Diffusion-based Real-World Super-Resolution

Project overview

Real-world image super-resolution (Real-ISR) has to restore images whose degradations are unknown, mixed and of varying severity. This project studies how to make large pretrained diffusion models, such as Stable Diffusion, work for that setting along two complementary directions: representing what is wrong with an image and how badly, and using that knowledge to decide how a diffusion model should restore it.

Ordinal degradation representations

The first paper, Learning Ordinal Degradation Representations with Textual Priors for Diffusion-Based Blind Image Super-Resolution (arXiv:2512.10340), was first released as ZePHiR: zero-shot adaptation of Stable Diffusion through a plug-in hierarchical degradation representation.

• Decoupled degradation: an HD-CLIP component separates degradation type from degradation intensity using composition-based ordinal text embeddings.
• Plug-in design: the representation is attached to a pretrained diffusion model rather than training a restoration network from scratch.

One-step diffusion with manifold inversion

The second paper, Bridging Restoration and Generation in One-step Diffusion for Real-World Image Super-Resolution (arXiv:2604.24136), targets the cost of iterative diffusion sampling. Distilling a diffusion model into a single step is fast but runs into a perception-distortion trade-off: common one-step methods turn the generative prior into a deterministic mapping and lose its stochastic nature, while methods that bring randomness back by shifting timesteps or adding noise only control one side, so restoration and generation agree at a few preset timesteps and drift apart once steered.

The proposed IDaS-SR (Inversion and Degradation-aware Sampling for Real-ISR) is a one-step framework with two parts:

• Manifold Inversion Noise Estimator: anchors the low-quality latent onto the pretrained diffusion trajectory.
• CHARIOT: reintroduces controlled stochasticity, so a single scalar moves smoothly along the fidelity-realism trade-off at inference.

The authors report state-of-the-art results under explicit control in a single inference step.

Related publications