On Distillation of Guided Diffusion Models

Text-guided generation (2 steps) Text-guided generation (4 steps) Image to image translation (3 steps) Class-conditional generation (1 step) Image inpainting (2 steps) Input Input Mask Result 1 Result 2 Output (different styles) Text-guided generation (1 step) Figure 1 . Distilled Stable Diffusion samples generated by our method. Our two-stage distillation approach is able to generate realistic images using only 1 to 4 denoising steps on various tasks. Compared to standard classifier-free guided diffusion models, we reduce the total number of sampling steps by at least 20⇥.