Representation alignment encourages intermediate diffusion features to match a frozen pretrained encoder, but not all encoder information is equally useful.
Purifying alignment targets improves diffusion generation, especially for reconstructive encoders
The authors' local appearance purification method removes locally predictable content before alignment, cutting MAE-based FID from 53.1 to 42.3 on ImageNet-256 with SiT-B/2.
Top University
Yingheng Wang · Yaoqiang Li · Yaqin Wu · Junwen Bai · Jiatao Gu · Christopher De Sa · +1 more
Cornell University · University of Pennsylvania
Research Digest··2 min read
The authors introduce local appearance purification (LAP), which strips locally predictable content from diffusion alignment targets, and show it improves generation across five pretrained encoders, with the largest gain for MAE.
Why this paper
From University of Pennsylvania and Cornell University
In one line
Aligning to the residual after removing locally predictable content (LAP) improves diffusion generation quality and turns MAE into an effective alignment target.
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 (2 noted)
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§