Research map / Image generation / Image restoration
Restoration by Generation with Constrained Priors
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Restores degraded images by adding noise to the input and denoising with a pretrained diffusion model after fine-tuning it on an anchor album (generative album or personal album) that constrains the generative space.
Problem
Blind image restoration without paired training data or known degradation models, where restored outputs must stay faithful to the input identity while being high quality.
Approach
Project the degraded input into a diffusion model's generative space by adding noise to step K and denoising, after fine-tuning the model on anchor images (a generative album produced via skip guidance from the input, or a personal album) so the sampling space is constrained toward the input's content.
Key results
- On Wider-Test, FID 46.38 and MUSIQ 58.73 versus CodeFormer's 48.57 and 55.70, surpassing all supervised baselines on all four real-world face datasets.
- On Deblur-Test, FID 135.33 and MUSIQ 60.20 versus CodeFormer's 163.47 and 57.09, showing better generalization to unseen motion blur.
- Personalized identity score on Obama reached 0.716 versus 0.558 for CodeFormer and 0.574 for ASFFNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use restoration-by-generation with a generative album rather than supervised baseline methods | Blind face restoration · Wider-Test | FID: 48.57 -> 46.38 | Helped |
| Use restoration-by-generation with a generative album rather than supervised baseline methods | Blind face restoration · Wider-Test | MUSIQ: 56.48 -> 58.73 | Helped |
| Use restoration-by-generation with a generative album rather than supervised baseline methods | Blind face restoration · WebPhoto-Test | FID: 98.55 -> 96.44 | Helped |
| Use restoration-by-generation with a generative album rather than supervised baseline methods | Blind face restoration · WebPhoto-Test | MUSIQ: 55.20 -> 57.71 | Helped |
| Use restoration-by-generation with a generative album rather than supervised baseline methods | Blind face restoration · LFW-Test | FID: 64.10 -> 56.32 | Helped |
| Use restoration-by-generation with a generative album rather than supervised baseline methods | Blind face restoration · LFW-Test | MUSIQ: 60.46 -> 60.68 | Helped |
| Use restoration-by-generation with a generative album rather than supervised baseline methods | Blind face restoration · Deblur-Test | FID: 163.47 -> 135.33 | Helped |
| Use restoration-by-generation with a generative album rather than supervised baseline methods | Blind face restoration · Deblur-Test | MUSIQ: 58.03 -> 60.20 | Helped |
| Use a personal album to constrain the generative space | Personalized face restoration · Subject A | identity score: 0.694 -> 0.731 | Helped |
| Use a personal album to constrain the generative space | Personalized face restoration · Obama | identity score: 0.574 -> 0.716 | Helped |
| Use a personal album to constrain the generative space | Personalized face restoration · Hermione | identity score: 0.522 -> 0.664 | Helped |
| Add a generative album close to the input to constrain the generative space (ablation) | Single-image blind face restoration · Single-image restoration | image quality and input fidelity | Helped |
Surprising findings
- Fine-tuning on a generative album of images sampled from the input itself is enough to beat supervised face restoration methods without any paired training data.
Limitations
- Single-image restoration requires per-input fine-tuning, which is slow compared to feed-forward methods.
- Validated mainly on class-specific domains (faces, dogs, cats) due to lack of a high-quality pretrained diffusion model for natural images.
Experimental setup
Datasets: FFHQ AFHQ Dog AFHQ Cat Wider-Test LFW-Test WebPhoto-Test Deblur-Test (67 images from [25])
Benchmarks: Wider-Test WebPhoto-Test LFW-Test Deblur-Test
Metrics: FID MUSIQ identity score (ArcFace cosine similarity)
Baselines: GFPGAN CodeFormer VQFR DR2 ASFFNet
Builds on: DDPM DDPM-based posterior sampling (DPS) CodeFormer
Closest papers by meaning
- Fast constrained sampling in pre-trained diffusion models (NeurIPS 2025 · similarity 0.91)
- IRBridge: Solving Image Restoration Bridge with Pre-trained Generative Diffusion Models (ICML 2025 · similarity 0.90)
- Improving Diffusion-Based Image Restoration with Error Contraction and Error Correction (AAAI 2024 · similarity 0.90)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.89)
- A Variational Perspective on Solving Inverse Problems with Diffusion Models (ICLR 2024 · similarity 0.88)
- Consistency Posterior Sampling for Diverse Image Synthesis (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.