPL terms. | Yes | Ready ONNX in the stock YOLOv8 layout; demo-grade and unmeasured. |
| KuBaN1900/rtdetr-v2-r50vd-hardhat | With conditionsTraining set has no licence tag. | Not yetNeeds a loader first. | Helmet versus bare head only; two classes. |
| Qualcomm PPE-Detection (gear_guard_net) | Allowed | Not yetNeeds a loader first. | Tiny detector for Qualcomm-class edge boxes; the source reportedly describes a person/face detector. |
| PP-YOLOE+ (PaddleDetection)Unconfirmed | With conditionsWeights terms not checked. | Not yetNeeds a loader first. | Apache code route if the Paddle toolchain is acceptable; needs its own loader. |
Avoid 6
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| SH17 YOLOv8 / v9 / v10 weightsAvoid | Not rated | Not rated | Trained on CC BY-NC-SA 4.0 data and published for evaluation only. |
| leeyunjai/yolo11-ppeAvoid | Not rated | Not rated | No licence tag, no data or metrics, card copied from another project. |
| YOLO-NAS PPE fine-tunes (SuperGradients)Avoid | Not rated | Not rated | The Deci weights licence forbids distribution and any fine-tune inherits it. |
| Community uploads with a PPE name but no card or files (for example VirnectUnityTeam/rfdetr-ppe-detection, gtensorapp/worksite-ppe-*)Avoid | Not rated | Not rated | An apache-2.0 tag on an empty repo tells you nothing. |
| YOLOv9 (WongKinYiu) and YOLOv7AvoidUnconfirmed | Not rated | Not rated | GPL-3.0 by the GitHub record, no easier than ultralytics AGPL. |
| YOLOv10 (THU-MIG)AvoidUnconfirmed | Not rated | Not rated | AGPL-3.0 with no enterprise route. |
PPE datasets13 modelsStart hereHard Hat Workers (andrewmvd, Kaggle) and Voxel51 mirrorAllowed
What it does: Labelled pictures for teaching a PPE model your own site. Data, not a model.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| Hard Hat Workers (andrewmvd, Kaggle) and Voxel51 mirrorStart here | Allowed | Data, not a model | Cleanest licence for helmet and bare-head examples; close-up construction images. |
| Open Images V7 (Helmet, Goggles, Glove, Boot, Hat)Licence-safe alternativeUnconfirmed | Allowed | Data, not a model | Hard negatives and helmets from varied scenes; Google does not vouch for each image. |
| Roboflow "Hard Hats" via keremberke/hard-hat-detection | Allowed | Data, not a model | Largest permissive hard-hat set; the licence is the Roboflow project's own claim. |
| Roboflow "ppes-kaxsi" via keremberke/protective-equipment-detection | Allowed | Data, not a model | Gear plus explicit negatives; labels are noisy. |
| Roboflow "PPE Combined Model" (roboflow-universe-projects)Unconfirmed | With conditionsLicence only confirmed second-hand. | Data, not a model | Broad class set with negatives; check the Universe page before use. |
| Roboflow "construction-site-safety" via keremberke/construction-safety-object-detection | Allowed | Data, not a model | Tiny set that mixes PPE and site vehicles; useful for tests. |
| Construction-PPE (Ultralytics) | With conditionsUltralytics Enterprise licence. | Data, not a model | Good negative-class design to imitate; Ultralytics publishes no weights for it. |
| SHWD (Safety Helmet Wearing Dataset)Unconfirmed | With conditionsPicture copyright not stated. | Data, not a model | Many bare-head negatives from CCTV-like views. |
| CPPE-5 (medical PPE) | With conditionsLicence unknown. | Data, not a model | Only medical PPE (hospital preset). |
| Hairnet and kitchen PPE sets (Roboflow Universe, IEEE DataPort)Unconfirmed | With conditionsLicences not re-verified. | Data, not a model | Hygiene preset (hairnet, mask, gloves, apron), where nothing else exists. |
Avoid 3
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| SH17Avoid | Not rated | Not rated | CC BY-NC-SA 4.0 plus Pexels licence, research only. Copy its class list, not its images. |
| CHV (ZijianWang-ZW/PPE_detection) and Pictor-PPE (ciber-lab/pictor-ppe)Avoid | Not rated | Not rated | No licence file; the GitHub record shows license null for pictor-ppe. |
| Hugging Face-uploaded PPE sets with a licence tag (for example VincentGOURBIN/ppe-detection, HB1204/PPE_Detection)AvoidUnconfirmed | Not rated | Not rated | The card holds only a cc-by-4.0 tag and no source. |
Open-vocabulary detection19 models · 2 load todayStart hereGrounding DINO (open Swin-T and Swin-B)Allowed
What it does: Finds things you name in plain words with no training, for labelling and rare items.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| Grounding DINO (open Swin-T and Swin-B)Start here | Allowed | Not yetNeeds a loader first. | Standard permissive text-prompt labeller; slow, use offline. |
| OmDet-TurboStart here | Allowed | Not yetNeeds a loader first. | Fast enough for near frame rate; quality on PPE untested. |
| MM-Grounding-DINO (tiny / base / large)Licence-safe alternative | Allowed | Not yetNeeds a loader first. | Open re-implementation with Transformers support; PPE quality untested. |
| OWLv2 (and OWL-ViT v1)Licence-safe alternative | Allowed | Not yetNeeds a loader first. | Text or image-example queries; heavy ViT, not real-time. |
| LLMDet | Allowed | Not yetNeeds a loader first. | Newer labeller; training data licence not confirmed. |
| Florence-2 (base / large) | Allowed | Not yetNeeds its own integration. | Grounding, captioning and OCR from one small model; boxes less precise than a detector. |
| SAM 3 / 3.1 (text and exemplar concept prompts) | With conditionsMeta may change the licence; ITAR and weapons-use bar. | Not yetNeeds its own integration. | Best-quality labeller for masks and boxes; keep offline and out of weapon-detection workflows. |
| SAM 2 / 2.1 | Allowed | Not yetNeeds its own integration. | Turns a box into a mask or tracks a mask through video; no text prompts. |
| Grounded-SAM / Grounded-SAM-2Unconfirmed | Allowed | Not yetNeeds its own integration. | Ready text to box to mask to track pipeline; glue-repo licence not re-read. |
| Roboflow AutodistillUnconfirmed | Allowed | Not yetNeeds its own integration. | The pattern: label with a foundation model, ship only the student; its YOLO-World and YOLOE plug-ins are copyleft. |
| MobileSAM and EfficientSAMUnconfirmed | Allowed | Yes | Cheap box-to-mask on CPU; upstream LICENSE not fetched; EfficientSAM not checked. |
| YOLOE (THU-MIG YOLOE and Ultralytics YOLOE-11 / YOLOE-26) | Not allowed | Yes | Fastest text-prompt detector, for labelling only; bake prompts offline and do not ship it. |
| Detic (Meta)Unconfirmed | With conditionsWeights licence not confirmed. | Not yetNeeds a loader first. | Historic baseline, superseded by Grounding DINO and OWLv2. |
| GLIP (Microsoft)Unconfirmed | With conditionsWeights licence not checked. | Not yetNeeds a loader first. | Historic baseline, superseded by Grounding DINO. |
Avoid 5
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| YOLO-World (Tencent; Ultralytics port yolov8-world)Avoid | Not rated | Not rated | GPL-3.0 (Tencent) and AGPL-3.0 (port); a commercial licence on request is from the survey only. |
| Grounding DINO 1.5 / 1.6, DINO-X (IDEA Research API)AvoidUnconfirmed | Not rated | Not rated | Cloud API only, no weights; it would also send site frames off the site. |
| Rex-Omni and T-Rex / T-Rex2 (IDEA Research)Avoid | Not rated | Not rated | IDEA License 1.0, non-commercial research. |
| OV-DEIM / LibreOVDEIMAvoid | Not rated | Not rated | CC-BY-NC-4.0 weights. |
| FastSAMAvoid | Not rated | Not rated | AGPL-3.0, looks like a cheap segmenter but is copyleft. |
Pose16 models · 2 load todayStart hereRTMPose (MMPose) via rtmlibWith conditions
What it does: Finds the body's joints, the basis for falls, posture and reach.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| RTMPose (MMPose) via rtmlibStart here | With conditionsTraining-set terms not read. | Yes | Per-person keypoints on crops; the paired RTMDet detector is trained on Objects365. |
| RF-DETR keypoints (preview)Licence-safe alternative | Allowed | Not yetNeeds a loader first. | Detector and pose in one pass once the DETR decoder exists; it is a preview. |
| RTMO (one-stage multi-person) | With conditionsBody7 training-set terms not read. | Not yetNeeds a loader first. | For crowded scenes; heavier than RTMPose crops for few people. |
| ViTPose / ViTPose+ | With conditionsAIC terms unread. | Not yetNeeds a loader first. | Highest accuracy when speed is secondary; per-person crops. |
| LibreYOLO pose checkpoints (YOLO9-s-pose, RF-DETR n/s/m/l/x-pose) | With conditionsYOLO9 variants derive from YOLOv9, believed GPL-3.0. | Not yetNeeds a loader first. | The cards call them "EXTREMELY experimental". |
| DWPose / RTMW whole-body | With conditionsRTMW cocktail14 data terms not read. | Not yetNeeds a loader first. | Adds hand and face keypoints. |
| MediaPipe Pose / BlazePose LandmarkerUnconfirmed | With conditionsWeights licence and data unconfirmed. | Not yetNeeds a loader first. | One close subject, 33 landmarks; single-person. |
| Skeleton action models for falls: PoseC3D, ST-GCN (MMAction2, PYSKL) | With conditionsStock weights carry NTU terms. | Not yetNeeds a loader first. | Learn falls from pose sequences if a site has labelled falls; most teams use a rule instead. |
| HRNet, Intel OMZ human-pose-estimation, PP-TinyPose, Lightweight OpenPose | With conditionsNo weights licence stated. | Not yetNeeds a loader first. | References and edge targets; older and less accurate than RTMPose. |
| YOLO-pose (Ultralytics YOLO11 / YOLO26 pose) | With conditionsEnterprise License or AGPL terms. | Yes | Only pose model that loads today, for development. |
| MoveNet (Google; Qualcomm AI Hub port)Unconfirmed | With conditionsRead through a third-party port; data cannot be audited. | Not yetNeeds a loader first. | Single-person, very fast keypoints on edge devices. |
Avoid 5
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| EdgeCrafter ECPose (and the LibreEC-pose mirrors)Avoid | Not rated | Not rated | EdgeCrafter License, non-commercial; the mirror cards' Apache-2.0 tag contradicts upstream. |
| OpenPose and AlphaPoseAvoid | Not rated | Not rated | Academic non-profit licence for OpenPose; AlphaPose non-commercial per the survey only. |
| Sapiens (Meta)Avoid | Not rated | Not rated | CC BY-NC 4.0. |
| YOLO-NAS-pose (Deci) and ED-Pose (IDEA)Avoid | Not rated | Not rated | The YOLO-NAS licence bars commercial use; ED-Pose is under the non-commercial IDEA License. |
| openpifpafAvoid | Not rated | Not rated | AGPL-3.0 with a commercial licence on request from EPFL-TTO, plus a "may not be used to harm any person deliberately" line. |
Crowd density and counting10 models · Not in the platform yetStart hereRF-DETR person detection with tilingAllowed
What it does: Counts and density maps where a detector merges bodies. Needs: Elevated view, validation against manual counts. Main risk: Large published errors; research-only datasets.
| Model | Commercial use | Good to know |
|---|---|---|
| RF-DETR person detection with tilingStart here | Allowed | Counts up to a few dozen visible people with exact boxes. Yes applies to N to L only.Would run on: Consumer GPU |
| CLIP-EBC | With conditionsWeights built on research datasets; fine-tune on site frames. | Best open density model. NWPU-Crowd MAE 61.2 (ViT-B/16), 75.8 (ResNet-50).Would run on: Consumer GPU |
| DM-Count | With conditionsWeights terms not stated. | Simple baseline that is easy to retrain on site data.Would run on: Consumer GPU |
| PET (point-query transformer) | With conditionsLicence template unfilled; ask the authors. | Gives crowd points, not just a count.Would run on: Consumer GPU |
| APGCC | With conditionsWeights and dataset terms unread. | Point-based localisation, stable under noise.Would run on: Consumer GPU |
| STEERER | With conditionsLicence text still names another owner. | Handles large size change within one view.Would run on: Consumer GPU |
| CountGD, DAVE, LOCA, GeCo2 and CounTR (class-agnostic counters)Unconfirmed | With conditionsWeights terms differ; CountGD builds on Grounding DINO. | Counts any item from examples or text: boxes on a pallet, bags, tyres. Not for dense crowds.Would run on: Consumer GPU to large GPU |
| Open crowd datasets (ShanghaiTech, UCF-QNRF, NWPU-Crowd, JHU-Crowd++)Unconfirmed | With conditionsTerms unread; research-only is usual. | Benchmarking only. Label site frames instead.Would run on: n/a |
Avoid 2
| Model | Commercial use | Good to know |
|---|---|---|
| P2PNet (Tencent)Avoid | Not rated | Academic research only. |
| CSRNet (leeyeehoo port)Avoid | Not rated | No licence file, so no right to use. |
Hands and gestures5 models · Not in the platform yetStart hereRF-DETR or D-FINE fine-tuned on "hand" and the hazard, plus zone geometryAllowed
What it does: Warns when a hand enters a hazard zone; spots a help signal. Needs: Close cameras, drawn zones, machine-state signal. Main risk: Not a safeguard; gloves and small hands.
| Model | Commercial use | Good to know |
|---|---|---|
| RF-DETR or D-FINE fine-tuned on "hand" and the hazard, plus zone geometryStart hereUnconfirmed | Allowed | Simplest and most robust with gloves. Needs a few hundred labelled gloved hands at real distance. Advisory only.Would run on: Consumer GPU |
| MediaPipe Hand Landmarker and Gesture RecognizerUnconfirmed | With conditionsWeights licence unconfirmed. | 21 landmarks and a built-in open-palm or fist classifier. Fails on small hands in wide views.Would run on: CPU |
| RTMPose hand and RTMW / DWPose whole-bodyUnconfirmed | With conditionsHand and whole-body training-set terms not read. | Hand points inside a whole-body pass. Only as good as the pixels available.Would run on: Consumer GPU |
| HaGRID gesture dataset (and models trained on it) | With conditionsShareAlike may bind derived models; read the PDF closely. | Starting images for a gesture classifier. Indoor, near-camera photos.Would run on: n/a |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| HaMeR and other MANO-based hand mesh modelsAvoidUnconfirmed | Not rated | Need the MANO model, which is for non-commercial research. |
Ergonomic risk from pose5 models · Not in the platform yetStart hereRTMPose / RTMW keypoints plus REBA or RULA angle rulesWith conditions
What it does: Station or task trend of REBA or RULA scores. Needs: Side camera, tall enough person, ergonomist thresholds. Main risk: 2D angle error; health inference on named workers.
| Model | Commercial use | Good to know |
|---|---|---|
| RTMPose / RTMW keypoints plus REBA or RULA angle rulesStart hereUnconfirmed | With conditionsPose training-set terms not read. | Joint angles from a side view. An ergonomist must review the REBA table.Would run on: Consumer GPU, or CPU for few people |
| RF-DETR keypoints (preview) plus angle rulesUnconfirmed | With conditionsKeypoint weights licence not read from a primary source. | Permissive person detection and 17 keypoints in one pass. Preview; fewer foot and hand points.Would run on: Consumer GPU |
| MotionBERT (2D to 3D lifting) | With conditionsWeights terms unstated and data is academic; treat as no until authors answer. | 3D joints for better trunk and neck angles. Needs 243-frame clips; mocap data does not look like a plant.Would run on: Consumer GPU |
| MediaPipe Pose Landmarker (BlazePose)Unconfirmed | With conditionsWeights licence not found. | One close subject with 33 landmarks and a 3D estimate. Single person per crop.Would run on: CPU |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| MeTRAbs, VideoPose3D and SMPL-based body models (4D-Humans, WHAM)Avoid | Not rated | Non-commercial weights and academic training data; the best-known 3D pose models cannot be sold here. |
Action and behaviour recognition from video9 models · Not in the platform yetStart hereMMAction2 skeleton models (PoseC3D, ST-GCN, CTR-GCN) trained on site clipsAllowed
What it does: Spots a person down, climbing, a fight or entry into a danger zone. Needs: Pose stage, GPU, site-labelled clips, review loop. Main risk: False alarms; not a safety function; no ranking of workers.
