Research map / Architectures and efficiency / Mixture of experts
Complexity Experts are Task-Discriminative Learners for Any Image Restoration
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 MoCE-IR, a mixture-of-complexity-experts image restoration model whose experts vary in computational complexity and receptive field, with a complexity-biased router that yields task-discriminative expert allocation for all-in-one restoration.
Problem
All-in-one image restoration models waste parameters because task-specific parameters stay inactive for other tasks, and existing MoE routing yields inconsistent expert behavior (some experts over-generalize, others underperform), limiting efficient bypassing of irrelevant experts.
Approach
Replace uniform MoE experts with nested experts of increasing channel dimension and window size, and add a complexity-proportional bias (normalized parameter counts) to the load-balancing importance loss so that routing favors lower-complexity experts, which naturally produces task-specific allocation.
Key results
- MoCE-IR-S (11M) reaches 32.57 dB average PSNR on the three-degradation setting, above AirNet's 31.20 dB.
- On five degradations, MoCE-IR (25M) averages 30.58 dB PSNR versus 29.55 dB for InstructIR-5D.
- On CDD11 composite degradations, MoCE-IR-S averages 29.05 dB, a 0.64 dB gain over OneRestore.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use complexity experts with complexity-biased routing (ablation) | MoCE-IR-S · AIO-3 average | PSNR: 32.30 -> 32.57 | Helped |
| use complexity-biased routing instead of load balancing (ablation) | MoCE-IR-S · AIO-3 average | PSNR: 32.39 -> 32.57 | Helped |
| use nested rather than exponential expert scaling (ablation) | MoCE-IR-S · AIO-3 average | PSNR: 32.47 -> 32.57 | Helped |
| normalize router weights by the maximum rather than minimum expert capacity (ablation) | MoCE-IR-S · AIO-3 average | PSNR: 32.44 -> 32.57 | Helped |
| replace complexity-biased routing with AMIR routing (ablation) | MoCE-IR-S · AIO-3 average | PSNR: 32.57 -> 32.28 | Hurt |
| replace complexity-biased routing with WM-MoE routing (ablation) | MoCE-IR-S · AIO-3 average | PSNR: 32.57 -> 31.99 | Hurt |
Surprising findings
- Complexity-biased routing without any explicit task labels yields task-discriminative expert allocation, with some experts shared across tasks and others task-specific.
Limitations
- Image-level routing limits scalability and fine-grained spatial adaptivity.
- Evaluated only on synthetic degradations; extension to real-world and video is left to future work.
Experimental setup
Datasets: BSD400 WED Rain100L SOTS GoPro LOL-v1 BSD68 CDD11
Benchmarks: SOTS Rain100L BSD68 GoPro LOLv1 CDD11
Metrics: PSNR SSIM parameters FLOPs runtime
Baselines: AirNet PromptIR MPRNet Gridformer DA-CLIP InstructIR UniProcessor IDR OneRestore AMIR WM-MoE MoE with load balancing
Builds on: sparse mixture-of-experts (Riquelme et al.) Restormer PromptIR FFTformer
Closest papers by meaning
- Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration (NeurIPS 2026 · similarity 0.90)
- Parameter Efficient Adaptation for Image Restoration with Heterogeneous Mixture-of-Experts (NeurIPS 2024 · similarity 0.87)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.87)
- IRPO: Boosting Image Restoration via Post-training GRPO (NeurIPS 2026 · similarity 0.87)
- Multilinear Mixture of Experts: Scalable Expert Specialization through Factorization (NeurIPS 2024 · similarity 0.87)
- Debiased All-in-one Image Restoration with Task Uncertainty Regularization (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.