Stroke-Size Control for Improved Diffusion Training

A new method addresses low signal-to-noise challenges in diffusion models by implementing stroke-size control. This technique adjusts the roughness of supervised targets and perturbations across timesteps to simplify prediction tasks during high-noise regimes.