A Frozen Pixel-Space Diffusion Model Can Guide Itself with Its Own Samples

Researchers propose Synthetic Self-Guidance, a lightweight method that improves a frozen pixel-space diffusion model during sampling.

It adds a small head on an intermediate layer and uses the gap between intermediate and final predictions to steer generation.

The head can be trained on model-generated synthetic images instead of real data, with especially strong gains on high-frequency detail.

Across several ImageNet pixel diffusion models, the approach delivers better sample quality and lower FID with very low extra compute.