| Model | Commercial use | Good to know |
|---|---|---|
| MMAction2 skeleton models (PoseC3D, ST-GCN, CTR-GCN) trained on site clipsStart here | Allowed | Person down, climbing, running, crouching over 1 to 3 seconds. Viewpoint-specific, which suits fixed CCTV.Would run on: CPU for the classifier; consumer GPU for the pose stage |
| V-JEPA 2 / 2.1 (frozen backbone plus a small probe) | With conditionsPretraining data mix has research-oriented terms. | Strongest open video features; a small head on 50 to 200 clips per action. Benchmarks are gym-style, not CCTV.Would run on: Consumer GPU (ViT-L); large GPU (ViT-g) |
| X3D and SlowFast (PyTorchVideo)Unconfirmed | With conditionsWeights licence silent; Kinetics provenance. | Small classifiers to fine-tune when pose is not enough (carrying, ladder use). Older than V-JEPA 2.Would run on: CPU or small GPU (X3D-XS, S); consumer GPU (SlowFast) |
| ViViT (google/vivit-b-16x2-kinetics400) | With conditionsKinetics provenance. | Plainly MIT-tagged video transformer; slow, no reason to prefer it over V-JEPA 2.Would run on: Consumer GPU |
| X-CLIP (microsoft/xclip-*) zero-shot | With conditionsKinetics training data. | Name an action in English and score a clip with no training; for prototypes. No Sinhala or Tamil prompts.Would run on: Consumer GPU |
| Qwen3-VL (2B, 8B) asked about a short clip | Allowed | Second opinion on a clip an alert already flagged, and alert text. Not alarm-grade; assume English.Would run on: Consumer GPU |
| SmolVLM2 (256M, 500M, 2.2B) | Allowed | Smallest video-aware model; weak on safety detail. Scraped video terms unread; record the risk decision.Would run on: CPU or small GPU |
Avoid 2
| Model | Commercial use | Good to know |
|---|---|---|
| VideoMAE, VideoMAE V2 and TimeSformer (stock weights)Avoid | Not rated | CC BY-NC 4.0 weights. VideoMAE V2 code is MIT but its weights are not. |
| Community violence, fight and fall detectors (for example HappyGook/videomae-violence-detector)Avoid | Not rated | Card tag MIT, but fine-tuned from a CC BY-NC base on UCF-Crime. |
Tracking and identity
Tracking18 modelsStart heretrackers (Roboflow): SORT, ByteTrack, OC-SORT, BoT-SORT, C-BIoU, McByteAllowed
What it does: Follows each person from picture to picture, so one person is one track and one event.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| trackers (Roboflow): SORT, ByteTrack, OC-SORT, BoT-SORT, C-BIoU, McByteStart here | Allowed | Not yetNeeds its own integration. | Default tracker library for a closed-source product; no ReID branches; McByte needs SAM and Cutie. |
| norfair (Tryolabs)Licence-safe alternative | Allowed | Not yetNeeds its own integration. | Simple distance-based tracking; last release 2025-04-30. |
| BoT-SORT (NirAharon) | Allowed | Not yetNeeds its own integration. | Best for cameras that move or shake; use the trackers port. |
| supervision ByteTrack (Roboflow) | Allowed | Not yetNeeds its own integration. | Ready ByteTrack plus zones, lines and counters; overlaps with trackers. |
| ByteTrack (ifzhang original) | Allowed | Not yetNeeds its own integration. | Reference code; use the maintained trackers package in production. |
| OC-SORT (noahcao) | Allowed | Not yetNeeds its own integration. | Better recovery after occlusion; also inside trackers. |
| Deep OC-SORT | Allowed | Not yetNeeds its own integration. | Adds appearance for crowds; research code. |
| BoostTrack | Allowed | Not yetNeeds its own integration. | For experiments with a newer association scheme; small project. |
| TrackTrack | Allowed | Not yetNeeds its own integration. | Benchmark comparison on your own footage; scores not checked. |
| deep_sort_realtime | Allowed | Not yetNeeds its own integration. | DeepSORT-style tracker with a permissive licence; an old method. |
| CAMELTrack | With conditionsWeights licence and data terms not checked. | Not yetNeeds its own integration. | Learned fusion of motion and appearance; needs a ReID model. |
| MOTRv2 / MOTIP (end-to-end transformer trackers) | With conditionsWeights licence unknown. | Not yetNeeds its own integration. | Research only; heavy, one model per dataset. |
| SAM 2 video object tracking (SAMURAI builds on it) | Allowed | Not yetNeeds its own integration. | Tracks one or a few chosen objects, not every person; far heavier than box trackers. |
| DeepStream trackers (NvDCF, NvSORT, NvDeepSORT, MV3DT)Unconfirmed | With conditionsNVIDIA EULA terms not read. | Not yetNeeds its own integration. | Only if the product moves to DeepStream on NVIDIA GPUs. |
Avoid 4
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| SORT (abewley, original) and DeepSORT (nwojke)Avoid | Not rated | Not rated | GPL-3.0; the DeepSORT embedder was trained on MARS. Use SORT from the trackers package. |
| StrongSORT (dyhBUPT)Avoid | Not rated | Not rated | GPL-3.0, bundled in many YOLO-tracking repos. |
| BoxMOTAvoid | Not rated | Not rated | AGPL-3.0; fine for prototyping, never shipped. |
| Ultralytics built-in trackers (ByteTrack, BoT-SORT)Avoid | Not rated | Not rated | AGPL-3.0, covered only by the Ultralytics commercial licence. |
Re-identification11 modelsStart hereOSNet architecture (torchreid), trained on your own trackletsAllowed
What it does: Recognises the same person again on another camera or later the same day.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| OSNet architecture (torchreid), trained on your own trackletsStart here | Allowed | Not yetNeeds its own integration. | Start from random or ImageNet initialisation, not the stock Market-1501 checkpoints, or the row becomes With conditions. |
| DINOv2 embeddings as a re-ID baseline | Allowed | Not yetNeeds its own integration. | Zero-effort appearance embedding to test against OSNet; not trained for clothing-based identity. |
| OSNet / OSNet-AIN stock zoo weights | With conditionsTrained on research-only datasets. | Not yetNeeds its own integration. | For quick experiments only; weights are on Google Drive, not a stable URL. |
| FastReID (JD AI) | With conditionsStock weights carry research-only dataset terms. | Not yetNeeds its own integration. | Complete training pipeline for your own weights; last push about 2024-07. |
| TransReID (ViT re-ID) | With conditionsStock weights carry dataset terms. | Not yetNeeds its own integration. | Higher accuracy than OSNet when GPU time is not tight; heavy for many cameras. |
| CLIP-ReID and SOLIDER-REID | With conditionsWeights licences and LUPerson terms not read. | Not yetNeeds its own integration. | Benchmark leaders if you train your own on clean data. |
| Intel Open Model Zoo person-reidentification-retail-0277 / 0286 / 0287 / 0288 | With conditionsTraining data undisclosed. | Not yetNeeds its own integration. | Tiny CPU embedding for a quick demo; accuracy below OSNet. |
| NVIDIA TAO ReIdentificationNetUnconfirmed | With conditionsNGC page could not be read. | Not yetNeeds its own integration. | Only on the NVIDIA / DeepStream path. |
| DINOv3 and SigLIP 2 as embedding baselines | With conditionsDINOv3 needs approval and has conditions. | Not yetNeeds its own integration. | Neither is trained for person identity; SigLIP 2 is the clean one. |
Avoid 2
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| BPBreID (body-part re-ID)Avoid | Not rated | Not rated | Hippocratic License 3.0, an ethical-source licence with use restrictions a closed product cannot easily meet. |
| Community Hugging Face re-ID checkpoints (for example PersonViT MSMT17, OSNet MSMT17 copies)Avoid | Not rated | Not rated | The uploader's tag cannot override the MSMT17 terms; the same weights are tagged MIT by one uploader and GPL-3.0 by another. |
Face detection15 models · 1 load todayStart hereYuNet (OpenCV Zoo)With conditions
What it does: Finds faces in the picture, the first step before any face is matched.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| YuNet (OpenCV Zoo)Start here | With conditionsTrained on WIDER FACE. | Not yetNeeds a loader first. | Small, fast default with 5 landmarks for alignment; WIDER FACE terms not re-read. |
| MediaPipe BlazeFace / Face DetectorUnconfirmed | With conditionsWeights licence rests on an unread model card. | Not yetNeeds a loader first. | Permissive, no-WIDER option for close faces (doors, kiosks); misses small CCTV faces. |
| RetinaFace (biubug6 Pytorch_Retinaface; yakhyo/retinaface-pytorch) | With conditionsWeights licence not stated; WIDER FACE. | Not yetNeeds a loader first. | Accurate 5-point detection when YuNet misses; the yakhyo fork was not read. |
| libfacedetection | With conditionsWIDER FACE provenance. | Not yetNeeds a loader first. | Source project of YuNet, for retraining. |
| AMD RetinaFace int8 ONNX | With conditionsWIDER FACE provenance. | Not yetNeeds a loader first. | Ready quantised ONNX, built for AMD NPUs. |
| UltraFace (Linzaer) and CenterFace (Star-Clouds) | With conditionsWIDER FACE provenance. | Not yetNeeds a loader first. | Very light CPU detection, less accurate than YuNet; CenterFace licence not read. |
| MTCNN (ipazc/mtcnn, facenet-pytorch) | With conditionsWeights provenance. | Not yetNeeds a loader first. | Familiar and easy, but slow and dated; ipazc/mtcnn licence not read. |
| Intel OMZ face-detection-0200 / 0202 / 0204 / 0205 / 0206, adas-0001, retail-0004 | With conditionsTraining data undisclosed. | Not yetNeeds a loader first. | For OpenVINO on Intel CPUs. |
| dlib mmod CNN face detector | With conditionsSource image terms not itemised. | Not yetNeeds a loader first. | Explicit public-domain grant on the model files; needs dlib. |
| HF community YOLO face models (AdamCodd/YOLOv11n-face-detection, Bingsu/adetailer face_yolov8*, arnabdhar/YOLOv8-Face-Detection) | With conditionsEnterprise License or AGPL terms. | Yes | Only face detector that loads today, for internal tests; an Apache tag does not remove the AGPL. |
Avoid 5
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| dlib 68-point landmark model (also used by the face_recognition package)Avoid | Not rated | Not rated | The author states it cannot be used commercially. Use the 5-point landmarks from YuNet. |
| SCRFD (InsightFace)Avoid | Not rated | Not rated | Non-commercial research weights. |
| YOLO face models (YOLOv8-face by derronqi and akanametov, YOLOv5-Face by deepcam-cn)Avoid | Not rated | Not rated | GPL-3.0 repos on Ultralytics / YOLOv5 code. |
| LibreYOLO face-rfdetr-nanoAvoid | Not rated | Not rated | A toy checkpoint trained on 100 images. |
| OpenCV res10 SSD face detector (Caffe)AvoidUnconfirmed | Not rated | Not rated | A forum thread says no licence is given for the weights; replaced by YuNet. |
Face recognition11 modelsStart hereSFace (OpenCV Zoo, MobileFaceNet)With conditions
What it does: Matches a face to the people you have named.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| SFace (OpenCV Zoo, MobileFaceNet)Start here | With conditionsTraining data may include research-only sets. | Not yetNeeds its own integration. | Small baseline embedder through OpenCV FaceRecognizerSF. |
| AuraFace-v1 (glintr100)Start here | With conditionsTraining dataset undisclosed. | Not yetNeeds its own integration. | Stated commercial intent; the repo also holds InsightFace-looking files that must not be loaded. |
| FaceNet (davidsandberg, facenet-pytorch) | With conditionsVGGFace2 / CASIA-WebFace terms not read. | Not yetNeeds its own integration. | A familiar 512-d baseline; dated. |
| dlib ResNet v1 face recognition (and face_recognition package) | With conditionsSource dataset terms (FaceScrub, VGG). | Not yetNeeds its own integration. | CPU-friendly fallback; the wrappers also load the non-commercial 68-point model. |
| Intel OMZ face-reidentification-retail-0095 | With conditionsTraining data undisclosed. | Not yetNeeds its own integration. | For Intel CPU deployments; accuracy not read. |
Avoid 6
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| InsightFace packs (buffalo_l/s/m/sc, antelopev2, ArcFace R50/R100, w600k_r50)Avoid | Not rated | Not rated | Non-commercial research weights; commercial licence sold by InsightFace; the python package auto-downloads them. |
| EdgeFace (Idiap)Avoid | Not rated | Not rated | CC BY-NC-SA 4.0 weights. |
| AdaFaceAvoid | Not rated | Not rated | Weights trained on research-only face sets and published with no licence. |
| ONNX Model Zoo ArcFace (LResNet100E-IR) and OMZ face-recognition-resnet100-arcface-onnxAvoid | Not rated | Not rated | Apache-2.0 label on weights trained on MS-Celeb-1M. |
| MagFace, GhostFaceNets, CurricularFace and similar research reposAvoid | Not rated | Not rated | Weights trained on MS1M data and published without a licence. |
| Weight bundles (UniFace weights mirror, DeepFace)Avoid | Not rated | Not rated | Mixed per file; they contain InsightFace non-commercial files, GPL-3.0 YOLOv5-Face and unlicensed weights. |
Anti-spoofing6 modelsStart hereSilent-Face-Anti-Spoofing (MiniFASNetV1SE / V2)Allowed
What it does: Tells a real face from a photo or a screen held up to the camera.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| Silent-Face-Anti-Spoofing (MiniFASNetV1SE / V2)Start here | Allowed | Not yetNeeds its own integration. | Print and screen-replay checks; RGB only and needs a near-frontal face. |
| MiniFASNet ONNX and LiteRT ports (UniFace mirror, litert-community) | Allowed | Not yetNeeds its own integration. | Runs without PyTorch; the ONNX file sits in the mixed-terms UniFace bundle, so take only that file. |
Avoid 4
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| CDCN / CDCN++ (ZitongYu)Avoid | Not rated | Not rated | MIT text overridden by "just for research purpose, and commercial use should be contacted with authors first". |
| light-weight-face-anti-spoofing (kprokofi) and hairymax/Face-AntiSpoofingAvoid | Not rated | Not rated | Weights trained on CelebA-Spoof; hairymax has no licence file. |
| Hugging Face community liveness models (nguyenkhoa ViT/DINOv2, biometric-ai-lab/Antispoofing)Avoid | Not rated | Not rated | Trained on an unknown dataset, so the numbers and the licence mean little. |
| Spoof datasetsAvoid | Not rated | Not rated | CelebA-Spoof, OULU-NPU, CASIA-FASD (unconfirmed): terms not read; the survey records them as research and non-commercial. |
Safety checks
Phone use6 models · 1 load todayStart hereUltralytics YOLO11 / YOLO26 COCO weights, class 67 "cell phone"With conditions
What it does: Notices a phone in a person's hand where phones are not allowed.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| Ultralytics YOLO11 / YOLO26 COCO weights, class 67 "cell phone"Start here | With conditionsEnterprise License or AGPL terms. | Yes | In-house spike: no false alarms on 72 negatives, about half the phones found at about 220 px person height. |
| YOLOX (Megvii) COCO weightsLicence-safe alternative | Allowed | Not yetNeeds a loader first. | Clean-tier detector with ready ONNX files; the ONNX provider cannot decode its output as is. |
| RF-DETR N/S/M/L COCO weights (class "cell phone") | With conditionsPretraining data not stated. | Not yetNeeds a loader first. | Licence-clean phone detector from the same model that finds people; fine-tune on Open Images phone boxes. |
| D-FINE / DEIM / RT-DETRv2 COCO weights | Allowed | Not yetNeeds a loader first. | Alternative to RF-DETR with a COCO-only checkpoint; same decoder gap. |
| Open Images V7 "Mobile phone" boxes (fine-tuning data)Unconfirmed | Allowed | Data, not a model | Clean extra phone examples to raise recall; mostly hand-held close-ups. |
Avoid 1
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| sreenathsree1578/mobile-phone-usage-detectorAvoid | Not rated | Not rated | A whole-image classifier with unknown data, no licence and no phone location. |
Smoking5 models · 3 load todayStart hereBeehzod/smoking-detection-finetunedWith conditions
What it does: Notices smoking where it is not allowed.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| Beehzod/smoking-detection-finetunedStart here | With conditionsEnterprise License or AGPL terms. | Yes | In the model library as "Cigarette — YOLO11m (experimental)", pinned to commit d81c4396. The smoking check runs it on person crops at conf 0.5, and a phone or food box overlapping the hit vetoes it. Finds about half of site-scale cigarettes. |
| Pose rule (hand-to-mouth) on RTMPose keypoints plus a small own-trained cigarette/vape detectorLicence-safe alternative | With conditionsBody7 training-set terms not read. | Yes | Licence-clean way to build smoking; needs your own labelled clips. |
| Beehzod/smoke_cigarette-detection2-yolo11m and best_smoke_cigarette-detection | With conditionsEnterprise License or AGPL terms. | Yes | Second cigarette detector to compare; a 0.994 mAP50 on near-duplicate frames says little. |
