The authors trained diffusion backbones from scratch in controlled architectural comparisons, focusing on the presence or absence of long skip connections.
Long skip connections help diffusion transformers learn cleaner visual representations
Controlled training and diagnostic experiments suggest skip connections keep trajectory-related diffusion noise out of the network layers used as image features.
Chinese Tech
Benyuan Meng · Yiliang Zhang · Jin-Wen Wu · Qianqian Xu · Qingming Huang · Longtao Huang
Institute of Information Engineering, CAS · University of Chinese Academy of Sciences · Alibaba Group · Key Lab. of Intelligent Information Processing, Institute of Computing Technology, CAS · Beijing Academy of Artificial Intelligence
Research Digest··2 min read
Meng and colleagues investigate why Diffusion Transformers, or DiTs, can generate strong images while producing weaker visual representations than U-Net diffusion models.
Why this paper
From University of Chinese Academy of Sciences and 5 others
In one line
Diffusion Transformers without long skip connections produce weaker visual features because diffusion noise pollutes the main path; adding long skip connections restores feature quality.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (3 noted)
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§