Stroke-Size Control Improves Diffusion Model Performance
A new paper on arXiv proposes stroke-size control as a method to improve diffusion models in low signal-to-noise regimes. The authors draw an analogy to oil painting, where using a fine stroke throughout is ineffective. By adjusting the effective roughness of supervised targets, predictions, and perturbations across timesteps, they aim to ease the challenge of pixel-level predictions under high noise. The paper is available at arXiv:2603.26783.
Key facts
- arXiv:2603.26783v2
- Announce Type: replace-cross
- Diffusion models face challenges in low signal-to-noise regimes
- Stroke-size control is proposed as a controlled intervention
- Analogy to oil painting with finest stroke
- Changes effective roughness of supervised target, predictions, and perturbations across timesteps
- Aims to ease low signal-to-noise challenge
- Published on arXiv
Entities
Institutions
- arXiv