| mbehbooei/vit-base-patch16-224-in21k-finetuned-smoking | With conditionsUnknown training data. | Not yetNeeds its own integration. | Whole-crop second opinion; no localisation, a lead only. |
| Roboflow Universe cigarette sets (xu-wei/cigarette-all and others)Unconfirmed | With conditionsLicence per set unknown. | Data, not a model | Only source of cigarette boxes the community models use; read each page before training. |
Fire and smoke13 models · 3 load todayStart hererabahdev/fire-smoke-yolov8nWith conditions
What it does: Notices flames or smoke in the picture.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| rabahdev/fire-smoke-yolov8nStart here | With conditionsEnterprise License or AGPL terms. | Yes | Works today; use conf 0.5 or more for fire. |
| RF-DETR or D-FINE fine-tuned on D-Fire + Pyro-SDIS + site negatives (recipe) | With conditionsRF-DETR pretraining data unstated; D-Fire authors hold no copyright over the images. | Not yetNeeds a loader first. | The licence-clean detector the product can own; the D-FINE COCO-only base avoids the RF-DETR question. |
| pyronear/rexnet1_0x and other Pyronear classifiers (mobilenet_v3, resnet18/34) | Allowed | Not yetNeeds its own integration. | Cheap whole-frame wildfire yes/no gate for an expensive detector; no boxes, outdoor imagery. |
| D-Fire dataset | Allowed | Data, not a model | Main permissive fire and smoke set. |
| Pyro-SDIS dataset | Allowed | Data, not a model | Distant smoke from fixed cameras; outdoor wildfire views only. |
| xiazh0219/yolo26m-UAV-fire-smoke-detection | With conditionsEnterprise License or AGPL terms. | Yes | Aerial and outdoor smoke; not for ceiling or wall cameras. |
| Pyronear YOLO (yolov8s, yolov11s/n, yolo11s_*) | With conditionsEnterprise License or AGPL terms. | Yes | Early smoke at long range from fixed towers; indoor steam and dust will trigger it. |
| fireviewer/rtdetr-v2-r50-fire-smoke | With conditionsNo redistribution, public credit; get a written agreement first. | Not yetNeeds a loader first. | Fine for a hosted service you run yourself; shipping it inside installed software is redistribution. |
| EdBianchi/vit-fire-detection and fireviewer/dinov2-small-fire-associated-smoke-v2.5 | With conditionsFine-tune data unclear. | Not yetNeeds its own integration. | Whole-frame fire classifiers to confirm a detector alarm. |
| Intel/fire-and-smoke-detection (Qwen2-VL-2B, OpenVINO) | Allowed | Not yetNeeds its own integration. | Language-prompted verifier for an alarm candidate on Intel hardware; slow and gives no box. |
| FASDDUnconfirmed | With conditionsLicence carried from the survey, not re-verified. | Data, not a model | Large aerial and ground fire/smoke set. |
Avoid 2
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| Simuletic CCTV smoke/fire and wildfire smoke datasetsAvoidUnconfirmed | Not rated | Not rated | CC BY-NC 4.0, even though other Simuletic sets are CC BY 4.0. |
| Community fire/smoke YOLO fine-tunes (mfranzon, nobita111222, fireviewer yolo11m, Beehzod yolov26 ...)Avoid | Not rated | Not rated | All AGPL-3.0 by inheritance; a CC BY tag does not lift the framework's AGPL. |
Falls11 models · 4 load todayStart hereRTMPose (rtmlib / MMPose) plus fall rulesWith conditions
What it does: Notices a person who has fallen and stays down.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| RTMPose (rtmlib / MMPose) plus fall rulesStart here | With conditionsBody7 training-set terms not read. | Yes | Loads now; the person-down pose rule is built on it. |
| ViTPose (usyd-community, Hugging Face) | With conditionsPose training-set terms not read. | Not yetNeeds a loader first. | More accurate keypoints than RTMPose when the box is large; top-down, heavier. |
| Ultralytics YOLO11-pose | With conditionsEnterprise License or AGPL terms. | Yes | Lets you try the fall rules before the Apache route exists. |
| Ultralytics YOLO26-pose | With conditionsEnterprise License or AGPL terms. | Yes | In the model library as "Pose — YOLO26 (official)" (weights from ultralytics assets v8.4.0). Feeds the Person down check, method pose; s is the default. |
| Simuletic CCTV Incident Dataset: Fall / Lying Down (synthetic)Unconfirmed | Allowed | Data, not a model | Overhead CCTV lying-down poses to test the rules; small and synthetic. |
| MediaPipe Pose Landmarker (BlazePose)Unconfirmed | Allowed | Not yetNeeds a loader first. | Tiny single-person pose on a person crop, runs on CPU. |
| melihuzunoglu/human-fall-detection | With conditionsEnterprise License or AGPL terms. | Yes | Quick fall-by-posture detector; cannot tell lying down from a fall without a time rule. |
| pat2echo/fall-detection-posture-classification (dataset)Unconfirmed | Allowed | Data, not a model | Clean clips with consent stated; a single subject in one setting. |
| PoseC3D / ST-GCN skeleton classifiers (MMAction2)Unconfirmed | With conditionsPublic checkpoints carry research terms. | Not yetNeeds a loader first. | A learned classifier once you have your own clips; do not use the public checkpoints. |
Avoid 2
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| kamalchibrani/yolov8_fall_detectionAvoid | Not rated | Not rated | No documentation, and the mit tag is unreliable on an AGPL-3.0 base. |
| UR Fall Detection and OmniFall datasetsAvoidUnconfirmed | Not rated | Not rated | CC BY-NC-SA 4.0 and cc-by-nc-4.0, non-commercial. Retire the internal URFD test. |
Forklift6 models · 3 load todayStart hereRF-DETR or D-FINE fine-tuned on forklift + person site frames (recipe)With conditions
What it does: Finds forklifts, and people too close to one.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| RF-DETR or D-FINE fine-tuned on forklift + person site frames (recipe)Start here | With conditionsRF-DETR pretraining data unstated. | Not yetNeeds a loader first. | The licence-clean route; D-FINE COCO-only is the alternative base. |
| mustafahacioglu/yolo11s-forklift-person-detectionStart here | With conditionsEnterprise License or AGPL terms. | Yes | Works today; its card adds a box-containment rule for operator presence; the map_up licence is unread. |
| yihong1120/Construction-Hazard-Detection | With conditionsEnterprise License or AGPL terms. | Yes | Works today as an alternative forklift source: it reports forklifts as "machinery", and the catalogue ignores its "vehicle" class because it false-alarms on site. |
| keremberke/yolov8{n,s,m}-forklift-detection (and yolov5 versions) | With conditionsEnterprise License or AGPL terms. | Yes | Tiny model for a first demo; a Roboflow-derived forklift model once boxed the whole frame. |
| Roboflow forklift datasets (keremberke/forklift-object-detection, forklift-dsitv)Unconfirmed | Allowed | Data, not a model | Seed images; add your own site frames. |
Avoid 1
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| akdve/ot-forklift-person-yolo11s-smokeAvoid | Not rated | Not rated | A one-epoch, two-image pipeline test. |
Blocked exits and clear walkways5 models · Not in the platform yetStart hereZone comparison against an empty reference (OpenCV differencing or DINOv2 patch features)Allowed
What it does: Alerts when an exit, aisle or lane is blocked, with a dated record. Needs: Fixed camera view, zone, empty reference. Main risk: People walking through; light and camera shift.
| Model | Commercial use | Good to know |
|---|---|---|
| Zone comparison against an empty reference (OpenCV differencing or DINOv2 patch features)Start here | Allowed | Is the exit zone different from its empty reference for 3 minutes with nobody passing? Needs no labels.Would run on: CPU |
| RF-DETR-Seg fine-tuned on "obstruction" (pallet, trolley, boxes)Unconfirmed | Allowed | Names the blocker and gives its mask for the audit record. Needs labelled images of real clutter.Would run on: Consumer GPU |
| SAM 2 for labelling and mask clean-up | Allowed | Speeds up labelling; not run live on every camera.Would run on: Consumer GPU |
| OneFormer (ADE20K) and Mask2Former (COCO) | With conditionsADE20K terms unread; Mask2Former licence unresolved. | Labels "floor" and "wall" for walkable-area masks without site labels.Would run on: Consumer GPU |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| SegFormer (NVIDIA)Avoid | Not rated | Non-commercial NVIDIA licence; a popular default in tutorials. |
Left-behind and removed objects4 models · Not in the platform yetStart hereOpenCV background subtractors (MOG2, KNN)Allowed
What it does: Alerts when something appears and stays, or something that should stay is gone. Needs: Reference frame, zone, dwell time; CPU. Main risk: Light changes and moved stock cause false alarms.
| Model | Commercial use | Good to know |
|---|---|---|
| OpenCV background subtractors (MOG2, KNN)Start here | Allowed | A slow and a fast background: a blob only in the slow one is left behind, one missing from it is removed.Would run on: CPU |
| DINOv2 patch-feature comparison with a reference | Allowed | Zone-level "same as the reference" check that tolerates lighting change. Coarse (patch size 14).Would run on: Consumer GPU or CPU (small) |
| RF-DETR (COCO classes such as backpack, suitcase, handbag, bottle) | Allowed | Names the thing left behind. Yes for N to L only; a pallet or tool is not a COCO class.Would run on: Consumer GPU |
| ChangeFormer | With conditionsWeights licence not stated. | Change between two pictures; trained on satellite pairs, a poor match to ground-level CCTV.Would run on: Consumer GPU |
Objects left behind or removed6 models · Not in the platform yetStart hereDetector plus tracker plus dwell-time rule (RF-DETR or D-FINE with the trackers library)Allowed
What it does: Alerts when a bag, box or tool stays or disappears. Needs: Fixed camera, zones, dwell time, reference image. Main risk: Shadows and pausing people cause false alarms.
| Model | Commercial use | Good to know |
|---|---|---|
| Detector plus tracker plus dwell-time rule (RF-DETR or D-FINE with the trackers library)Start hereUnconfirmed | Allowed | Reuses the person detector and tracker. Needs a person-proximity test so someone sitting by a bag is not an alert.Would run on: Consumer GPU |
| OpenCV background subtraction (MOG2, KNN) with static-foreground logic | Allowed | Finds anything new and still in a zone, whatever it is. Cheapest first filter.Would run on: CPU |
| BGSLibrary (40 or more algorithms) | With conditionsBundles third-party algorithm code with its own terms; some patented. | Try many algorithms to find one that survives your lighting.Would run on: CPU |
| SAM 2 / SAM 2.1 (object masks for "what changed") | Allowed | Cuts a clean mask of a changed region and drafts labels. Not for every frame on every camera.Would run on: Consumer GPU |
| DINOv2 patch-feature difference against an empty-zone reference | Allowed | More robust to light and shadow than pixel differencing. Needs a reference frame per camera.Would run on: Consumer GPU, or CPU at low rates |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| CoTracker3 (point tracking)Avoid | Not rated | CC BY-NC 4.0 weights. |
Distance to a hazard from one camera (depth)5 models · Not in the platform yetStart hereDepth Anything V2 Small (relative depth)Allowed
What it does: A rough near-or-far hint without a floor map. Needs: Depth model; floor reference for metres. Main risk: Relative depth only; never a safe-distance decision.
| Model | Commercial use | Good to know |
|---|---|---|
| Depth Anything V2 Small (relative depth)Start here | Allowed | Which of two things is nearer. Larger models are non-commercial. Training data not itemised; record the risk decision.Would run on: CPU at low frame rates, or small GPU |
| MiDaS and DPT (Intel), and ZoeDepth | With conditionsMixed training-data terms; KITTI is CC BY-NC-SA. | Mature baselines; older and heavier than Depth Anything V2 Small.Would run on: Consumer GPU |
| Depth Anything V2 Metric, SmallUnconfirmed | With conditionsVirtual KITTI 2 and Hypersim terms. | Metres directly, indoors. Outdoor weights likely inherit non-commercial terms.Would run on: Small GPU |
| Metric3D v2 | With conditionsWeights terms ambiguous. | Metric depth using known intrinsics; heavy.Would run on: Large GPU |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| Depth Anything V2 Base, Large, Giant and Apple Depth ProAvoid | Not rated | CC BY-NC 4.0 and research-only licences. |
Vehicles and plates
Vehicles23 models · 2 load todayStart hereRF-DETR (Nano, Small, Medium, Large)Allowed
What it does: Finds vehicles and says what kind, for gates, yards and loading bays.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| RF-DETR (Nano, Small, Medium, Large)Start here | Allowed | Not yetNeeds a loader first. | Default detector once the DETR decoder exists; the weights licence rests on the README table. |
| D-FINE (N, S, M, L, X)Licence-safe alternative | Allowed | Not yetNeeds a loader first. | Second detector to benchmark against RF-DETR; use the plain COCO files. |
| YOLOX (nano to X)Licence-safe alternative | Allowed | Not yetNeeds a loader first. | Runs well on CPU at small sizes; needs a small decoder for its ONNX output. |
| DINOv2 features plus a small classifier head (train on detector crops) | Allowed | Not yetNeeds its own integration. | Vehicle type and colour, including local classes, from a few hundred crops per class; do not swap in DINOv3. |
| RT-DETR and RT-DETRv2 | Allowed | Yes | Easiest DETR to prototype with in Transformers; avoid the AGPL ultralytics port. |
| YOLOv9 MIT re-implementation (MultimediaTechLab/YOLO, mirrored as LibreYOLO9) | With conditionsWeights licence only asserted by a mirror. | Not yetNeeds a loader first. | Nearest licence-clean drop-in for ultralytics weights; not the GPL-3.0 WongKinYiu repo. |
| RF-DETR Atto, Femto, Pico, XL, 2XL (rfdetr_plus)Unconfirmed | With conditionsPML 1.0 text not found. | Not yetNeeds a loader first. | Only with a Roboflow plan; the Apache sizes cover the same need. |
| LW-DETR (T, S, M, L, X) | Allowed | Not yetNeeds a loader first. | The architecture RF-DETR builds on; research reference only. |
| DAMO-YOLO (Nano to L)Unconfirmed | Allowed | Not yetNeeds a loader first. | No reason to prefer it over YOLOX or RF-DETR. |
| PP-YOLOE+ and PP-Vehicle (PaddleDetection)Unconfirmed | Allowed | Not yetNeeds a loader first. | Only if the team already runs Paddle; PaddleYOLO ports (YOLOv5/6/7/8) are GPL-3.0, do not use those. |
| RTMDet (MMDetection)Unconfirmed | With conditionsNo weights licence stated. | Not yetNeeds a loader first. | Mature CPU-friendly detector; last push 2024-08. |
| Ultralytics YOLO11 and YOLO26 | With conditionsUltralytics Enterprise licence. | Yes | Only family that loads today; AGPL applies even to internal use or training from scratch. |
| NVIDIA TrafficCamNetUnconfirmed | With conditionsLicence text unread. | Not yetNeeds a loader first. | Cars, persons, road signs and two-wheelers; US training data. |
| NVIDIA VehicleTypeNet and VehicleMakeNetUnconfirmed | With conditionsLicence text not read. | Not yetNeeds its own integration. | Six US-style classes; no motorcycle, three-wheeler or minibus taxi. |
| DINOv3 (facebook/dinov3-*)Unconfirmed | With conditionsLicence text not read. | Not yetNeeds its own integration. | Stronger features than DINOv2; DINOv2 is the safe choice. |
| OWLv2 (google/owlv2-*) | Allowed | Not yetNeeds a loader first. | Text-prompted pre-labelling of local vehicle types; heavy, not for live streams. |
| Grounding DINO tiny (IDEA-Research) | With conditionsTraining sets not checked. | Not yetNeeds a loader first. | Text-prompted boxes to bootstrap labels for local classes; not a live detector. |
Avoid 6
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| YOLOv10, YOLOv12 and YOLOv13 research reposAvoid | Not rated | Not rated | AGPL-3.0 with no commercial route short of the repo owners. |
| YOLOv7, official YOLOv9 and YOLOv6Avoid | Not rated | Not rated | GPL-3.0; the same authors' MultimediaTechLab/YOLO is MIT, do not mix them up. |
| YOLO-NASAvoid | Not rated | Not rated | The Deci licence bars commercial use; the Apache code does not cover the weights. |
| DEIMv2 (and EdgeCrafter)Avoid | Not rated | Not rated | DEIMv2 License, non-commercial. |
| Community YOLO vehicle detectors on Hugging Face (for example Intel/vehicle-detection)Avoid | Not rated | Not rated | The card says MIT but the base is Ultralytics, so the AGPL stays. |
| YOLOv5 and YOLO-World (AGPL and GPL)Avoid | Not rated | Not rated | Copyleft. |
Plate detection12 modelsStart hereRF-DETR fine-tuned on local plate boxesAllowed
