Proposes edge-preserving diffusion via hybrid noise with edge-aware scheduler to enhance structural fidelity in diffusion and flow-matching models.
For latent-space diffusion (Rombach et al., 2022), we tested on CelebA (2562) and AFHQ-Cat (5122)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2024 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Edge-preserving noise for diffusion models
Proposes edge-preserving diffusion via hybrid noise with edge-aware scheduler to enhance structural fidelity in diffusion and flow-matching models.