Research map / Image generation / Image restoration
FoundIR-v2: Optimizing Pre-Training Data Mixtures for Image Restoration Foundation Model
Xiang Chen, Jinshan Pan, Jiangxin Dong, Jian Yang, Jinhui Tang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add data equilibrium scheduling (DES) instead of mixed training (ablation) | FoundIR-v2 · Four-task avg (deblur, dehaze, low-light, SR) | PSNR: 18.91 -> 20.41 | Improved |
| add DES instead of sequential learning (ablation) | FoundIR-v2 · Four-task avg (deblur, dehaze, low-light, SR) | PSNR: 18.69 -> 20.41 | Improved |
| add DES instead of incremental learning (ablation) | FoundIR-v2 · Four-task avg (deblur, dehaze, low-light, SR) | PSNR: 19.93 -> 20.41 | Improved |
| add DES to PromptIR (ablation) | PromptIR · Four-task avg (deblur, dehaze, low-light, SR) | PSNR: 15.89 -> 16.92 | Improved |
| add DES to FoundIR (ablation) | FoundIR · Four-task avg (deblur, dehaze, low-light, SR) | PSNR: 16.77 -> 17.95 | Improved |
| replace FoundIR model with FoundIR-v2 under DES (no SR task) (ablation) | FoundIR-v2 · Four-task avg (deblur, dehaze, low-light, SR) | PSNR: 17.95 -> 20.09 | Improved |
| use equal or inappropriate data mixture proportions (ablation) | all-in-one restoration model · Deblur/dehaze/low-light/SR | PSNR per task | Hurt |
| vary evaluation interval T (10000, 25000, 30000) (ablation) | DES · Four-task avg (deblur, dehaze, low-light, SR) Avg PSNR | Avg. PSNR: 20.17 / 20.53 / 20.41 | No clear effect |
| vary D_ref size (10, 20, 30) (ablation) | DES · Four-task avg (deblur, dehaze, low-light, SR) Avg PSNR | Avg. PSNR: 20.41 / 20.48 / 20.36 | No clear effect |
| add soft MoE-driven scheduler to SDXL (ablation) | FoundIR-v2 · Four sub-tasks (radar chart) | task PSNR/MUSIQ | Improved |
| filter low-quality GT images from training data (ablation) | FoundIR-v2 · Four-task average | MUSIQ | Improved |
| use FoundIR-v2 versus best prior on UHD-LL and FoundIR-L | FoundIR-v2 · UHD-LL low-light | PSNR: 22.52 (AutoDIR) vs 20.08 | Hurt |
Later papers that compared with this method
1 later accepted paper reports FoundIR-v2 as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- TPGDiff : Hierarchical Triple-Prior Guided Diffusion for Image Restoration (ICML 2026 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 4 on most shared results.
- AutoDIR — AutoDIR: Automatic All-in-One Image Restoration with Latent Diffusion (ECCV 2024) · better on 5 of 7 shared results
- FoundIR — FoundIR: Unleashing Million-scale Training Data to Advance Foundation Models for Image Restoration (ICCV 2025) · better on 3 of 3 shared results
- FaithDiff — FaithDiff: Unleashing Diffusion Priors for Faithful Image Super-resolution (CVPR 2025) · better on 1 of 2 shared results
- SeeSR — SeeSR: Towards Semantics-Aware Real-World Image Super-Resolution (CVPR 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Parameter Efficient Adaptation for Image Restoration with Heterogeneous Mixture-of-Experts (NeurIPS 2024 · similarity 0.88)
- DreamClear: High-Capacity Real-World Image Restoration with Privacy-Safe Dataset Curation (NeurIPS 2024 · similarity 0.88)
- Debiased All-in-one Image Restoration with Task Uncertainty Regularization (AAAI 2025 · similarity 0.87)
- ReFIR: Grounding Large Restoration Models with Retrieval Augmentation (NeurIPS 2024 · similarity 0.87)
- IRPO: Boosting Image Restoration via Post-training GRPO (NeurIPS 2026 · similarity 0.86)
- EnsIR: An Ensemble Algorithm for Image Restoration via Gaussian Mixture Models (NeurIPS 2024 · similarity 0.86)
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.