What it does: Finds number plates in the picture.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| RF-DETR fine-tuned on local plate boxesStart here | Allowed | Not yetNeeds a loader first. | Clean route for local plate shapes and camera angles; needs labelled plates. |
| autolane/yolox-s-alprLicence-safe alternative | With conditionsDataset terms unverified. | Not yetNeeds a loader first. | Permissive detector with an INT8 TFLite; US plates only in training. |
| open-image-models YOLOv9 plate detectorsUnconfirmed | With conditionsWeights licence and data not written down. | Not yetNeeds a loader first. | Small, fast first detector (t-384 is 7.8 MB); ONNX has NMS baked in, not YOLO-style raw output. |
| autolane/rfdetr-alpr | With conditionsTraining data undisclosed. | Not yetNeeds a loader first. | Shows an RF-DETR plate detector already exported to ONNX and INT8. |
| justjuu/rtdetr-v2-license-plate-detection | With conditionsDataset licence not read. | Not yetNeeds a loader first. | Transformers-loadable plate detector to compare against; small, narrow training set. |
| HyperLPR3Unconfirmed | With conditionsWeights licence and data not read. | Not yetNeeds a loader first. | Only for Chinese plates. |
| NVIDIA LPDNetUnconfirmed | With conditionsLicence text not retrievable. | Not yetNeeds a loader first. | Only in a TAO and TensorRT workflow for US plates. |
Avoid 5
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| Community ultralytics plate detectors on Hugging Face (YOLOv5, v8, v11, v26)Avoid | Not rated | Not rated | Cards vary (agpl-3.0, mit, none) but all sit on AGPL-3.0 code. |
| Intel license-plate-recognition (YOLOv8 plus PP-OCRv4)Avoid | Not rated | Not rated | MIT card label over an ultralytics-trained detector; use the recogniser idea, not the detector. |
| Nomeroff-NetAvoid | Not rated | Not rated | GPL-3.0 copyleft covers any product that links it. |
| OpenALPRAvoid | Not rated | Not rated | AGPL-3.0, old, and the maintainer sells commercial products. |
| alpr-unconstrained / WPOD-NETAvoid | Not rated | Not rated | GPL-3.0 research code from 2018. |
Plate reading16 models · 1 load todayStart herefast-plate-ocr (cct-s-v2-global, cct-xs-v2-global and older)With conditions
What it does: Reads the characters on a number plate.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| fast-plate-ocr (cct-s-v2-global, cct-xs-v2-global and older)Start hereUnconfirmed | With conditionsWeights licence and data not written down. | Not yetNeeds its own integration. | Fast Latin-script reader that can be fine-tuned on local plates; Latin letters and digits only. |
| RapidOCRLicence-safe alternative | Allowed | Not yetNeeds its own integration. | Runs PP-OCR models without Paddle installed; same non-plate tuning. |
| PaddleOCR PP-OCRv5 recognisers (default, latin, arabic, devanagari, cyrillic, ta, te, th, korean, el) | Allowed | Yes | Only permissive route to non-Latin scripts; not tuned for plates, so fine-tune on cropped plates. |
| Awiros-ANPR-OCR (Indian plates) | With conditionsUpstream data licences not stated. | Not yetNeeds its own integration. | Indian plates, one or two rows; self-reported accuracy. |
| EasyOCR | Allowed | Not yetNeeds its own integration. | Quick baseline with Bengali, Hindi, Marathi, Nepali, Tamil, Telugu, Urdu, Arabic and Persian; no Sinhala. |
| Tesseract with tessdata_best | Allowed | Not yetNeeds its own integration. | Ready packs for Bengali, Hindi, Sinhala, Tamil, Nepali and Urdu; poor on small blurry plate crops. |
| docTRUnconfirmed | With conditionsWeights licences and data never read. | Not yetNeeds its own integration. | Document-style OCR; not plate-specific. |
| PaddleOCR-VL | Allowed | Not yetNeeds its own integration. | Second-opinion read of hard plates on a GPU, not every frame. |
| GOT-OCR2.0 | Allowed | Not yetNeeds its own integration. | General OCR second opinion; mostly English and Chinese. |
| DeepSeek-OCR | Allowed | Not yetNeeds its own integration. | Heavy GPU fallback for odd plates; too slow for live reading. |
| NVIDIA LPRNetUnconfirmed | With conditionsLicence text not retrievable. | Not yetNeeds its own integration. | Only US or Chinese plates inside a TAO workflow. |
| LPRNet_Pytorch (sirius-ai)Unconfirmed | With conditionsWeights licence not stated. | Not yetNeeds its own integration. | A small architecture to retrain on local plates; shipped weights read Chinese plates only. |
| TrOCR (microsoft/trocr-base-printed) and plate fine-tunesUnconfirmed | With conditionsNo weights licence stated. | Not yetNeeds its own integration. | Appears in many plate tutorials; plate fine-tunes carry no stated licence. |
| hezarai crnn-fa-license-plate-recognition-v2 (Persian plates)Unconfirmed | With conditionsNo licence field. | Not yetNeeds its own integration. | Only non-Latin (Persian) plate reader found; a pattern for an Arabic-script model. |
Avoid 2
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| Surya OCRAvoid | Not rated | Not rated | Modified AI Pubs Open RAIL-M weights, free for research, personal use and startups under 5 M funding or revenue. |
| HunyuanOCRAvoid | Not rated | Not rated | Tencent Hunyuan Community License does not apply in the EU, UK and South Korea. |
Picture and camera
Camera health10 models · 1 load todayStart hereOpenCV classical checks (Laplacian variance, histograms, edge density, ORB or phase-correlation shift, SSIM)Allowed
What it does: Tells the operator when a camera is blurred, blocked, moved, dark or frozen. Main risk: False alarms at dusk, rain and glare; reference must not refresh itself.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| OpenCV classical checks (Laplacian variance, histograms, edge density, ORB or phase-correlation shift, SSIM)Start here | Allowed | Yes | Blur, dark, blocked, moved and frozen frames. Tune thresholds per camera and for day and night. |
| DINOv2 embedding drift against a reference frameLicence-safe alternative | Allowed | Not yetNo. | Scene-level check that survives lighting change. Cannot tell a real scene change from tamper. |
| HyperIQA | With conditionsWeights licence not stated; trained on KonIQ-10k. | Not yetNo. | Learned quality score as a second opinion beside the classical checks. |
| BRISQUE (OpenCV contrib quality module) | With conditionsTrained on LIVE-R2; its terms not read. | Not yetNo. | One no-reference sharpness number. Built for photos, not CCTV; plain Laplacian variance is usually enough. |
| MUSIQ (Google Research) | With conditionsCheckpoint data terms not read. | Not yetNo. | Scores images at any resolution. Porting effort is real. |
| LIQE | With conditionsWeights and dataset terms unclear. | Not yetNo. | Quality score plus a distortion label (blur, noise, over-exposure) that explains the flag. |
| ARNIQA | With conditionsNo weights licence in the README. | Not yetNo. | Learns quality without human scores; a second opinion. |
Avoid 3
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| Q-Align / OneAlignAvoid | Not rated | Not rated | S-Lab code licence (non-commercial); a 7B model to find a smudged lens is far too heavy. |
| IQA-PyTorch (pyiqa)Avoid | Not rated | Not rated | PolyForm Noncommercial code and CC BY-NC-SA weights; easy to pull in by accident. Use the individual MIT or Apache repositories. |
| DOVER and CLIP-IQA (NTU S-Lab)Avoid | Not rated | Not rated | S-Lab License 1.0, non-commercial. |
Camera tamper and picture quality6 models · 2 load todayStart hereOpenCV image checks (Laplacian blur, brightness, SSIM, frozen frame)Allowed
What it does: Flags blocked, dirty, moved or frozen cameras before an incident. Main risk: Rain, glare and infrared switching cause false alarms.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| OpenCV image checks (Laplacian blur, brightness, SSIM, frozen frame)Start here | Allowed | Yes | Near-zero cost on every camera every minute. Thresholds per camera. |
| OpenCV ORB or AKAZE matching with a homography | Allowed | Yes | Detects that the camera moved and by how much, so zones can be marked invalid. Needs background features. |
| XFeat | With conditionsMegaDepth and COCO terms unread. | Not yetNo. | Learned features that survive light change better than ORB; ORB is usually enough. |
| DINOv2 linear probe for fog, rain, lens dirt and glare | Allowed | Not yetNo. | Names the cause of a poor picture; needs labelled examples per condition. |
| LightGlue (with DISK or ALIKED) | With conditionsNever ship the SuperPoint variant; training data unread. | Not yetNo. | Accurate view-change estimate when the scene is hard for ORB. |
Avoid 1
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| pyiqa (MUSIQ, CLIP-IQA, TOPIQ, NIMA)Avoid | Not rated | Not rated | PolyForm Noncommercial library; classical checks do the job. |
Picture improvement20 models · Not in the platform yetStart hereOpenCV classical enhancement (CLAHE, gamma, non-local-means, dehaze)Allowed
What it does: Easier review of dark, noisy or hazy footage. Needs: CPU for classical steps; GPU for learned models. Main risk: Learned models invent detail; never evidence.
| Model | Commercial use | Good to know |
|---|---|---|
| OpenCV classical enhancement (CLAHE, gamma, non-local-means, dehaze)Start here | Allowed | Contrast lift and denoise with no invented detail. Dehaze struggles with steam and bright sky.Would run on: CPU |
| Real-ESRGAN (x4plus, general-x4v3) | With conditionsWeights licence unstated; source photo terms unread. | Upscales a small crop for a human viewer. Invents texture; no identification or evidence.Would run on: Consumer GPU |
| SwinIR and Swin2SR | Allowed | More faithful than GAN upscalers; Swin2SR loads in Transformers. Training set terms unread.Would run on: Consumer GPU |
| SPAN and community SPAN models | With conditionsTraining photo sets unread; no LICENSE file at HEAD. | Near real-time upscaling and denoising of a live view. Read each model card.Would run on: Consumer GPU |
| SCUNet | With conditionsWeights terms not stated. | Blind denoising of high-ISO night footage.Would run on: Consumer GPU |
| NAFNet | With conditionsWeights terms unstated; benchmark sets are research-only. | Fast denoise and motion-blur removal; best speed-quality trade-off.Would run on: Consumer GPU |
| Retinexformer | With conditionsWeights terms not stated. | Low-light enhancement with a tiny network. Photographic sets, not infrared CCTV.Would run on: Consumer GPU |
| OpenCV videostab (opencv_contrib) | Allowed | Stabilises a shaking pole-mounted camera. Cropping cuts the field of view.Would run on: CPU |
| HAT (Hybrid Attention Transformer) | With conditionsWeights licence unstated. | Highest still-image super-resolution among open transformers; heavy for video.Would run on: Consumer GPU (slow) |
| Restormer | With conditionsWeights terms not stated. | Deblur, denoise, deraining for stills; too slow for many streams.Would run on: Consumer GPU (heavy) |
| HVI-CIDNet | With conditionsWeights terms not stated. | Low-light with fewer colour shifts; newer, less tested.Would run on: Consumer GPU |
| DehazeFormer | With conditionsTrained on RESIDE; weights terms not stated. | Haze on outdoor yards. Never dehaze the stream a fire detector reads.Would run on: Consumer GPU |
| BasicVSR++ (via MMagic) | With conditionsWeights terms not stated. | Offline clean-up of a short clip. The licence shown is MMagic's.Would run on: Consumer GPU (offline) |
| vid.stab | With conditionsLGPL; dynamic linking only, take advice. | Mature two-pass stabiliser used inside ffmpeg.Would run on: CPU |
Avoid 6
| Model | Commercial use | Good to know |
|---|---|---|
| Zero-DCE, Zero-DCE++ and SCIAvoid | Not rated | No licence file in either repository, so no right to use. Use Retinexformer or CLAHE. |
| VRT and RVRTAvoid | Not rated | CC BY-NC 4.0, non-commercial. |
| DiffBIRAvoid | Not rated | Diffusion invents plausible detail, the worst fit for evidence; built on Stable Diffusion weights with OpenRAIL++-M terms. |
| GFPGANAvoid | Not rated | Generates a face that is not the person's; a confident wrong identity is serious harm. |
| CodeFormerAvoid | Not rated | S-Lab non-commercial licence and the same invented-face problem. |
| DUT video stabilisationAvoid | Not rated | LICENSE says MIT but the README says research only; the conflict is the risk. |
Image restoration for CCTV frames4 models · Not in the platform yetStart hereReal-ESRGANWith conditions
What it does: Denoise, deblur and upscale saved frames. Needs: A GPU; side-by-side original; enhanced watermark. Main risk: Invented detail; face restorers are banned.
| Model | Commercial use | Good to know |
|---|---|---|
| Real-ESRGANStart here | With conditionsWeights trained on DIV2K, which is academic-only. | Upscales and cleans a still. Mark restored images; it invents texture.Would run on: Consumer GPU (CPU for single stills) |
| SwinIR, HAT, NAFNet, Restormer, SCUNet, BSRGANUnconfirmed | With conditionsTraining sets are academic-only. | Denoise, deblur and super-resolution. Retrain on licensed or own data.Would run on: Consumer GPU |
| RealBasicVSR, FlashVSR and DiffBIRUnconfirmed | With conditionsWeights and data not read. | Cleans short clips. Diffusion restoration hallucinates the most.Would run on: Consumer GPU |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| CodeFormer, GFPGAN and SUPIRAvoid | Not rated | They rebuild faces and can make a stranger look like a suspect; avoid on ethical grounds whatever the licence. |
Depth and geometry from one camera17 models · Not in the platform yetStart hereGround-plane homography from clicked floor points (OpenCV)Allowed
What it does: Distances and floor-plan positions from one fixed camera. Needs: One floor calibration per camera; floor plan. Main risk: Model depth is off by percent to tens of percent; location tracking of workers.
| Model | Commercial use | Good to know |
|---|---|---|
| Ground-plane homography from clicked floor points (OpenCV)Start here | Allowed | The reliable way to measure floor distance in metres. Flat floor only; redo when the camera moves.Would run on: CPU |
| Depth Anything V2 Small | Allowed | Relative depth only. Base, Large and Giant are CC-BY-NC-4.0.Would run on: CPU or consumer GPU |
| Depth Anything 3 (Small, Base, Metric-Large, Mono-Large) | Allowed | Metric depth (DA3METRIC-LARGE) for yards. Take only the Apache checkpoints; check metres against a tape.Would run on: Consumer GPU |
| GeoCalib | Allowed | Estimates focal length, distortion and tilt from one picture. Weak on featureless scenes.Would run on: Consumer GPU or CPU |
| MoGe and MoGe-2 | Allowed | Point map with metric scale, normals and field of view in one pass. Training-data terms unread.Would run on: Consumer GPU |
| MiDaS and DPT (Intel) | Allowed | Older relative depth with first-class Transformers support. Yes is at licence level only.Would run on: CPU or consumer GPU |
| Depth Anything V2 metric (Hypersim, Virtual KITTI 2)Unconfirmed | With conditionsDataset terms unread; outdoor inherits a non-commercial set. | Metres directly. Prefer DA3METRIC-LARGE or MoGe-2.Would run on: Consumer GPU |
| ZoeDepth | With conditionsKITTI is non-commercial. | Early metric depth; superseded.Would run on: Consumer GPU |
| Marigold (diffusion depth) | With conditionsTraining data includes a non-commercial set. | Very sharp relative depth for one still. Slow.Would run on: Consumer GPU (slow) |
| Metric3D v2 | With conditionsWeights terms unconfirmed; authors invite commercial enquiries. | Metric depth that needs true focal length.Would run on: Large GPU |
| Video Depth Anything | Allowed | Steady depth for video; use Small only.Would run on: Consumer GPU |
| WildCamera | With conditionsWeights licence not found. | Focal length from an in-the-wild image; an alternative to GeoCalib.Would run on: Consumer GPU |
Avoid 5
| Model | Commercial use | Good to know |
|---|---|---|
| Apple Depth ProAvoid | Not rated | Apple research-only licence. |
| UniDepth (V1, V2)Avoid | Not rated | CC BY-NC 4.0. |
| VGGT (including VGGT-1B-Commercial)AvoidUnconfirmed | Not rated | VGGT-1B is CC BY-NC 4.0; the commercial licence text is gated and unread; multi-view anyway. |
| DUSt3R and MASt3RAvoid | Not rated | CC BY-NC-SA 4.0. |
| Perspective FieldsAvoid | Not rated | Adobe noncommercial research licence. |
Segmentation for zones and masks14 models · Not in the platform yetStart hereSAM 2 and SAM 2.1 (Tiny, Small, Base+, Large)Allowed
What it does: One click on a floor or door gives a clean zone outline. Needs: A setup screen where a human accepts each outline. Main risk: Rough edges at shadows and glass; some models are non-commercial or AGPL.
