A team of researchers from Cornell and NUS asks whether diffusion models can borrow a trick from human painters: start with rough strokes, then refine. Their proposed MultiStroke operator linearly blends coarse and fine representations in a coarse-to-fine schedule, aiming to make the prediction target easier to learn, especially at high-noise timesteps where current models struggle.

In experiments, the authors trained DDPM and flow-matching models from scratch with MultiStroke and report FID reductions of up to 70.2% compared with baselines. The gains are attributed to simplifying the regression target during the early, noisy phase of the denoising chain, rather than changing the model architecture or adding computation.

The paper is a preprint that has been revised several times since March 2026, with the latest version posted in October 2026. As with any preprint, the results have not yet been peer-reviewed, and the method's effectiveness on larger, state-of-the-art diffusion systems remains to be demonstrated.