huggingface/pytorch-image-models

PyTorch Image Models, known as timm, is the largest open-source collection of image recognition model architectures and pretrained weights for the PyTorch deep learning framework. It solves a practical problem in computer vision research and production: researchers and engineers frequently need to swap between dozens of different neural network architectures for image tasks (classification, feature extraction, object detection backbones), and building each from scratch or hunting across separate repositories is time-consuming and error-prone. The library provides a unified API for loading any supported model, over 1,000 architectures including ResNet, EfficientNet, Vision Transformer (ViT), Swin Transformer, ConvNeXt, MobileNet, and many others, with pretrained weights automatically downloaded from the Hugging Face Hub. You call timm.create_model("resnet50", pretrained=True) and you have a working, weight-loaded model ready for training or inference. The key abstraction is that all models share the same interface for feature extraction, so you can use any architecture as a backbone for downstream tasks like object detection or segmentation without rewriting glue code. The library also ships production-quality training scripts, augmentation pipelines, and a suite of optimizers, making it usable as an end-to-end training toolkit rather than just a model zoo. You would use timm when benchmarking different architectures, fine-tuning a pretrained model on your own dataset, or building a computer vision system that needs a strong image encoder. It is the standard first stop in the computer vision research community for reproducing published results. The tech stack is Python with PyTorch as the only hard dependency, pretrained weights live on the Hugging Face Hub.