| Model | Commercial use | Good to know |
|---|---|---|
| SAM 2 and SAM 2.1 (Tiny, Small, Base+, Large)Start here | Allowed | Click a point to get a zone outline; tracks a mask through a clip. The dataset chain is a theoretical risk.Would run on: Consumer GPU (Tiny and Small on a modest one) |
| RF-DETR Seg (N to 2XL) | Allowed | Person and vehicle masks for foot-in-zone tests. COCO classes only.Would run on: Consumer GPU |
| Segment Anything (SAM 1: ViT-B, L, H) | Allowed | Well-known still-image click-to-mask; SAM 2.1 is the better start.Would run on: Consumer GPU (ViT-H heavy) |
| MobileSAM | Allowed | Click-to-mask without a strong GPU. Rougher edges.Would run on: CPU or any GPU |
| EfficientViT-SAM and EfficientSAM | Allowed | Faster alternatives to SAM. Port effort.Would run on: Consumer GPU |
| SAM 3 | With conditionsCustom licence with termination and litigation clauses; read it closely. | Segments every instance from a text phrase.Would run on: Consumer GPU to large GPU |
| Florence-2 (base, large) | Allowed | Referring segmentation in one small model; masks less accurate than SAM.Would run on: CPU or consumer GPU |
| Grounded-SAM-2 recipe (Grounding DINO plus SAM 2) | Allowed | Type "conveyor belt" and get a mask for a zone proposal. A human confirms every zone.Would run on: Consumer GPU |
| OneFormer and Mask2Former (ADE20K semantic) | With conditionsADE20K and Cityscapes terms are research-oriented. | Labels floor, wall, door and stairs to suggest walkable zones.Would run on: Consumer GPU |
| EoMT (Encoder-only Mask Transformer) | With conditionsADE20K and COCO terms unconfirmed. | Fast plain-ViT segmentation.Would run on: Consumer GPU |
| CLIPSeg | With conditionsConflicting statements about the weights. | Cheap text-prompt segmentation. Get a clear answer first.Would run on: CPU or consumer GPU |
Avoid 3
| Model | Commercial use | Good to know |
|---|---|---|
| SegFormer (NVIDIA)Avoid | Not rated | NVIDIA Source Code License, research or evaluation only; easy to adopt by accident. |
| FastSAMAvoid | Not rated | AGPL-3.0 (YOLOv8-seg based); MobileSAM does the job under Apache. |
| EdgeSAM and Bria RMBGAvoid | Not rated | S-Lab and Bria non-commercial terms. BiRefNet (MIT) covers background removal. |
Zone outlines and depth hints3 models · 2 load todayStart hereSAM 2.1 Hiera Tiny (zone outline)Allowed
What it does: Two helpers the Models page installs. Neither runs on every frame and neither enables a safety check by itself.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| SAM 2.1 Hiera Tiny (zone outline)Start here | Allowed | Yes | No pinned ONNX file exists, so the install downloads the commit-pinned checkpoint, checks its SHA-256, and exports the encoder and decoder on the site with the trainer's torch. Small, base+ and large are not offered: too slow on a processor. |
| Depth Anything V2 Small (depth hint) | Allowed | Yes | One commit-pinned ONNX file from the onnx-community mirror, checked by SHA-256. Relative depth only, never metres. |
Avoid 1
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| Depth Anything V2 Base and LargeAvoid | Not rated | Not rated | CC BY-NC-4.0, non-commercial. They are not offered. |
Optical flow and motion10 models · Not in the platform yetStart hereOpenCV DIS and Farneback dense optical flowAllowed
What it does: Direction and speed of movement, a stopped conveyor, a moved camera. Needs: Frame pairs; floor calibration for real units. Main risk: Noise in rain and glare; must not score workers.
| Model | Commercial use | Good to know |
|---|---|---|
| OpenCV DIS and Farneback dense optical flowStart here | Allowed | Direction and speed per zone, camera shake, conveyor stopped. Noisy on plain surfaces and in rain.Would run on: CPU |
| SEA-RAFT (S, M, L) | With conditionsPick after a licence answer; KITTI and Sintel terms. | Accurate learned flow at modest cost for vehicle speed and direction.Would run on: Consumer GPU |
| RAFT (original and torchvision port) | With conditionsKITTI is non-commercial; synthetic-only weights avoid it. | The reference learned flow with first-class torchvision support.Would run on: Consumer GPU |
| NeuFlow v2Unconfirmed | With conditionsWeights licence not found. | Flow on edge hardware. The two speed claims disagree; test on the real board.Would run on: Jetson-class board or consumer GPU |
| WAFT | With conditionsWeights licence not found. | Newer warping-based flow from the RAFT group.Would run on: Consumer GPU |
| UniMatch (flow, stereo, depth) | With conditionsWeights licence unstated. | One codebase for flow, stereo and depth.Would run on: Consumer GPU |
| GMFlow, FlowFormer and MemFlowUnconfirmed | With conditionsWeights terms not stated. | Research-grade; heavier than SEA-RAFT for little gain on CCTV.Would run on: Consumer GPU to large GPU |
| PTLFlow (library of flow models) | With conditionsEvery bundled model keeps its own licence. | Try many flow models behind one interface.Would run on: Consumer GPU |
| TAPIR and TAPNext (point tracking)Unconfirmed | With conditionsWeights licence not found. | Follows chosen points to measure vibration or drift.Would run on: Consumer GPU |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| CoTrackerAvoid | Not rated | CC BY-NC 4.0, non-commercial. |
Reading text in a scene and documents14 models · Not in the platform yetStart herePaddleOCR PP-OCRv5 (detection and recognition)Allowed
What it does: Reads wagon numbers, signs, displays and scanned forms. Needs: Text detector and reader on a crop; check-digit rules. Main risk: Confident wrong digits; Sinhala reading is weak.
| Model | Commercial use | Good to know |
|---|---|---|
| PaddleOCR PP-OCRv5 (detection and recognition)Start here | Allowed | Wagon, container and sign text in a crop; English and Tamil (94.2 on Baidu's own Tamil set, vendor figure). No Sinhala recogniser.Would run on: CPU (mobile); consumer GPU (server) |
| PaddleOCR-VL (0.9B) and PaddleOCR-VL-1.5 | Allowed | Whole-page forms and tables. Card says 109 languages but names neither Sinhala nor Tamil; test both.Would run on: Consumer GPU |
| Tesseract 5 with tessdata_best (sin, tam, eng) | Allowed | The only engine checked with a ready Sinhala model. Weak on photos and low light; Sinhala accuracy unmeasured.Would run on: CPU |
| GLM-OCR (0.9B) | Allowed | Document parsing with tables; small enough for a site GPU. Sinhala and Tamil not read.Would run on: Consumer GPU |
| dots.ocr | Allowed | Layout plus text for many scripts; worth a Sinhala and Tamil test. Use the original rednote-hilab repo.Would run on: Consumer GPU |
| Granite-Docling-258M with Docling | Allowed | PDF and scanned permits to tables on a CPU-only server. English-centric.Would run on: CPU possible; small GPU faster |
| DeepSeek-OCR | Allowed | Long dense documents. Heavier than PaddleOCR-VL; Sinhala and Tamil not read.Would run on: Consumer GPU |
| olmOCR 2 (7B) | Allowed | Scanned pages and forms to clean text; English-centric and heavy.Would run on: Consumer GPU (16 GB class) |
| EasyOCRUnconfirmed | With conditionsNo weights licence stated. | Easy API with a Tamil model; no Sinhala.Would run on: CPU or consumer GPU |
| docTR (Mindee)Unconfirmed | With conditionsPretrained-weight terms not read. | Latin-script documents and receipts only.Would run on: CPU or consumer GPU |
Avoid 4
| Model | Commercial use | Good to know |
|---|---|---|
| Surya and Chandra (Datalab)Avoid | Not rated | Modified OpenRAIL-M weights, free only below a funding or revenue cap; a closed product sold to many customers needs a paid licence. |
| HunyuanOCR (Tencent)Avoid | Not rated | Licence excludes the EU, UK and South Korea and caps users at 100 million a month. |
| Nanonets-OCR (s and 2-3B)Avoid | Not rated | No licence stated and the base model is research-only. |
| Nougat (Meta)Avoid | Not rated | CC BY-NC 4.0 weights; academic PDFs only. |
Barcodes and QR codes4 models · Not in the platform yetStart hereZXing-cppAllowed
What it does: Gate staff scan permits, passes and labels with any camera. Needs: A sharp frame; crop for distant codes. Main risk: Codes can be copied; LGPL needs care.
| Model | Commercial use | Good to know |
|---|---|---|
| ZXing-cppStart here | Allowed | Reads QR, Data Matrix, Code 128, EAN and more, with Python and C++ bindings.Would run on: CPU |
| WeChat QR code detector (OpenCV contrib)Unconfirmed | With conditionsWeights licence not read. | QR only; handles tiny or blurred codes. Built on ZXing.Would run on: CPU |
| ZBar | With conditionsLGPL needs dynamic linking. | Old, fast and widely packaged; use ZXing-cpp first.Would run on: CPU |
| OpenCV QRCodeDetector | Allowed | Used only at kiosk cameras while a visitor's pass is being scanned; no extra install.Would run on: CPU |
Visual anomaly detection against a normal reference6 models · Not in the platform yetStart herePatchCoreWith conditions
What it does: Flags anything that differs from a normal view. Needs: 100 to 500 clean frames per zone; a GPU to train. Main risk: Alert fatigue; benchmarks do not carry over to CCTV.
| Model | Commercial use | Good to know |
|---|---|---|
| PatchCoreStart here | With conditionsImageNet backbone terms are research-oriented; use a DINOv2 backbone. | Few-shot "something differs" on a fixed view, with a heat map. Up to 99.6 percent image AUROC on MVTec AD (not re-read).Would run on: Consumer GPU to build; CPU or GPU to infer |
| Anomalib (framework) | Allowed | Runs PatchCore, EfficientAD, Dinomaly, WinCLIP and AnomalyDINO. Each model inside has its own backbone terms.Would run on: CPU or consumer GPU (OpenVINO export) |
| Dinomaly | Allowed | One model for many zones. Yes only if trained on site frames; do not ship the published checkpoints.Would run on: Consumer GPU |
| EfficientAD | With conditionsImageNet images in training; check the teacher weights. | Very fast inference, suits many streams.Would run on: CPU or consumer GPU |
| WinCLIP and AnomalyDINO (zero or few-shot)Unconfirmed | With conditionsPorts not read; CLIP trained on undisclosed web data. | Few or no normal examples needed, to start. Less precise than PatchCore.Would run on: Consumer GPU |
| RD4AD (reverse distillation) | With conditionsImageNet backbone terms unread. | A classic reconstruction-style baseline. Older.Would run on: Consumer GPU |
Anomaly detection in video without labels5 models · Not in the platform yetStart hereAnomalib (PatchCore, EfficientAD, FastFlow, Dinomaly, STFPM and others)Allowed
What it does: A ranked review list of moments that look different. Needs: Days of normal footage per fixed camera; review screen. Main risk: Many false alarms; it says different, not dangerous.
| Model | Commercial use | Good to know |
|---|---|---|
| Anomalib (PatchCore, EfficientAD, FastFlow, Dinomaly, STFPM and others)Start here | Allowed | "This view looks different" on a still scene. Image-level, not motion-aware. Never use the MVTec-trained checkpoints; record the ImageNet risk decision.Would run on: CPU for tens of images; GPU speeds it; OpenVINO export |
| AI-VAD (in anomalib, video) | With conditionsComponent weight terms unread. | Flags unusual movement against a fitted normal; explainable. Heavy; report per zone, never per person.Would run on: Consumer GPU |
| AnomalyDINO (few-shot, DINOv2 patch features) | Allowed | Works from 1 to 10 normal reference images; quick to try on one zone. Sensitive to camera shift and light.Would run on: Consumer GPU, or CPU for small inputs |
| WinCLIP (zero and few-shot, in anomalib)Unconfirmed | With conditionsCLIP weights licence not read. | Scores from a text description plus a few normal images. English prompts only.Would run on: Consumer GPU |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| Supervised crime and anomaly video models (VadCLIP, Holmes-VAD and others trained on UCF-Crime, XD-Violence)AvoidUnconfirmed | Not rated | "crime" classes do not match plant events and the data is research-only. |
Industrial inspection and condition9 models · Not in the platform yetStart hereAnomalib PatchCore or EfficientAD trained on good partsAllowed
What it does: Watches the plant: stopped conveyor, guard state, lamps, spills. Needs: Fixed cameras and site labels; thermal camera for hot spots. Main risk: A wrong all-clear is worse than no alert.
| Model | Commercial use | Good to know |
|---|---|---|
| Anomalib PatchCore or EfficientAD trained on good partsStart here | Allowed | Surface defects and odd states on a fixed-pose station from 20 to 200 good images. Needs identical pose and lighting.Would run on: CPU to consumer GPU |
| RF-DETR or RF-DETR-Seg fine-tuned on site states (lamp colour, open door, missing guard, spill, jammed belt)Unconfirmed | Allowed | Lamp lit or not, door open, guard missing, liquid on the floor. Needs 100 to 500 labelled images per state.Would run on: Consumer GPU |
| OpenCV optical flow (DIS, Farneback) for conveyor stall and jam | Allowed | Is the belt moving at normal speed? Confirm with a PLC signal; a plain belt has little texture.Would run on: CPU |
| DINOv2 or SigLIP 2 linear probe for door, guard or valve state | Allowed | A two-class check from about 50 images per state. Camera shift ruins it; retrain per camera.Would run on: CPU for small models |
| Florence-2 and PaddleOCR on a cropped display, nameplate or digital gauge | Allowed | Reads digits on a seven-segment display or plate. Yes applies to Florence-2 only; needle dials are not its strength.Would run on: Consumer GPU or CPU |
| SEA-RAFT and RAFT optical flow | With conditionsWeights provenance (Sintel and others) unread. | Dense motion field for jams and spillage when classical flow is too weak. Heavy for a stall check.Would run on: Consumer GPU |
| Thermal datasets and models (Teledyne FLIR ADAS, Roboflow Universe thermal sets)Unconfirmed | With conditionsFLIR terms unread; per-project licences. | Pre-training a thermal person or vehicle detector. For equipment hot spots use the camera's radiometric threshold.Would run on: n/a (dataset) |
Avoid 2
| Model | Commercial use | Good to know |
|---|---|---|
| Analog gauge readers (ETH analog_gauge_reader, hcltech-robotics/gauge_reader)Avoid | Not rated | No code licence, an AGPL dependency and untraceable weights. Build your own on a detector plus keypoints. |
| MVTec AD and VisA anomaly datasetsAvoid | Not rated | MVTec AD is CC BY-NC-SA 4.0 and checkpoints trained on it are non-commercial; VisA is CC BY 4.0 but the row mixes the two, so no is the safe label. |
Small and far objects: slicing0 models
What it does: Nothing here needs a new model or library. The product cuts a large picture into overlapping tiles, runs the same detector on each tile and merges the boxes back (greedy IoU 0.6; people class-agnostic, and people, vehicles, and fire and smoke merged separately). It does this in three places:.
Sound and speech
Sound events11 models · Not in the platform yetStart hereEfficientAT (MobileNetV3 and dynamic MN, AudioSet)With conditions
What it does: Hears alarms, breaking glass, bangs and calls for help. Needs: A microphone per zone; per-site calibration. Main risk: False alarms; audio capture is regulated.
| Model | Commercial use | Good to know |
|---|---|---|
| EfficientAT (MobileNetV3 and dynamic MN, AudioSet)Start here | With conditionsWeights licence not stated; ask the author or retrain. | Small 527-class tagger; best accuracy per CPU cycle here. mAP 43.2 (mn04) to about 48 (paper).Would run on: CPU |
| YAMNetUnconfirmed | With conditionsWeights licence not confirmed. | Easiest start: 521 classes, 0.96 s windows, runs on a Raspberry Pi 3. Older and less accurate; needs TFLite or ONNX conversion.Would run on: CPU |
| AST (MIT/ast-finetuned-audioset-10-10-0.4593) | Allowed | Higher accuracy (mAP 0.459) and a strong base to fine-tune a site alarm. Same AudioSet caveat.Would run on: Consumer GPU |
| PANNs (CNN14 and relatives) | Allowed | Mature; gives frame-level timestamps. mAP 0.431 to 0.439. Attribution needed.Would run on: CPU |
| BEATs (Microsoft)Unconfirmed | With conditionsCheckpoint licence unconfirmed; ask Microsoft. | Top accuracy (mAP about 48.6, paper) and good embeddings for machine sound.Would run on: Consumer GPU |
| PaSSTUnconfirmed | With conditionsWeights licence not confirmed. | Fast transformer tagger, mAP 0.476 by checkpoint name. No updates since 2024-01.Would run on: Consumer GPU |
| CED (mispeech) | Allowed | Distilled for speed; mAP about 0.48 for base. Code licence not confirmed.Would run on: CPU (tiny, mini); GPU (base) |
| LAION CLAP | With conditionsTraining collection mixes source licences. | Type a sound and match audio with no training; good for testing new classes. Less precise than a fine-tuned tagger.Would run on: Consumer GPU |
| Microsoft CLAP | With conditionsSome training sets are non-commercial. | Audio-text retrieval and captioning; repository archived 2024-09.Would run on: Consumer GPU |
| Qwen2-Audio-7B-Instruct | Allowed | Ask questions about a flagged clip; too heavy for always-on use and answers can be invented.Would run on: Large GPU (about 16 GB in half precision) |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| Community AST fine-tunes for gunshotAvoidUnconfirmed | Not rated | 0 to 14 downloads each, no false-alarm rate, undocumented data. Train your own on site recordings. |
Speech to text17 models · Not in the platform yetStart hereWhisper (large-v3, large-v3-turbo, medium, small, base, tiny)Allowed
What it does: Spoken notes become searchable text on the incident log. Needs: Push-to-talk button, ASR on the server, site tests. Main risk: Push-to-talk only; Sinhala is the weak spot; invented text on noise.
