Input
Predicted
32.7
PSNR
1
Steps
180M
Params
Abstract
A distilled diffusion model for blind image restoration that runs in a single forward pass. We train a teacher diffusion model on a curated mixture of degradations and distill it into a student that matches its output in one step.
Why one step
Diffusion models are powerful but slow. For restoration in real-time pipelines, we need single-step inference. Distillation lets us keep the quality of multi-step diffusion at a fraction of the cost.