Research map / Image generation / Image restoration
DreamClear: High-Capacity Real-World Image Restoration with Privacy-Safe Dataset Curation
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.
Proposes GenIR, a dual-prompt fine-tuned T2I pipeline that synthesizes a privacy-safe one-million-image restoration dataset, and DreamClear, a DiT-based (PixArt-α) restoration model with a Mixture of Adaptive Modulator for degradation-aware expert fusion.
Problem
Real-world image restoration suffers from small, copyright- and privacy-problematic web-scraped datasets and a lack of high-capacity models that generalize across diverse degradations.
Approach
Synthesize IR training images with a T2I model adapted via learned positive/negative prompt tokens, MLLM-generated scene prompts and filtering; and build a DiT restoration model (ControlFormer) whose MoAM routes tokens to restoration experts using token-wise degradation maps derived from LQ-reference cross-attention.
Key results
- DreamClear achieves best LPIPS (0.3657), DISTS (0.1637) and FID (20.61) on DIV2K-Val among compared methods.
- On COCO detection, DreamClear reaches AP^b 19.3 vs 18.7 for the next best (DiffBIR), and mIoU 31.9 on ADE20K.
- A user study with 256 evaluators gave DreamClear over 45% of votes and top choice for 80% of images.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DreamClear instead of SeeSR | Real-world image restoration · DIV2K-Val | FID: 24.98 -> 20.61 | Helped |
| Use DreamClear instead of SeeSR | Real-world image restoration · LSDIR-Val | FID: 25.91 -> 22.06 | Helped |
| Use DreamClear instead of DiffBIR | Real-world image restoration · COCO val2017 | Object Detection AP^b: 18.7 -> 19.3 | Helped |
| Use DreamClear instead of DASR | Real-world image restoration · ADE20K | mIoU: 30.4 -> 31.9 | Helped |
| Use DreamClear instead of the pre-trained T2I model to generate training images | SwinIR-GAN · LSDIR-Val | LPIPS: 0.4819 -> 0.4578 | Helped |
| Add degradation-aware expert mixtures to the AM module (ablation) | DreamClear · DIV2K-Val | LPIPS: 0.3981 -> 0.3657 | Helped |
| Use Mixture of AM instead of cross-attention (ablation) | DreamClear · DIV2K-Val | LPIPS: 0.4177 -> 0.3657 | Helped |
| Use Mixture of AM instead of a zero-linear layer (ablation) | DreamClear · DIV2K-Val | LPIPS: 0.4082 -> 0.3657 | Helped |
| Add the Reference Branch (ablation) | DreamClear · DIV2K-Val | LPIPS: 0.4207 -> 0.3657 | Helped |
| Use detailed MLLM text prompts instead of null prompts (ablation) | DreamClear · DIV2K-Val | LPIPS: 0.3521 -> 0.3657 | Hurt |
| Use detailed MLLM text prompts instead of null prompts (ablation) | DreamClear · DIV2K-Val | MANIQA: 0.4230 -> 0.4320 | Helped |
Surprising findings
- DreamClear has lower PSNR/SSIM than several baselines, which the authors attribute to these metrics poorly reflecting visual quality.
- Null prompts slightly outperform MLLM detailed prompts on perceptual metrics (LPIPS, DISTS, FID).
Limitations
- Diffusion-based multi-step inference is too slow for real-time applications.
- Under severe degradation, synthesized textures may not match ground truth.
- Lower PSNR/SSIM than several baselines.
Experimental setup
Datasets: DIV2K Flickr2K LSDIR DIV8K COCO val2017 ADE20K RealSR DRealSR RealLQ250
Benchmarks: DIV2K-Val LSDIR-Val RealSR DRealSR RealLQ250 COCO val2017 ADE20K
Metrics: PSNR SSIM LPIPS DISTS FID NIQE MANIQA MUSIQ CLIPIQA AP^b
Baselines: BSRGAN Real-ESRGAN SwinIR-GAN DASR StableSR DiffBIR ResShift SinSR SeeSR SUPIR
Builds on: ControlNet PixArt-α SDXL AdaLN Real-ESRGAN
Closest papers by meaning
- GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration (CVPR 2025 · similarity 0.89)
- ReFIR: Grounding Large Restoration Models with Retrieval Augmentation (NeurIPS 2024 · similarity 0.89)
- Acquire and then Adapt: Squeezing out Text-to-Image Model for Image Restoration (CVPR 2025 · similarity 0.88)
- Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild (CVPR 2024 · similarity 0.88)
- IRBridge: Solving Image Restoration Bridge with Pre-trained Generative Diffusion Models (ICML 2025 · similarity 0.87)
- RAP-SR: RestorAtion Prior Enhancement in Diffusion Models for Realistic Image Super-Resolution (AAAI 2025 · similarity 0.87)
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.