Representation Entanglement for Generation: Training Diffusion Transformers Is Much Easier Than You Think
Shen Zhang, Ruijing Shi, Shanghua Gao, Zhenyuan Chen, Lei Wang, Zhaowei Chen Hongcheng Gao Yao Tang Jian Yang Ming-Ming Cheng Xiang Li, [0.04cm] VCIP, CS, Nankai University, NKIARI, Shenzhen Futian, JIIOV Technology · 2025
Read Representation Entanglement for Generation: Training Diffusion Transformers Is Much Easier Than You Think by Shen Zhang, Ruijing Shi, Shanghua Gao, Zhenyuan Chen, Lei Wang, Zhaowei Chen Hongcheng Gao Yao Tang Jian Yang Ming-Ming Cheng Xiang Li, [0.04cm] VCIP, CS, Nankai University, NKIARI, Shenzhen Futian, JIIOV Technology (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-01467.txt.
Description
Teaching an image machine to clean noise from a picture is slow, and a known trick lends it eyes: align its inner views with a trained model's clean views. The authors of a 2025 paper argue that trick stops too soon. The lent eyes guide training, then vanish at use, when the machine works from pure noise.
Their fix tangles the loan into the work itself. Low-level picture codes are woven with one high-level class token from the trained model, and the pair is denoised as one. The machine then births matched picture-and-class pairs from noise alone. Birth quality and training speed both jump, at near-no added cost in use. Simple tangles beat side lessons.
Catalog record
| CREATOR | Shen Zhang, Ruijing Shi, Shanghua Gao, Zhenyuan Chen, Lei Wang, Zhaowei Chen Hongcheng Gao Yao Tang Jian Yang Ming-Ming Cheng Xiang Li, [0.04cm] VCIP, CS, Nankai University, NKIARI, Shenzhen Futian, JIIOV Technology |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-01467