pytorch-image-models vs supervision
Two of the top projects in ML frameworks and libraries, side by side: who leads on each score part and who each one suits, then setup, license, activity and what each review found.
pytorch-image-models
PyTorch library of image models, pretrained weights and training scripts
supervision
Python toolkit for loading, annotating and counting computer vision detections
Verdict
pytorch-image-models ranks higher, 93 to 92. Of the 4 score parts we measure for both, pytorch-image-models leads on 1, supervision on 1 and 2 are even.
- Adoptionsupervision leads by 3 points, 94 to 91: close to even.
- FreshnessEven at 100.
- Maintenancepytorch-image-models leads by 7 points, 97 to 90.
- Agent-readyEven at 70.
Pick pytorch-image-models if you are
ML engineers training or using image backbones in PyTorch
- Broad model zoo with pretrained weights for nearly every listed architecture family
- Includes train.py, validate.py, inference.py and sweep.py for reproducible ImageNet-style runs
Pick supervision if you are
Python developers building computer vision apps around detection models
- Connectors for Ultralytics, Transformers, MMDetection and Inference; rfdetr returns sv.Detections directly
- Dataset utils load, split, merge and convert YOLO, Pascal VOC and COCO
| What we compare | pytorch-image-models | supervision |
|---|---|---|
| Score parts, out of 100 | ||
| Adoption | 91, widely used | 94, widely used |
| Freshness | 100, active | 100, active |
| Maintenance | 97, healthy | 90, healthy |
| Agent-ready | 70, partly | 70, partly |
| Facts from GitHub and the README | ||
| Stars | 37.3k | 51.6k |
| License | Apache-2.0 (permissive) | MIT (permissive) |
| Last commit | Nov 2026 | Nov 2026 |
| Last release | Nov 2026 | Nov 2026 |
| Language | Python | Python |
| Docker | No | No |
| GPU | Optional | Optional |
| arm64 or Apple Silicon | Not stated | Not stated |
pytorch-image-models
timm is a Python library of image model architectures (ViT, EVA, ConvNeXt, Swin, EfficientNet and many more) with pretrained weights loaded from the Hugging Face Hub. It also ships optimizers, LR schedulers, data loaders and augmentations, plus reference train, validate, inference and sweep scripts for reproducing ImageNet results. Recent releases add multi-label training, NaFlexViT variable-resolution input, and forward_intermediates support in every model.
Who it is for: ML engineers training or using image backbones in PyTorch
Strengths
- Broad model zoo with pretrained weights for nearly every listed architecture family
- Includes train.py, validate.py, inference.py and sweep.py for reproducible ImageNet-style runs
- forward_intermediates and prune_intermediate_layers supported across all models
- Large optimizer set, including Muon, Kron, LAMB and foreach implementations
Weaknesses
- Image models only; no text or LLM support
- Some model variants ship without pretrained weights
- RoPE and low-precision fixes can change existing model outputs between releases
- Jan 2026 compat break in ParallelScalingBlock bias handling may affect downstream code
- GPU optional
- Needs PyTorch
- Models: PyTorch, Hugging Face Hub weights
supervision
Supervision is a model-agnostic Python library that wraps detection, classification and segmentation outputs in a common sv.Detections object. It has connectors for Ultralytics, Transformers, MMDetection and Inference, annotators for drawing results, and utilities to load, split, merge and convert YOLO, Pascal VOC and COCO datasets. Tracking, zone counting and speed estimation appear in the linked tutorials.
Who it is for: Python developers building computer vision apps around detection models
Strengths
- Connectors for Ultralytics, Transformers, MMDetection and Inference; rfdetr returns sv.Detections directly
- Dataset utils load, split, merge and convert YOLO, Pascal VOC and COCO
- Configurable annotators such as BoxAnnotator for drawing detections on frames
- Installs with pip; docs, cookbooks and a cheatsheet are available
Weaknesses
- Does not train or run models itself; a separate detection model is required
- Requires Python 3.10 or newer
- Inference and dataset download examples need a Roboflow API key
- Tracking and zone features are shown in video tutorials, not the README text
- GPU optional
- Needs Python>=3.10, rfdetr (optional), Roboflow Inference (optional), Roboflow API key (optional)
- Models: Ultralytics, Transformers, MMDetection, Inference, rfdetr
More in ML frameworks and libraries
Building with these?
We help teams pick an open-source AI stack and run it in production. Client work never changes a rank.