pytorch-image-models

The largest collection of PyTorch image encoders and backbones keeps a dated changelog, and the recent entries are a decent signal of where vision model work is drifting.

ResNet, ResNeXT, EfficientNet, NFNet, ViT, MobileNetV4, MobileNetV3 and V2, RegNet, DPN, CSPNet, Swin, MaxViT, CoAtNet, ConvNeXt: the collection spans old convolutional workhorses and newer transformer designs, all with train, eval, inference, and export scripts plus pretrained weights. Documentation is hosted at huggingface.co, the code is Python under Apache-2.0, and the repo holds around 37,000 stars.

Updates are logged by date, and the August 2026 entry is typical of the recent direction. It adds model definitions and pretrained weights for CPUBone, an efficient vision backbone aimed at devices with weak parallelization capabilities, and improves NaFlexViT. A July entry tuned optimizer support and cut version 1.0.28.

Before that, the log picked up EUPE ViT and ConvNeXt models, TIPSv2 definitions, and Gemma4 ViT encoders. Attention mask handling for ViT and EVA models got fixes, and a patch representation refinement pooling option landed. If you want a quick sense of what image model research produced over a stretch of months, this file is an honest proxy.

One entry deserves its own note. After the author's departure from Hugging Face, the first maintenance release concentrated on pickle checkpoint handling and security, exactly the kind of unglamorous work a collection like this needs to stay trustworthy. Versioned releases with new model support and fixes have continued since.

For anyone doing image work in PyTorch, this repo functions as the reference shelf: when a paper ships a backbone, its definitions and weights tend to show up here. The scripts cover the boring parts too, training, evaluation, inference, and export, so a model from the list can move from download to experiment without much glue code.

• Source 1 · Project repository
• Source 2 · README