We’ve all seen photos that are rarely perfect. They might be blurry and too dark at the same time, or low-resolution with missing pixels. Until now, fixing these problems meant building a separate AI model for each type of damage—expensive, time-consuming, and difficult to scale.

Dolby researchers set out to change that by building SD-URM: Stable Diffusion-based Zero-Shot Universal Restoration Model—a single AI system that restores degraded frames across multiple damage types without needing to be retrained for each new scenario.

Instead of teaching an AI how to fix every possible problem, SD-URM first learns how an image was degraded—whether by blur, low light, low resolution, missing pixels, or any combination — and uses that understanding to guide the restoration process intelligently.

SD-URM: An Intelligent AI Restoration Model

The engine behind SD-URM is Stable Diffusion, a generative AI model that has already learned what natural, high-quality images should look like. Using only a small set of learned parameters tied to the specific degradation, SD-URM steers the diffusion process without training the data from scratch. This keeps computational costs low and makes the system faster to adapt.

Our researchers also identified that SD-URM handled multiple types of image damage simultaneously—something traditional models, including GAN-based approaches, consistently struggle with. It converged faster and delivered strong results across diverse real-time conditions.

The results indicate a promising future where a single adaptable system can handle the unpredictable, overlapping damage that real images carry without having to maintain an extensive library of specialized models. As generative AI continues to mature, SD-URM points toward image restoration that is more practical, easier to deploy, and better built for the complexity of the real world.

How Dolby Is Rethinking Video Restoration

{{ pageEducationalSlot.headingContent[index] }}