google-research/vision_transformer Alternatives

choose when you need ViT/MLP-Mixer or many other backbones in PyTorch with pretrained weights instead of JAX code.

The largest collection of PyTorch image encoders / backbones. Including train, eval, inference, export scripts, and pretrained weights -- ResNet, ResNeXT, EfficientNet, NFNet, Vision Transformer (ViT), MobileNetV4, MobileNet-V3 & V2, RegNet, DPN, CSPNet, Swin Transformer, MaxViT, CoAtNet, ConvNeXt, and more

choose for a maintained JAX research library with broader CV training utilities rather than this minimal fine-tuning repo.

Scenic: A Jax Library for Computer Vision Research and Beyond

choose for the original higher-throughput JAX training/inference framework that produced the seed's checkpoints.

Official codebase used to develop Vision Transformer, SigLIP, MLP-Mixer, LiT and more.

choose for a hybrid Mamba-transformer backbone when you want a more efficient modern alternative to ViT.

[CVPR 2025] Official PyTorch Implementation of MambaVision: A Hybrid Mamba-Transformer Vision Backbone

choose if you want an all-in-one PyTorch CV toolkit rather than a ViT-specific release.

An all-in-one toolkit for computer vision

choose for a concise, self-contained PyTorch ViT implementation without JAX dependency, mainly for experimentation.

Implementation of Vision Transformer, a simple way to achieve SOTA in vision classification with only a single transformer encoder, in Pytorch

choose if you are standardizing on PaddlePaddle and want classification pretrained models/toolkit instead of JAX.

A treasure chest for visual classification and recognition powered by PaddlePaddle

choose for a fast hierarchical ViT variant with improved throughput/accuracy over plain ViT.

[ICLR 2024] Official PyTorch implementation of FasterViT: Fast Vision Transformers with Hierarchical Attention

choose it for a lightweight PyTorch implementation of MLP-Mixer if seed's JAX stack is unwanted.

An All-MLP solution for Vision, from Google AI

choose for a lightweight MetaFormer baseline if you avoid ViT and don't mind an unmaintained repo.

PoolFormer: MetaFormer Is Actually What You Need for Vision (CVPR 2022 Oral)

choose it when you specifically need Swin's hierarchical transformer design in PyTorch; the repo is stable but no longer updated.

This is an official implementation for "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows".

choose if you are in the OpenMMLab ecosystem and need a training toolbox; the repo is unmaintained recently.

OpenMMLab Pre-training Toolbox and Benchmark

choose for Apple's PyTorch CV library with deployment-ready models if you don't mind no recent maintenance.

CVNets: A library for training computer vision networks

choose for an efficient convolution-transformer hybrid backbone in PyTorch if you don't mind an older unmaintained repo.

[ICLR2022] official implementation of UniFormer

choose for simple Swin-like transformer designs in PyTorch if you want a lighter alternative, though it is old and unmaintained.

Two simple and effective designs of vision transformer, which is on par with the Swin transformer

older PyTorch-only MLP-Mixer implementation, useful only if you need a minimal no-dependency Mixer in legacy code.

Unofficial implementation of MLP-Mixer: An all-MLP Architecture for Vision

choose for a tokens-to-token ViT variant that trains from scratch better than plain ViT; PyTorch but older.

ICCV2021, Tokens-to-Token ViT: Training Vision Transformers from Scratch on ImageNet

old PyTorch reimplementation of ViT; only if you need a tiny standalone example, use timm or vit-pytorch instead.

Pytorch reimplementation of the Vision Transformer (An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale)