Diffusion models can restore blurry and damaged images without retraining — just by learning better prompts

Researchers discovered that pre-trained image generation models already know how to fix blurry, degraded, or damaged images — you just have to ask them the right way. Instead of rebuilding the model from scratch, you can unlock this ability by teaching the model new prompt instructions, making image restoration faster and cheaper to deploy.

For years, using pre-trained image models for real-world restoration tasks required expensive retraining or bolting on new specialized modules. This finding means that capability was already there — dormant but present. That changes the economics: a developer can now take an off-the-shelf model, spend compute time optimizing prompts instead of retraining weights, and get a working restoration system. The practical effect is that image restoration becomes cheaper and faster to deploy at scale.