One model for all image restoration: balancing speed and realism

Generative models like diffusion produce beautiful, realistic images but are slow and imprecise. Classical regression methods are fast and faithful to the input but dull. This work unifies both by splitting the interpolation process into independent generation and regression paths, letting you smoothly trade off fidelity for perceptual quality at inference time. A dual-branch transformer handles both modes efficiently, achieving competitive results on denoising, inpainting, and super-resolution.