IAAILoopREG
LoopREG: Representation Supervision Enables Test-Time Recurrent Depth in Diffusion Transformers
A recurrent diffusion Transformer that repeatedly applies a shared block at a state aligned to a pretrained visual encoder after the final recurrent visit, so inference can extend the same operator to additional visits.
Inside the method
[ Shared DiT block ] × visits → representation alignment (training)
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Core recurrence
- Depth control
- Extra inference visits beyond training
- KV / state strategy
- Not applicable (diffusion)
Reading note
On ImageNet 256×256, SiT-B/2 FID drops from 15.08 to 4.91 at 400K steps versus the REG baseline; one extra inference visit on a two-visit checkpoint improves FID from 14.71 to 5.34. Holds for SiT-L/2 and SiT-XL/2.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{loopreg2026,
title = {LoopREG: Representation Supervision Enables Test-Time Recurrent Depth in Diffusion Transformers},
author = {Zipeng Wang and Anping Xiong and Jing Zhang and Xiaoya Zhu and Yi Jiang and Yanbing Xiao and Zige Zeng and Zezhou Chen and Xiang Liu and Huan Hu and Ping Chen and Kai Wang and Zhaoxiang Liu and Shiguo Lian and Ye LV and Fang Zhao and Huanlin Gao and Zhengrui Huang and Bo Shen and Yuren You},
year = {2026},
eprint = {2610.17877},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.17877}
}