| Model | Commercial use | Good to know |
|---|---|---|
| Whisper (large-v3, large-v3-turbo, medium, small, base, tiny)Start here | Allowed | Mature English and radio-quality notes. Tamil is weaker; Sinhala is unreliable out of the box. Run voice-activity gating first.Would run on: CPU (tiny to small); consumer GPU (large) |
| faster-whisper and whisper.cpp (runtimes)Unconfirmed | Allowed | Runs Whisper several times faster in less memory. Licence files not re-read.Would run on: CPU or consumer GPU |
| Omnilingual ASR (Meta) | Allowed | The only permissive model found that lists Sinhala and Tamil. 7B CER: Sinhala 7.2, Tamil 5.2 (researcher's figure; CER is not WER). Needs fairseq2.Would run on: Consumer GPU (300M to 1B); large GPU (7B) |
| Moonshine (Useful Sensors) | Allowed | Very small English streaming ASR for a kiosk. Do not use the non-commercial language models.Would run on: CPU |
| NVIDIA Parakeet TDT 0.6B and Canary-1B-v2 | Allowed | Best permissive English accuracy and speed. No Sinhala or Tamil. Needs NeMo; attribution required.Would run on: Consumer GPU |
| AI4Bharat IndicConformer 600M | With conditionsGated download; training-data terms not read. | Strong Tamil candidate. Sinhala is not covered.Would run on: Consumer GPU |
| sherpa-onnx (k2-fsa)Unconfirmed | Allowed | One CPU runtime for ASR, TTS, VAD and keyword spotting. Check each model you load.Would run on: CPU |
| IBM Granite Speech (3.3 8B, 4.1 2B) | Allowed | English-first speech with a clear licence trail.Would run on: Consumer GPU (2B); large GPU (8B) |
| distil-whisper large-v3 | Allowed | English only; about six times faster than large-v3.Would run on: Consumer GPU or fast CPU |
| Kyutai STT (stt-2.6b-en, stt-1b-en_fr) | Allowed | Low-latency streaming English and French.Would run on: Consumer GPU |
| Qwen3-ASR (0.6B, 1.7B) | Allowed | 30 languages; no Tamil or Sinhala on the card. A community Sinhala fine-tune exists, quality and licence unknown.Would run on: Consumer GPU |
| Mistral Voxtral (Mini 3B, Small 24B) | Allowed | Speech plus questions about audio. No Sinhala or Tamil.Would run on: Consumer GPU (3B); large GPU (24B) |
| Community Sinhala fine-tunes of Whisper and XLS-RUnconfirmed | With conditionsCC BY-SA training data; WER unverified. | Shows Sinhala fine-tuning is feasible. Fine-tune your own on data you hold.Would run on: CPU (small); consumer GPU (large) |
| VoskUnconfirmed | With conditionsWeights licence not read. | Small offline recognisers on weak CPUs. Tamil or Sinhala models not confirmed.Would run on: CPU |
| SenseVoice Small (Alibaba FunASR) | With conditionsCustom weights licence. | Fast Chinese, English, Japanese and Korean. No Sinhala or Tamil; switch off its emotion and event tags.Would run on: CPU or consumer GPU |
Avoid 2
| Model | Commercial use | Good to know |
|---|---|---|
| Meta MMS (mms-1b-all)Avoid | Not rated | CC-BY-NC-4.0 weights. Covers Sinhala and Tamil, so use only as a benchmark to beat. |
| SeamlessM4T v2 largeAvoid | Not rated | CC-BY-NC-4.0, non-commercial. |
Text to speech13 models · Not in the platform yetStart hereKokoro-82MAllowed
What it does: The kiosk or a site speaker says a message aloud. Needs: Speaker or paging unit; recorded phrases for fixed alarms. Main risk: No good licensed Sinhala voice; no voice cloning.
| Model | Commercial use | Good to know |
|---|---|---|
| Kokoro-82MStart here | Allowed | Small, fast, natural English voice. No Sinhala or Tamil. Data includes audio from closed TTS models.Would run on: CPU |
| Indic Parler-TTS (AI4Bharat) | With conditionsGated; training-data terms not read. | Tamil (voices Kavitha and Jaya) and English. No Sinhala. Takes seconds, so pre-render fixed messages.Would run on: Consumer GPU (CPU slow) |
| Piper | With conditionsActive code is GPL-3.0; voice licences differ. | Very fast; 38 English voices; community Sinhala voices exist (OpenSLR 30, CC BY-SA). No official Sinhala or Tamil voice. Check the GPL terms.Would run on: CPU |
| Parler-TTS Mini v1 | Allowed | English voices with controllable tone.Would run on: Consumer GPU |
| Chatterbox (Resemble AI) | Allowed | Very natural English. Leave voice cloning disabled.Would run on: Consumer GPU |
| Orpheus TTS 3B | With conditionsBuilt on Llama 3.2; the Llama licence may bind. | Expressive English.Would run on: Consumer GPU |
| Dia 1.6B and VibeVoice 1.5B | Allowed | Long multi-speaker English dialogue; not needed for alarms. Can make convincing fake voices.Would run on: Consumer GPU |
| espeak-ngUnconfirmed | With conditionsGPL-3.0; call as a separate process and read the licence first. | Last-resort robotic Sinhala and Tamil for variable words.Would run on: CPU |
| Community Sinhala TTS (VITS and Piper on OpenSLR 30)Unconfirmed | With conditionsCC BY-SA data; unknown quality and voice consent. | Shows Sinhala neural TTS exists; hobby quality. Better to commission your own recordings.Would run on: CPU |
Avoid 4
| Model | Commercial use | Good to know |
|---|---|---|
| Meta MMS-TTS (Tamil, Sinhala)Avoid | Not rated | CC-BY-NC-4.0 weights; benchmark only. |
| Coqui XTTS-v2Avoid | Not rated | Coqui Public Model License, non-commercial; the company stopped. |
| F5-TTS and Fish OpenAudio S1-miniAvoid | Not rated | CC-BY-NC weights, trained on Emilia (non-commercial). |
| OpenVoice (voice cloning)Avoid | Not rated | MIT and commercial use is allowed, but avoid on purpose. Cloning a supervisor's voice is impersonation and a way to fake an emergency order. |
Voice activity and wake-word7 models · Not in the platform yetStart hereSilero VADAllowed
What it does: A hands-free kiosk and a speech gate. Needs: A kiosk microphone; a trained wake word. Main risk: Always-on microphone; shipped wake-word models are non-commercial.
| Model | Commercial use | Good to know |
|---|---|---|
| Silero VADStart here | Allowed | Detects speech, not words, so it works for Sinhala and Tamil. Under 1 ms per chunk on one CPU thread (README). The separate silero-models repo has an unclear licence.Would run on: CPU |
| openWakeWord | With conditionsShipped models are non-commercial; use only the training pipeline. | Train your own wake word, even in Sinhala or Tamil. 15 to 20 models on one Pi 3 core (README). Check the licences of the data you feed it.Would run on: CPU |
| WebRTC VAD (py-webrtcvad)Unconfirmed | Allowed | Cheapest speech gate, no weights. Noisier than Silero on a loud floor.Would run on: CPU |
| microWakeWordUnconfirmed | With conditionsPre-trained model terms unconfirmed. | Tiny wake-word models for microcontrollers.Would run on: CPU (microcontroller) |
| TEN VAD (Agora) | With conditionsMay not compete with Agora's offerings. | Very low latency. The competition clause is vague for a white-label product; Silero does the job.Would run on: CPU |
| Picovoice PorcupineUnconfirmed | With conditionsPaid plan needed; key may phone home. | Best wake-word accuracy, but a poor fit for an offline site.Would run on: CPU |
| pyannote segmentation-3.0 | With conditionsGated; training-corpus terms unclear. | Voice activity and overlap detection inside a recording; Silero is simpler.Would run on: CPU |
Machine sound anomaly detection3 models · Not in the platform yetStart hereDasheng (mispeech) audio encoderAllowed
What it does: Flags a machine that sounds different from its normal. Needs: Close microphone and weeks of normal audio per machine. Main risk: Per-machine tuning; alarms on every shift change.
| Model | Commercial use | Good to know |
|---|---|---|
| Dasheng (mispeech) audio encoderStart here | Allowed | General embeddings; measure distance from each machine's normal. No published machine-fault accuracy found.Would run on: Consumer GPU (CPU works) |
| EfficientAT or BEATs embedding with a per-machine outlier scoreUnconfirmed | With conditionsWeights licences of both unconfirmed. | The sound-event models used as frozen feature extractors. Needs per-machine thresholds.Would run on: CPU |
| MIMII dataset and DCASE baselinesUnconfirmed | With conditionsShare-alike; record your own machines. | Public normal and faulty valves, pumps, fans and rails. Test a pipeline with it, do not ship it.Would run on: n/a |
Search and language
Video search and embeddings20 modelsStart hereSigLIP 2 (google/siglip2-*)Allowed
What it does: Turns pictures into numbers, so recorded video can be searched by words or by example.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| SigLIP 2 (google/siglip2-*)Start here | Allowed | Not yetNeeds its own integration. | Default text-to-image search over sampled frames; the base model runs on CPU. |
| SigLIP (original, google/siglip-*)Start here | Allowed | Not yetNeeds its own integration. | Mature, widely exported predecessor of SigLIP 2; English text only. |
| Perception Encoder PE-Core (Meta) | Allowed | Not yetNeeds its own integration. | Best-tested option for clips and actions; needs Meta's perception_models code; do not confuse with Perception-LM. |
| OpenAI CLIP | Allowed | Not yetNeeds its own integration. | Cheapest baseline (B/32 runs on CPU); older and weaker than SigLIP 2, English only. |
| OpenCLIP LAION models | Allowed | Not yetNeeds its own integration. | Open recipe and many sizes if you fine-tune CLIP on site footage. |
| Qwen3-VL-Embedding (2B, 8B) | Allowed | Not yetNeeds its own integration. | One space for text, frames and short clips; heavy, use it on candidate clips. |
| DINOv2 (search by example image) | Allowed | Not yetNeeds its own integration. | Image-to-image lookups (find this vehicle or person again); no text encoder. |
| bge-m3 and multilingual-e5-large (text embeddings) | Allowed | Not yetNeeds its own integration. | Multilingual text embeddings for event notes and plain-words questions. |
| EVA-CLIP | Allowed | Not yetNeeds its own integration. | Accuracy research; the file is 30.1 GB, too big for a site server. |
| TinyCLIP | Allowed | Not yetNeeds its own integration. | For CPU-only sites on a tight budget; accuracy clearly below SigLIP 2. |
| jina-clip-v1 | Allowed | Not yetNeeds its own integration. | Text-to-text and text-to-image in one model with an ONNX export; do not confuse with v2. |
| nomic-embed-vision-v1.5 | Allowed | Not yetNeeds its own integration. | Shares a vector space with nomic's text embeddings; smaller community. |
| GME-Qwen2-VL-2B | Allowed | Not yetNeeds its own integration. | Superseded by Qwen3-VL-Embedding. |
| LongCLIP-L | With conditionsBuilt on OpenAI CLIP; training data not checked. | Not yetNeeds its own integration. | Long, detailed text queries (about 248 tokens); the file is 1.7 GB. |
| X-CLIP (microsoft/xclip-*) | With conditionsTrained on Kinetics (YouTube clips, mixed rights). | Not yetNeeds its own integration. | Video clip to text matching on 8-frame inputs; built for action labels. |
| InternVideo2-CLIP | With conditionsGated download; web video training data. | Not yetNeeds its own integration. | Native video-text retrieval research; heavy for live use. |
Avoid 4
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| jina-clip-v2Avoid | Not rated | Not rated | CC BY-NC 4.0; commercial download needs contacting Jina sales. v1 is Apache-2.0. |
| MobileCLIP, MobileCLIP2 and DFN5B-CLIP (Apple)Avoid | Not rated | Not rated | Apple AMLR, research purposes only; tempting because it is fast on phones. |
| MetaCLIP and MetaCLIP 2 (facebook)Avoid | Not rated | Not rated | CC BY-NC 4.0. |
| jina-embeddings-v4 and nomic-embed-multimodal-3bAvoid | Not rated | Not rated | Qwen Research License, non-commercial, inherited from Qwen2.5-VL-3B. |
Vision-language models24 modelsStart hereQwen3-VL (2B, 4B, 8B, 32B, 30B-A3B, 235B-A22B)Allowed
What it does: Answers questions about a picture or a clip in plain words.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| Qwen3-VL (2B, 4B, 8B, 32B, 30B-A3B, 235B-A22B)Start here | Allowed | Not yetNeeds its own integration. | First model to try for event descriptions; 4B Q4 fits a 6 GB GPU, 8B Q4 an 8 to 12 GB GPU (file sizes only). |
| Gemma 4 (E2B, E4B, 12B, 26B-A4B, 31B)Licence-safe alternative | Allowed | Not yetNeeds its own integration. | Strong small model with an official 4-bit build; download Gemma 4, not 3. |
| Qwen3.5 (0.8B, 2B, 4B, 9B, 27B, 35B-A3B, 122B-A10B, 397B-A17B) | Allowed | Not yetNeeds its own integration. | Newer natively multimodal line; no official GGUF repo exists, check runtime support. |
| SmolVLM2 (256M, 500M, 2.2B) and SmolVLM | Allowed | Not yetNeeds its own integration. | Smallest video-capable model, runs on CPU; machine-generated caption terms are mixed. |
| MiniCPM-V 4.6 and 4.5 | Allowed | Not yetNeeds its own integration. | Image and video understanding small enough for a CPU box; only the 4.x weights were checked. |
| InternVL3.5 (1B to 38B) | Allowed | Not yetNeeds its own integration. | Alternative to Qwen3-VL at 2B to 8B; check the exact repo's licence. |
| Ministral 3 (3B, 8B, 14B) | Allowed | Not yetNeeds its own integration. | Image understanding plus chat in one model; image only, no video. |
| Qwen3.6 (27B, 35B-A3B) | Allowed | Not yetNeeds its own integration. | Strongest Apache option if the site has a 24 GB+ GPU; too big for small boxes. |
| Molmo2 (4B, 8B, O-7B) | Allowed | Not yetNeeds its own integration. | Pointing at and tracking objects in video; no official GGUF. |
| Florence-2 (base, large) | Allowed | Not yetNeeds its own integration. | Cheap captions, grounding and OCR on a still frame; not conversational, no video. |
| Phi-4-multimodal-instruct | Allowed | Not yetNeeds its own integration. | MIT image plus audio plus text; image only, weaker than Qwen3-VL at the same size. |
| Mistral Small 3.2 24B and Pixtral 12B | Allowed | Not yetNeeds its own integration. | Larger Apache vision model for a 24 GB GPU; Pixtral is older and was not re-read. |
| Granite Vision 4.1 4B | Allowed | Not yetNeeds its own integration. | For tables, forms and charts, not scenes. |
| GLM-4.6V-Flash and GLM-4.1V-9B-Thinking | Allowed | Not yetNeeds its own integration. | MIT at 9B for a 16 GB GPU; thinking variants write long reasoning first. |
| Moondream 2 | Allowed | Not yetNeeds its own integration. | Tiny captioner and pointer; older and weaker than newer small models. |
| Moondream 3 Preview | With conditionsMust not compete with M87 Labs' paid offerings; check local rules. | Not yetNeeds its own integration. | A CCTV analytics product is probably not competing, but embedding it in a paid product is the clause to check. |
| LFM2-VL and LFM2.5-VL (Liquid AI) | With conditionsRevenue threshold of 10 million US dollars. | Not yetNeeds its own integration. | Very efficient on CPU; a revenue cap is a poor fit for a reseller. |
| NVIDIA Cosmos-Reason1 and Cosmos-Reason2 | With conditionsRead the full agreement; Cosmos-Reason2-8B is gated. | Not yetNeeds its own integration. | Reasoning about what happens in a video clip (safety events). |
| Qwen2.5-VL 7B and 32B (Apache) versus 3B and 72B | Allowed | Not yetNeeds its own integration. | Older Apache option; Qwen3-VL replaces it. |
| DeepSeek-VL2 (tiny, small, base) | With conditionsUse-based restrictions (Attachment A) must be passed on. | Not yetNeeds its own integration. | Little reason to pick it over Apache models; the record shows license "other". |
Avoid 4
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| Gemma 3, Gemma 3n and PaliGemma 2Avoid | Not rated | Not rated | Gemma Terms of Use; Google may restrict usage remotely, awkward for an offline site. Use Gemma 4. |
| Llama 3.2 Vision (11B, 90B) and Llama 4 Scout and MaverickAvoid | Not rated | Not rated | The 3.2 licence withholds multimodal rights from EU-based persons and companies. |
| Qwen2.5-VL 3B (and 72B)Avoid | Not rated | Not rated | The 3B is under the Qwen Research License, non-commercial; the 72B licence text was not read. |
| Meta Perception-LMAvoid | Not rated | Not rated | FAIR Noncommercial Research License; easy to confuse with PE-Core. |
Language models20 modelsStart hereQwen3 (0.6B to 32B and 30B-A3B)Allowed
What it does: Writes summaries and answers questions about the site's records.
