Iwin Transformer    

Introduction

Iwin Transformer (the name Iwin stands for Interleaved window) is initially described in arxiv. It is a position-embedding-free hierarchical vision transformer, which can be fine-tuned directly from low to high resolution, through the collaboration of innovative interleaved window attention and depthwise convolution.

Results on ImageNet with Pretrained Models

ImageNet-1K and ImageNet-22K Pretrained Iwin Models

name	pretrain	resolution	acc@1	#params	FLOPs	22K model	1K model
Iwin-T	ImageNet-1K	224x224	82.0	30.2M	4.7G	-	github/config
Iwin-S	ImageNet-1K	224x224	83.4	51.6M	9.0G	-	github/config
Iwin-S	ImageNet-1K	384x384	84.3	51.6M	27.7G	-	github/config
Iwin-S	ImageNet-1K	512x512	84.4	51.6M	52.0G	-	github/config
Iwin-S	ImageNet-1K	1024x1024	83.8	51.6M	207.9G	-	github/config
Iwin-B	ImageNet-1K	224x224	83.5	91.2M	15.9G	-	github/config
Iwin-B	ImageNet-1K	384x384	84.9	91.2M	48.3G	-	github/config
Iwin-B	ImageNet-1K	512x512	85.1	91.3M	89.5G	-	github/config
Iwin-B	ImageNet-1K	1024x1024	85.0	91.3M	358.2G	-	github/config
Iwin-B	ImageNet-22K	224x224	85.5	91.2M	15.9G	github/config	github/config
Iwin-B	ImageNet-22K	384x384	86.6	91.2M	48.3G	-	github/config
Iwin-B	ImageNet-22K	512x512	86.1	91.2M	89.5G	-	github/config
Iwin-B	ImageNet-22K	1024x1024	85.6	91.2M	358.2G	-	github/config
Iwin-L	ImageNet-22K	224x224	86.4	204.3M	35.4G	github/config	github/config
Iwin-L	ImageNet-22K	384x384	87.4	204.3M	106.6G	-	github/config

Results on Downstream Tasks

COCO Object Detection (2017 val)

Backbone	Method	pretrain	Lr Schd	box mAP	mask mAP	#params	FLOPs	model
Iwin-T	Mask R-CNN	ImageNet-1K	1x	42.2	38.9	48M	268G	github
Iwin-S	Mask R-CNN	ImageNet-1K	1x	43.7	40.0	69M	358G	github
Iwin-T	Mask R-CNN	ImageNet-1K	3x	44.7	40.9	48M	268G	github
Iwin-S	Mask R-CNN	ImageNet-1K	3x	45.5	41.0	69M	358G	github
Iwin-T	Cascade Mask R-CNN	ImageNet-1K	1x	47.2	40.9	86M	747G	github
Iwin-T	Cascade Mask R-CNN	ImageNet-1K	3x	49.4	42.9	86M	747G	github
Iwin-S	Cascade Mask R-CNN	ImageNet-1K	3x	49.4	43.0	107M	837G	github

ADE20K Semantic Segmentation (val)

Backbone	Method	pretrain	Crop Size	Lr Schd	mIoU	#params	FLOPs	model
Iwin-T	UPerNet	ImageNet-1K	512x512	160K	44.70	61.9M	946G	github
Iwin-S	UperNet	ImageNet-1K	512x512	160K	47.50	83.2M	1038G	github
Iwin-B	UperNet	ImageNet-1K	512x512	160K	48.90	124.8M	1189G	github

Kinetics 400 Recognition

Kinetics-400 Video Recognition

Backbone	Pretrain	Lr Schd	spatial crop	acc@1	acc@5	#params	FLOPs	config	model
Iwin-T	ImageNet-1K	30ep	224	79.1	93.8	29.8M	74G	config	github
Iwin-S	ImageNet-1K	30ep	224	80.0	94.1	51.1M	140G	config	github

Image generation

We built an image generation model FlashDiT without position embedding using iwin 2D attention. It has less memory usage and faster training speed than Lightningdit.

Tokenizer	Generation Model	FID	FID cfg
VA-VAE	FlashDiT-win8-XL-56ep	5.37	2.30

Language Model

We built a language model "miniwin" based on the iwin's philosophy "no token is an island." See miniwin for more details.

Acknowledgements

This repo is mainly built on Swin. Thanks for the great work.

License

MIT.

Citation

If you find our work helpful, feel free to give us a cite and a star🌟.

@misc{huo2025iwin,
      title={Iwin Transformer: Hierarchical Vision Transformer using Interleaved Windows}, 
      author={Simin Huo and Ning Li},
      year={2025},
      eprint={2507.18405},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.18405}, 
}