microsoft/Swin-Transformer Alternatives

Choose it when you want an actively maintained single PyTorch library that includes Swin plus many other pretrained backbones.

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 it over Swin when global-attention ViT and JAX/TPU-style pretraining fit your stack better than hierarchical shifted windows.

Choose it when you prefer a maintained JAX/Flax research library over the PyTorch Swin repo and need a broad set of vision architectures.

Scenic: A Jax Library for Computer Vision Research and Beyond

Choose it for the original big_vision JAX codebase and large-scale ViT/SigLIP checkpoints instead of Swin.

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

Choose it when a hybrid mamba-transformer backbone gives better throughput/accuracy at high resolution than Swin.

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

Choose it when you want a higher-level all-in-one CV toolkit that wires backbones to downstream tasks, rather than assembling Swin yourself.

An all-in-one toolkit for computer vision

Choose it if a simple PoolFormer-style backbone is easier to adapt and you don't need shifted-window engineering, though the repo is old.

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

Choose it when you need a hierarchical ViT with faster factorized attention and good high-resolution scaling for dense prediction.

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

Choose it when you need a transformer with efficient local-global attention for image/video tasks and can settle for older, stably released code.

[ICLR2022] official implementation of UniFormer

Choose it for a tiny, readable, MIT-licensed ViT implementation if standard patch-based global attention beats Swin for your use case.

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

Choose it when your goal is semantic segmentation and you want an encoder/decoder pair designed together rather than a general Swin backbone plus separate head.

Official PyTorch implementation of SegFormer

Choose it when you want a simpler cross-attention vision transformer with official checkpoints competitive with Swin, despite an old unmaintained repo.

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

Choose it when you work in PaddlePaddle and want a maintained classification/backbone model zoo, including transformer variants, instead of the PyTorch official code.

A treasure chest for visual classification and recognition powered by PaddlePaddle

Choose it if you need OpenMMLab config-driven classifier/backbone zoo and accept an unmaintained codebase, but not as a current drop-in for official Swin.

OpenMMLab Pre-training Toolbox and Benchmark

Choose it only if you need Apple's mobile-oriented training pipeline for many CV backbones; it is stale and not a maintained drop-in.

CVNets: A library for training computer vision networks

Choose it for high-resolution CNN-based semantic segmentation if you prefer HRNet over a Swin transformer backbone, but it is a narrower one-task repo.

The OCR approach is rephrased as Segmentation Transformer: https://arxiv.org/abs/1909.11065. This is an official implementation of semantic segmentation for HRNet. https://arxiv.org/abs/1908.07919

Choose it only for experimenting with Bottleneck Transformer blocks; it is an unmaintained research snippet rather than a current backbone library.

Implementation of Bottleneck Transformer in Pytorch