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| Qwen3 (0.6B to 32B and 30B-A3B)Start here | Allowed | Not yetNeeds its own integration. | Plain-words questions and shift summaries; official GGUF builds exist. |
| gpt-oss-20b and gpt-oss-120b | Allowed | Not yetNeeds its own integration. | Strong reasoning on a single 16 GB GPU; replies carry thinking tokens; 120b not re-read. |
| Qwen3.5 and Qwen3.6 text use | Allowed | Not yetNeeds its own integration. | Newest Qwen generation; no official GGUF repos. |
| Gemma 4 (text use) | Allowed | Not yetNeeds its own integration. | Good multilingual chat at 4B to 12B with official 4-bit builds. |
| Phi-4 and Phi-4-mini | Allowed | Not yetNeeds its own integration. | Small MIT model for summaries and structured output; English-centric. |
| IBM Granite 4.x (micro, h-small, 3B, 8B, 30B) | Allowed | Not yetNeeds its own integration. | Enterprise-minded model for tool use and structured answers; smaller community. |
| Ministral 3 (3B, 8B, 14B) | Allowed | Not yetNeeds its own integration. | European-built, strong European language support; the older 2410 release is research-only. |
| Qwen3-Embedding (0.6B, 4B, 8B) | Allowed | Not yetNeeds its own integration. | Text embeddings so questions can find event descriptions; a retrieval model, not a chat model. |
| OLMo 3 (7B, 32B) | Allowed | Not yetNeeds its own integration. | For buyers who want fully documented training data; behind Qwen3 on benchmarks. |
| SmolLM3-3B | Allowed | Not yetNeeds its own integration. | Fully open 3B model for CPU-only sites; six languages. |
| GLM-4-9B-0414 | Allowed | Not yetNeeds its own integration. | MIT 9B with Chinese and English strength. |
| Qwen2.5 (7B and 14B Apache) | Allowed | Not yetNeeds its own integration. | Proven older model, widely quantised; Qwen3 supersedes it. |
| DeepSeek-R1-Distill-Qwen-7B | Allowed | Not yetNeeds its own integration. | Math-style reasoning; long chain-of-thought output makes it slow. |
| Meta Llama 3.1, 3.2 and 3.3 text models | With conditions"Built with Llama" attribution and Meta's terms. | Not yetNeeds its own integration. | Only where the buyer already approved Meta's terms; Apache alternatives are as good. |
| Falcon3 and Falcon-H1 (TII) | With conditionsTII licence text not read. | Not yetNeeds its own integration. | Nothing the Apache models do not cover. |
| NVIDIA Nemotron Nano and Nemotron 3 NanoUnconfirmed | With conditionsNemotron 3 licence text not read. | Not yetNeeds its own integration. | Efficient hybrid models if NVIDIA's terms are accepted. |
| Liquid LFM2 and LFM2.5 text models | With conditionsRevenue threshold of 10 million US dollars. | Not yetNeeds its own integration. | Very fast on CPU; a revenue cap suits a reseller poorly. |
Avoid 3
| Model | Commercial use | Loads today | Good to know |
|---|---|---|---|
| Ministral-8B-Instruct-2410Avoid | Not rated | Not rated | Mistral Research License, research only. Use Ministral 3. |
| Qwen2.5-3B-InstructAvoid | Not rated | Not rated | Qwen Research License; the 7B and 14B of the same family are Apache-2.0. |
| Cohere Command R7B and Aya ExpanseAvoid | Not rated | Not rated | CC BY-NC 4.0 (Aya Expanse not re-read). |
Embeddings and re-rankers15 models · Not in the platform yetStart hereQwen3-Reranker (0.6B, 4B, 8B)Allowed
What it does: Search notes, reports and manuals by meaning. Needs: Vector store, the text, small GPU or fast CPU. Main risk: Confident wrong results; Sinhala and Tamil retrieval unmeasured.
| Model | Commercial use | Good to know |
|---|---|---|
| Qwen3-Reranker (0.6B, 4B, 8B)Start here | Allowed | Best-scoring open Apache re-ranker in the card table (MTEB-R 65.8 for 0.6B). Sinhala and Tamil not benchmarked.Would run on: Consumer GPU; 0.6B on CPU for small lists |
| bge-reranker-v2-m3 | Allowed | Pairs with bge-m3, already named in Part 1. Behind Qwen3-Reranker (MTEB-R 57.0).Would run on: CPU for short lists; consumer GPU |
| gte-multilingual-reranker-base | Allowed | Small re-ranker for CPU-only sites; 70+ languages.Would run on: CPU |
| mxbai-rerank-base-v2 | Allowed | Strong English re-ranking in a small footprint; weaker outside English and Chinese.Would run on: Consumer GPU |
| Snowflake arctic-embed-l-v2.0 | Allowed | Multilingual retrieval embeddings; alternative to bge-m3. Sinhala and Tamil coverage not read.Would run on: Consumer GPU or CPU |
| gte-multilingual-base | Allowed | Small CPU embedding model with long context; 70+ languages claimed.Would run on: CPU |
| nomic-embed-text-v2-moe | Allowed | Multilingual embeddings in the same family as the nomic vision model in Part 1.Would run on: CPU or consumer GPU |
| LaBSE | Allowed | Matches a Sinhala note to an English manual line. Weaker as a question-to-passage retriever.Would run on: CPU |
| paraphrase-multilingual-MiniLM-L12-v2 | Allowed | Cheapest multilingual option for duplicate-note detection. Older; may omit Sinhala.Would run on: CPU |
| granite-embedding-278m-multilingual | Allowed | Clean provenance for English; Sinhala and Tamil are outside its 12 languages.Would run on: CPU |
| cross-encoder/ms-marco-MiniLM (L6, L12)Unconfirmed | With conditionsTrained on MS MARCO, whose data terms are not cleared. | Very fast English re-ranker. English only.Would run on: CPU |
Avoid 4
| Model | Commercial use | Good to know |
|---|---|---|
| ColPali v1.3 and ColQwen2.5AvoidUnconfirmed | Not rated | MIT adapters on PaliGemma or Qwen2.5-VL-3B backbones with restrictive terms; get exact terms before any use. |
| bge-reranker-v2-gemmaAvoid | Not rated | Gemma base brings the Gemma Terms; the Qwen3 re-ranker does the same job. |
| Jina re-rankers (v2, v3) and jina-embeddings-v3Avoid | Not rated | CC BY-NC 4.0. |
| EmbeddingGemma-300mAvoid | Not rated | Gemma Terms of Use with manual approval; does not suit an offline product. |
Translation12 models · Not in the platform yetStart hereMADLAD-400 (3B, 7B, 10B MT)Allowed
What it does: Alerts, reports and notes in English, Sinhala and Tamil. Needs: 8 GB+ GPU or 8-bit CPU; native-speaker review. Main risk: Sinhala quality unmeasured; best-known models are non-commercial.
| Model | Commercial use | Good to know |
|---|---|---|
| MADLAD-400 (3B, 7B, 10B MT)Start here | Allowed | Sinhala and Tamil are among its languages (not re-read for the exact pair). Card says research users and not assessed for production.Would run on: Consumer GPU; 3B on CPU slowly |
| Qwen3 and Gemma 4 prompted as translators | Allowed | Free-text translation with a glossary in the prompt; nothing extra to ship. Qwen3 names Sinhala among 119 languages; no score read.Would run on: Consumer GPU, already planned for the local language model |
| IndicTrans2 (en-indic 1B and 200M, indic-en 200M) | With conditionsGated download; training-data terms not read. | Strong Tamil to and from English. The card does not mention Sinhala.Would run on: CPU (200M); consumer GPU (1B) |
| M2M-100 (418M, 1.2B) and SMaLL-100 | Allowed | Permissive baseline for Sinhala and Tamil on small hardware. Older; test before use.Would run on: CPU (418M) or consumer GPU |
| CTranslate2 (inference engine) | Allowed | Runs translation models in 8-bit on CPU; makes MADLAD 3B practical. The model's own licence governs.Would run on: CPU or GPU |
| mBART-50 many-to-manyUnconfirmed | With conditionsWeights licence not stated. | Another candidate with both languages.Would run on: Consumer GPU |
| Opus-MT (Helsinki-NLP) multilingual models | With conditionsMixed OPUS training-data terms. | Tiny CPU translators; no dedicated Sinhala or Tamil model found, and multilingual ones are weak.Would run on: CPU |
Avoid 5
| Model | Commercial use | Good to know |
|---|---|---|
| NLLB-200Avoid | Not rated | CC BY-NC 4.0 weights. The tempting choice; use MADLAD-400. |
| SeamlessM4T v2Avoid | Not rated | CC BY-NC 4.0. |
| TranslateGemma (4B, 12B)Avoid | Not rated | Gemma Terms of Use with manual access approval; awkward for an offline product. |
| Sarvam-Translate, Tower+, Tiny AyaAvoidUnconfirmed | Not rated | Sarvam is GPL-3.0 on a Gemma base; the others are tagged non-commercial (not re-read). |
| LibreTranslate serverAvoid | Not rated | AGPL-3.0 reaches a closed product that serves it to users. |
Text classification and extraction12 models · Not in the platform yetStart hereGLiNER v2.1 (multi, large)Allowed
What it does: Types, severity, places and names pulled from notes. Needs: A few hundred labelled notes; spot-checked redaction. Main risk: Missed names; never label named workers.
| Model | Commercial use | Good to know |
|---|---|---|
| GLiNER v2.1 (multi, large)Start here | Allowed | Pulls names, dates, permit numbers and places with labels you choose at run time. Pick v2.1, not v0 or v1.Would run on: CPU |
| GLiNER PII models (urchade multi_pii-v1, knowledgator gliner-pii-base-v1.0) | Allowed | Masks names, phones, IDs and addresses; ONNX builds exist. English-first; spot-check a sample.Would run on: CPU |
| NuExtract 2.0 (2B and 8B) | Allowed | Fills a JSON template from a note, form or photo. Never use the 4B (Qwen Research License).Would run on: Consumer GPU |
| Microsoft Presidio | Allowed | Finds and redacts personal data; add patterns for national ID and phone formats. No Sinhala or Tamil recognisers.Would run on: CPU |
| MoritzLaurer zero-shot NLI classifiers (deberta-v3-large-zeroshot-v2.0, bge-m3-zeroshot-v2.0)Unconfirmed | With conditionsTraining-data terms vary. | Sorts a note into "near miss", "injury", "equipment" with no labels. The bge-m3 version is multilingual.Would run on: CPU or small GPU |
| XLM-RoBERTa (base, large) as a fine-tuning base | Allowed | Fine-tune on a few hundred labelled notes for a reliable incident-type classifier.Would run on: CPU or consumer GPU |
| spaCy English pipelinesUnconfirmed | With conditionsPipeline weights and data terms not read. | Fast English entities and rules. No Sinhala or Tamil pipeline checked.Would run on: CPU |
| IndicBERTv2 and SinBERT | Allowed | Fine-tuning bases for Tamil and Sinhala. SinBERT is from 2022 and news-domain.Would run on: CPU |
| mmBERT-base | Allowed | Newer multilingual encoder to fine-tune; Sinhala and Tamil coverage not read.Would run on: CPU |
Avoid 3
| Model | Commercial use | Good to know |
|---|---|---|
| piiranha-v1Avoid | Not rated | CC BY-NC-ND 4.0; non-commercial and no derivatives. |
| GLiNER v0 (gliner_base, gliner_multi) and NVIDIA gliner-PIIAvoid | Not rated | V0 is CC BY-NC 4.0; the NVIDIA licence text was not read. Use the Apache v2.1 or knowledgator models. |
| dslim/bert-base-NERAvoidUnconfirmed | Not rated | Trained on CoNLL-2003, whose data terms are unclear. |
Language identification3 models · Not in the platform yetStart hereGlotLID (v3)Allowed
What it does: Routes a note to the right translator or index. Needs: A short text sample; script checks cover most cases. Main risk: Romanised Sinhala and Tamil; short strings.
| Model | Commercial use | Good to know |
|---|---|---|
| GlotLID (v3)Start here | Allowed | Covers Sinhala and Tamil among 1,000+ languages. Keep the copyright notice on training data in the product notices.Would run on: CPU |
| fastText lid.176 | With conditionsShare-alike on the model file; legal review. | 176 languages including si and ta.Would run on: CPU |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| Facebook fasttext-language-identification (NLLB LID)Avoid | Not rated | CC BY-NC 4.0. |
Tool-calling models14 models · Not in the platform yetStart hereQwen3-Instruct-2507 (4B, 8B, 30B-A3B)Allowed
What it does: Plain-words requests become validated calls on product queries. Needs: Fixed tool catalogue, constrained decoding, a test set. Main risk: Wrong filters give a confident wrong list.
| Model | Commercial use | Good to know |
|---|---|---|
| Qwen3-Instruct-2507 (4B, 8B, 30B-A3B)Start here | Allowed | Start here: strong tool calling at 4B. BFCL-v3 61.9 (4B, card). Sinhala and Tamil quality not measured.Would run on: 4B at 4-bit on CPU or 6 GB GPU; 30B-A3B on 24 GB GPU |
| Gemma 4 (E2B, E4B, 12B, 26B-A4B, 31B) | Allowed | Native function calling plus image, video and audio input. Sinhala and Tamil not confirmed. Gemma 3 and FunctionGemma are not Apache.Would run on: E4B on 8 GB GPU at 4-bit; 26B-A4B on 16 to 24 GB |
| gpt-oss-20b | Allowed | Native function calling with adjustable reasoning; good for several calls in a row. English-first; reasoning adds latency.Would run on: 16 GB consumer GPU |
| IBM Granite 4.0 (Micro 3B, H-Small) | Allowed | Enterprise-tuned tool calling in low memory. Sinhala and Tamil not among its main languages.Would run on: Micro on CPU or 6 GB GPU |
| Mistral Small 3.2 (24B) and Ministral 3 (8B) | Allowed | Native function calling and JSON output. Lists Hindi, Bengali, Nepali, not Sinhala or Tamil.Would run on: 8B on 12 GB GPU; 24B at 4-bit on 24 GB |
| Microsoft Phi-4-mini-instruct | Allowed | Documented tool-calling format. Expect weak Sinhala and Tamil.Would run on: CPU or 6 GB GPU |
| Constrained decoding (XGrammar, Outlines, llama.cpp grammars) | Allowed | Forces output into a JSON schema so a 4B model never returns a malformed call. Still validate times, camera ids and ranges in code.Would run on: CPU |
| SmolLM3-3B and Hermes 4 14B | Allowed | Open-recipe 3B and a 14B tuned for tool use. Fine-tune data licences not read.Would run on: 3B on CPU; 14B on 16 GB GPU at 4-bit |
| Google FunctionGemma (270M) | With conditionsGated; Google may restrict use. | Tiny model to fine-tune for one fixed tool set. Gemma 4 E2B has Apache terms.Would run on: CPU |
| NVIDIA Nemotron Nano 9B v2 | With conditionsGuardrail-bypass termination clause. | Reasoning with tool calling on one GPU. No Sinhala or Tamil.Would run on: 12 to 16 GB GPU at 4-bit |
| Liquid LFM2 (1.2B-Tool, 8B-A1B) | With conditionsRevenue threshold of 10 million US dollars. | Very small and fast. Threshold applies to the licensee and affiliates.Would run on: CPU |
| Llama-based tool tuners (ToolACE-2-Llama-3.1-8B, MeetKai Functionary small v3.2) | With conditionsThe Llama licence still applies. | Mature parallel function calling. Qwen3 and Gemma 4 do the same without it.Would run on: 8B on 12 GB GPU |
Avoid 2
| Model | Commercial use | Good to know |
|---|---|---|
| Katanemo Arch-Function 3BAvoid | Not rated | Research-named licence on a research-licensed Qwen base. |
| Salesforce xLAM-2 and Hammer 2.1 (7B, 3B)Avoid | Not rated | CC-BY-NC-4.0 or Qwen research licence; top of leaderboards but not for sale. |
Omni models9 models · Not in the platform yetStart hereMiniCPM-o 4.5 (9B)Allowed
What it does: One model describes video and sound together. Needs: 24 GB GPU or larger; control of the audio path. Main risk: Audio is the legal problem; no Sinhala or Tamil speech.
| Model | Commercial use | Good to know |
|---|---|---|
| MiniCPM-o 4.5 (9B)Start here | Allowed | Smallest practical omni model: live video and audio in, text out (11 GB int4, 19 GB bf16, card). Disable or gate the audio path.Would run on: 12 GB GPU at int4; 24 GB at bf16 |
| Qwen3-Omni-30B-A3B (Instruct, Thinking, Captioner) | Allowed | Best open sound-plus-video reasoning; Captioner describes audio. Speech: 19 languages, no Sinhala or Tamil.Would run on: Large GPU (80 GB) at bf16; 24 GB at int4 for short clips |
| Gemma 4 E2B / E4B / 12B (audio and video input) | Allowed | Takes text, image, video frames and audio. The audio encoder is built for speech; use video only unless audio is cleared against local rules.Would run on: E4B on 8 GB GPU at 4-bit |
| Qwen2.5-Omni-7B | Allowed | Previous generation; superseded by Qwen3-Omni and MiniCPM-o 4.5.Would run on: 24 GB GPU at bf16 |
| Microsoft Phi-4-multimodal-instruct | Allowed | Compact image, audio and text with tool calling. Images not video; no Sinhala or Tamil.Would run on: 12 GB GPU |
| Ming-flash-omni Preview and Ming-Lite-Omni 1.5 | Allowed | Permissive omni with video, audio and image input. Smaller community; base-model data terms not read.Would run on: Large GPU (Preview) |
| Video-SALMONN 2 and Ola-7BUnconfirmed | With conditionsTraining-set terms unchecked (LLaVA-Video-178K and others). | Research-grade video with audio at 7B. Tag only.Would run on: 24 GB GPU |
Avoid 2
| Model | Commercial use | Good to know |
|---|---|---|
| NVIDIA OmniVinciAvoid | Not rated | The README says NVIDIA OneWay Noncommercial License while the record says apache-2.0; the tag contradicts the card. |
| Baichuan-Omni-1.5 (7B)Avoid | Not rated | Community licence excludes software and cloud service providers; the record says apache-2.0 but the card says otherwise. |
Assistant safety12 models · Not in the platform yetStart hereQwen3Guard (Gen 0.6B / 4B / 8B, Stream 4B)Allowed
What it does: Screens what the assistant is asked, answers and reads. Needs: CPU or small GPU; a written site policy. Main risk: False blocks; most injection filters are English only.
