 
 
 
This repo provides a simple PyTorch implementation for training and testing LiT, covering both class-conditional (C2I) and text-to-image (T2I) tasks.
LiT: Delving into a Simple Linear Diffusion Transformer for Image Generation
Jiahao Wang†, Ning Kang, Lewei Yao, Mengzhao Chen, Chengyue Wu, Songyang Zhang, Shuchen Xue, Yong Liu, Taiqiang Wu, Xihui Liu, Kaipeng Zhang, Shifeng Zhang, Wenqi Shao, Zhenguo Li†, Ping Luo†,
HKU, Shanghai AI Lab, Huawei Noah’s Ark Lab, UCAS, THUsz
- (✅ New) Oct. 7, 2025. 💥 LiT code and pretrained models is released!!
- (✅ New) June. 26, 2025. 💥 LiT is accepted by ICCV 2025!!
- (✅ New) Jan. 22, 2025. 💥 LiT paper is released!!
LiT is a systematic solution that helps you quickly and safely convert a pretrained DiT into a linear DiT.
LiT provides five practical guidelines, serving as building blocks that offer empirical insights to the community.
Guideline 1: Simply adding a depth-wise convolution in linear attention is sufficient for DiT-based image generation.
Guideline 2: Using few heads in the linear attention increases computation but not latency.
Guideline 3: Linear diffusion Transformer should be initialized from a converged DiT.
Guideline 4: Projection matrices of query, key, value, and output in linear attention should be initialized randomly.
Guideline 5: Hybrid distillation is neccessary for student linear diffusion Transformer. We distill not only the predicted noise but also variances of the reverse diffusion process, but in a moderate way.
We provide an instruction for class-conditional image generation using LiT.
We provide an instruction for text-to-image generation using LiT.
We sincerely thank the following excellent codebases for their generous open-source spirit. We borrowed parts of their code.
@article{wang2025lit,
title={LiT: Delving into a Simplified Linear Diffusion Transformer for Image Generation},
author={Wang, Jiahao and Kang, Ning and Yao, Lewei and Chen, Mengzhao and Wu, Chengyue and Zhang, Songyang and Xue, Shuchen and Liu, Yong and Wu, Taiqiang and Liu, Xihui and others},
journal={arXiv preprint arXiv:2501.12976},
year={2025}
}