| Model | Commercial use | Good to know |
|---|---|---|
| Qwen3Guard (Gen 0.6B / 4B / 8B, Stream 4B)Start here | Allowed | Three severity levels, 119 languages claimed. Best start for Sinhala, Tamil and English; accuracy in them unmeasured.Would run on: 0.6B on CPU or small GPU; 8B on a 12 GB+ GPU |
| ProtectAI deberta-v3-base-prompt-injection-v2 (and small v2) | Allowed | Fast detector of injected instructions in text the assistant reads. English only: a Sinhala or Tamil injection passes.Would run on: CPU |
| gpt-oss-safeguard (20B, 120B) | Allowed | Reads your own written policy at run time. Slower than a small classifier; competes with the main assistant for GPU memory.Would run on: 16 GB consumer GPU (20B) |
| IBM Granite Guardian (3.3 8B, 3.2 5B, 3.2 3B-A800M, HAP 38M) | Allowed | Jailbreak, groundedness and tool-call hallucination checks. English only.Would run on: HAP on CPU; 8B on a 12 GB+ GPU |
| PIGuard (leolee99) | Allowed | Tuned to reduce over-defence, useful when alert text has words like "stop". English only.Would run on: CPU |
| Guardrail frameworks (NeMo Guardrails, LLM Guard, Guardrails AI) | Allowed | Plumbing for input and output checks and tool allow-lists. A few hundred lines of own code may be enough.Would run on: CPU |
| Meta Llama Guard 4 (12B) and Llama Guard 3 (8B, 1B) | With conditionsGated; Llama licence (700 million users clause, attribution). | Published hazard taxonomy. English only; no Sinhala or Tamil. Qwen3Guard does the same job.Would run on: 1B on CPU; 8B on 12 GB GPU; 12B on 16 GB+ GPU |
| Google ShieldGemma (2B, 9B) and ShieldGemma 2 (4B, images) | With conditionsGated; Google may restrict usage. | Text and image safety classification. English only (secondary).Would run on: 2B on CPU or small GPU; 9B on 12 GB+ GPU |
| NVIDIA Nemotron Safety Guard 8B v3 and Content Safety Reasoning 4B | With conditionsStacks NVIDIA and Llama licences; guardrail-bypass termination. | Multilingual (nine languages) but no Sinhala or Tamil.Would run on: 8B on a 16 GB GPU |
| AI2 WildGuard (7B) | Allowed | Prompt harm, response harm and refusal detection. English; older. Dataset terms not read.Would run on: 12 GB+ consumer GPU at 4-bit |
| Meta Llama Prompt Guard 2 (86M, 22M) | With conditionsLlama terms (700M MAU, attribution); gated. | Small, fast injection classifier. No Sinhala or Tamil; ProtectAI and PIGuard cover the same ground.Would run on: CPU |
| Small injection classifiers (deepset deberta-v3-base-injection, fmops distilbert, jackhhao jailbreak-classifier) | Allowed | Baselines only; small English sets from 2023 to 2024.Would run on: CPU |
Data, privacy and forecasts
Personal-data detection and redaction10 models · Not in the platform yetStart hereMicrosoft Presidio (analyzer, anonymizer, image redactor)Allowed
What it does: Masks faces, plates, screens and personal text before data leaves. Needs: Face and plate detector, masks, Presidio, review queue. Main risk: Missed detections; no Sinhala or Tamil name models.
| Model | Commercial use | Good to know |
|---|---|---|
| Microsoft Presidio (analyzer, anonymizer, image redactor)Start here | Allowed | Framework for text masking with pattern and model recognisers. English out of the box; no Sinhala or Tamil names.Would run on: CPU |
| EgoBlur Gen 2 (face and licence plate)Unconfirmed | With conditionsRead the weights download agreement first. | Detects faces and plates for blurring (about 21 ms on GPU for 2 MP). Recall on tiny faces in wide CCTV views not published.Would run on: Consumer GPU live; CPU offline |
| OpenAI Privacy Filter | Allowed | Token classifier for 8 privacy categories. Weak on Sinhala and Tamil; fixed label set.Would run on: CPU or small GPU |
| GLiNER-PII (Knowledgator base and large) and gliner_multi_pii-v1 | Allowed | Zero-shot entity finder. F1 83.25 percent (large) on a synthetic set, not site logs. No Sinhala or Tamil claim.Would run on: CPU |
| SAM 2.1 (Hiera tiny to large) | With conditionsTrained on SA-1B, whose terms nobody read. | Pixel masks of a person or object tracked through a clip. Needs a prompt box.Would run on: Consumer GPU |
| docTR (text detection and OCR)Unconfirmed | With conditionsWeights licence and data unread. | Finds text on monitors and whiteboards to blank. Sinhala and Tamil not covered.Would run on: CPU or consumer GPU |
| NVIDIA GLiNER-PII | With conditionsNVIDIA licence text not read. | 55+ PII and health categories. English only; the Apache GLiNER models do the same job.Would run on: CPU or small GPU |
| Meta SAM 3 | With conditionsCustom licence with trade-control terms. | Text-prompted masks ("hard hat", "screen").Would run on: Consumer GPU |
| deface (CenterFace) | With conditionsTrained on WIDER FACE, which is CC BY-NC-ND. | Ready-made command-line face blurring; EgoBlur is cleaner.Would run on: CPU |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| Piiranha, ai4privacy and Isotonic PII modelsAvoid | Not rated | CC-BY-NC-ND-4.0 and CC-BY-NC-4.0 weights; popular but not usable in a sold product. |
Privacy masking and redaction5 models · Not in the platform yetStart hereRF-DETR Seg for person and vehicle masksAllowed
What it does: Covers faces, screens and bystanders in exports. Needs: Detectors, solid fill, human review, audit entry. Main risk: A missed face leaks identity; blur can be undone.
| Model | Commercial use | Good to know |
|---|---|---|
| RF-DETR Seg for person and vehicle masksStart here | Allowed | Covers everyone except the person of interest with a solid shape. Misses partly hidden people.Would run on: Consumer GPU |
| SAM 2.1 Tiny or Small (mask tracking through a clip) | Allowed | Click once on a screen or person and carry the mask through the clip (about 85 to 91 FPS per README). A reviewer checks the whole clip.Would run on: Consumer GPU |
| PaddleOCR PP-OCRv5 text detection | Allowed | Finds text on screens, papers and badges to cover. Also finds signs that should stay visible.Would run on: CPU or consumer GPU |
| OpenCV opaque fill, pixelate and blur | Allowed | Applies the mask. Use solid fill; it only applies what detectors found.Would run on: CPU |
| BiRefNet (and lite) | Allowed | Clean foreground cut-out to keep one person and hide the rest. Built for objects, not crowds.Would run on: Consumer GPU |
Synthetic data for training10 models · Not in the platform yetStart hereFLUX.2 [klein] 4BAllowed
What it does: Rare PPE and hazard pictures to widen training sets. Needs: 16 to 24 GB GPU, mask model, human review. Main risk: Wrong PPE geometry; generator provenance.
| Model | Commercial use | Good to know |
|---|---|---|
| FLUX.2 [klein] 4BStart here | Allowed | Text-to-image and editing in one model. Only the 4B is Apache; FLUX.2 [dev] is non-commercial.Would run on: 16 GB consumer GPU (about 13 GB VRAM) |
| Copy-paste compositing with SAM 2.1 masks (Albumentations) | With conditionsThe mask model inherits the SA-1B question; masks can be drawn by hand. | Cut real PPE from site frames and paste into others; labels stay exact. Do not paste a person's face into other scenes.Would run on: CPU for compositing; GPU for masks |
| FLUX.1 [schnell] | Allowed | Four-step text-to-image for scene variety. No editing; dev, Kontext and Fill are non-commercial.Would run on: 24 GB GPU, less when quantised |
| Qwen-Image (20B) and Qwen-Image-Edit-2511 | Allowed | Instruction editing ("put a hard hat on the second worker"). Run offline on a workstation.Would run on: Large GPU; 24 GB quantised |
| Z-Image-Turbo (6B) | Allowed | Fast text-to-image in 8 steps. Editing is a separate variant.Would run on: 16 GB consumer GPU |
| OmniGen2, BAGEL-7B-MoT, Step1X-Edit | Allowed | Unified generation and editing; alternatives to Qwen-Image-Edit. Base-model data chain not traced.Would run on: 24 GB GPU |
| HiDream-I1 and Sana 1.6B | With conditionsSana ships a Gemma text encoder under Gemma terms. | HiDream is high quality; Sana is very fast.Would run on: Sana on 8 GB; HiDream on a large GPU |
| Stable Diffusion XL and 3.5, with ControlNet | With conditionsUse-restriction clauses pass to customers; SD 3.5 free below 1 million dollars revenue. | Largest ecosystem: depth or pose-controlled scenes and mask inpainting. Older than FLUX.2 klein.Would run on: 12 GB consumer GPU (SDXL) |
| NVIDIA Cosmos Transfer 2.5 and Predict 2.5 (2B) | With conditionsRead the full agreement. | Turns a rendered scene into photoreal video with exact labels. Needs a 3D scene source.Would run on: Large GPU |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| FLUX.1 dev / Kontext / Fill, FLUX.2 dev, SDXL-Turbo, BRIA FIBOAvoid | Not rated | Non-commercial licences; outputs may not be cleared for training a product either. |
Image and video generation or editing6 models · Not in the platform yetStart hereFLUX.2 [klein] 4B (editing mode)Allowed
What it does: Clean a floor-plan photo for drawing cameras. Needs: Often no model; a mask and inpainting. Main risk: Altered pictures passed as evidence; invented pixels.
| Model | Commercial use | Good to know |
|---|---|---|
| FLUX.2 [klein] 4B (editing mode)Start here | Allowed | Instruction and multi-reference editing to clean or restyle a plan. Show a before and after.Would run on: 16 GB consumer GPU |
| Qwen-Image-Edit-2511 | Allowed | High-quality edits with fewer unwanted changes. Too heavy for a site server.Would run on: Large GPU |
| LaMa (big-lama) | With conditionsTrained on Places2 (non-commercial research terms); check the terms or retrain. | Paint a mask and LaMa fills it; small and fast.Would run on: CPU |
| Wan 2.2 (TI2V-5B, I2V-A14B) and Wan 2.1 VACE | Allowed | Text-to-video, image-to-video and video editing for offline demos only. Never on real footage in the product.Would run on: 24 GB GPU (5B); large GPU (A14B) |
| Other video generators (LTX-2, HunyuanVideo, CogVideoX-5B, Stable Video Diffusion, Mochi 1, Open-Sora v2) | With conditionsLicences differ; LTX-2 paid above 10 million dollars; Hunyuan excludes EU, UK, South Korea. | Not needed by the product.Would run on: Large GPU |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| ProPainter and DiffuEraser (video inpainting)Avoid | Not rated | ProPainter is S-Lab non-commercial and DiffuEraser needs its weights; altering video of real people also carries the evidence risk. |
Evidence integrity and AI-image labelling2 models · Not in the platform yetStart herec2pa-rs (Content Credentials library)Allowed
What it does: Signed exports that show who exported what and when. Needs: A signing certificate and key storage. Main risk: A signature proves no change, not a real scene.
| Model | Commercial use | Good to know |
|---|---|---|
| c2pa-rs (Content Credentials library)Start here | Allowed | Signs exports with who, when, camera and redaction state, and verifies later. Outside parties trust only C2PA-listed certificates.Would run on: CPU |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| AI-generated-image detectors (Ateeqq, haywoodsloan, umm-maybe, Organika sdxl-detector)Avoid | Not rated | Unreliable on camera pictures, and accusing a person of a faked image on a detector's score is unsafe; Organika is CC-BY-NC-3.0. |
Forecasting and anomaly detection11 models · Not in the platform yetStart hereChronos-2 (and Chronos-Bolt)Allowed
What it does: Forecasts footfall and queues; flags an odd hour. Needs: Weeks of counts, a site calendar. Main risk: Planned changes cause false alarms.
| Model | Commercial use | Good to know |
|---|---|---|
| Chronos-2 (and Chronos-Bolt)Start here | Allowed | Footfall and queue forecasts with day of week and shift plan as covariates; quantile ranges.Would run on: CPU (Bolt, small); consumer GPU |
| statsforecast, Prophet, darts and sktimeUnconfirmed | Allowed | Seasonal-naive, ETS and ARIMA baselines; often as good as foundation models on a stable site. Licences not re-read.Would run on: CPU |
| TimesFM 2.5 (200M) | Allowed | Zero-shot forecasts with long context. Training-data terms not read.Would run on: CPU or consumer GPU |
| Granite TinyTimeMixer (TTM-r2) | Allowed | Right size for a CPU-only site; fine-tunes on a few percent of the data.Would run on: CPU |
| PyOD, scikit-learn Isolation Forest, STUMPY, riverUnconfirmed | Allowed | Flags an unusual hour, odd stretches or drift on counts and camera health. Tune per site.Would run on: CPU |
| Toto (Datadog) | Allowed | Built for server metrics; useful for the product's own health numbers.Would run on: Consumer GPU |
| Sundial and Timer (Tsinghua) | Allowed | Alternative zero-shot forecasters. LOTSA has mixed terms.Would run on: CPU or consumer GPU |
| MOMENT-1 and Time-MoE | Allowed | MOMENT also does anomaly detection. Time-MoE licence not confirmed.Would run on: Consumer GPU |
| Lag-Llama | Allowed | Earlier probabilistic forecaster; behind Chronos-2 and TimesFM.Would run on: CPU |
| TiRex (NXAI) | With conditionsCommercial licence needed above 100 million EUR revenue. | Strong forecaster; the revenue condition scales with the owner and customers.Would run on: Consumer GPU |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| Moirai 1.1, 2.0 and MoE (Salesforce)Avoid | Not rated | CC BY-NC 4.0 weights. |
Risk scoring on tables6 models · Not in the platform yetStart hereLightGBM, XGBoost and CatBoostAllowed
What it does: Ranks zones and shifts by where problems cluster. Needs: Event table, months of history, per-score reason. Main risk: Never per person; small groups single out individuals.
| Model | Commercial use | Good to know |
|---|---|---|
| LightGBM, XGBoost and CatBoostStart hereUnconfirmed | Allowed | Strong default for counts per zone and shift. Needs an explanation layer such as SHAP. Places and shifts only.Would run on: CPU |
| Explainable Boosting Machine (InterpretML) and logistic regression (scikit-learn)Unconfirmed | Allowed | Glass-box models whose effects can be shown to an auditor; easier to defend.Would run on: CPU |
| SHAPUnconfirmed | Allowed | Explains each score from a boosted-tree model.Would run on: CPU |
| TabPFN v2 (Prior Labs) | With conditionsEnhanced attribution requirement. | Pretrained tabular model for small tables with no training. The 2.5 release is non-commercial.Would run on: Consumer GPU for larger tables |
| TabICLUnconfirmed | With conditionsOfficial weights licence not confirmed. | Pretrained tabular classifier; newer, less proven.Would run on: Consumer GPU |
Avoid 1
| Model | Commercial use | Good to know |
|---|---|---|
| TabPFN-2.5Avoid | Not rated | TabPFN-2.5 Non-Commercial License; no production use. |
Words used here
- Model
- A trained program that looks at a picture, a sound or a text and answers a question about it.
- Detector
- A model that draws a box around each thing it finds: a person, a helmet.
- Job
- One thing the platform does with a model: finding people, checking PPE, reading a number plate.
- Size
- How big the model is. Bigger is more careful and slower.
- Outline
- The exact shape of an object, following its edge, not a box around it. Click to outline draws a zone this way from one click.
- Rotated box
- A box turned to hug an object that sits at an angle.
- Body points
- The joints of a person, as points, for checks such as a person who has fallen.
- Whole-picture label
- One label for the whole picture, with no box around anything.
- Fit
- Whether a size runs on the processor (CPU), runs faster on a graphics card, or needs one. A size that needs a GPU warns you on a server without one.
- Heavy model
- A model that takes much of the server's computing power. It starts switched off until you turn it on.
- Type what to find
- For a model that takes typed words, the names of the things to look for. The platform makes a copy that looks only for those names.
- GPU
- A graphics card. It runs models many times faster than a normal processor (CPU).
- Licence
- The rules its makers set for using the model.
- Commercial use
- Using the model in a business. Some licences allow it, some allow it with conditions, some do not.
- Loader
- The part of the platform that knows how to run one kind of model. A model whose kind has no loader yet cannot run.
- Class
- One kind of thing a model can find, such as "helmet" or "no helmet".
- Class map
- The table that tells the platform which of a model's classes mean which PPE item. The platform suggests one when a model is installed.
- Activate
- To make an installed model the one that runs for its job.
- Test bench
- The page where you check photos and stored videos with a model before it runs on live cameras.
- Training
- Teaching a model with labelled pictures. Fine-tuning is training a ready-made model a little more on your own pictures.