Research map / Image generation / Super-resolution
Decoupling Direction and Magnitude: Language-Steered Flow Matching for Super-Resolution in the Dark
Jiaxin Gao, Ziyu Yue, Yaohua Liu, Danchen Cui, Zhihui Zhao, Zhixun Su
Read the paperCheck your idea against 137,217 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 a language-steered flow-matching approach for super-resolution in the dark that decouples update direction and magnitude to adapt integration to spatially varying exposure.
Problem
Image super-resolution in low-light (dark) conditions is challenged by extreme spatial heterogeneity: underexposed noisy regions need aggressive generative enhancement while well-exposed regions need conservative updates to avoid hallucinated textures.
Approach
Standard diffusion and flow-matching models use globally uniform integration trajectories; the paper instead decouples update direction and magnitude and uses language steering to adapt them spatially, so that under-exposed regions get aggressive generative updates and well-exposed regions get conservative ones.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add degradation language prompt (ablation) | DeLang-SR · RELLISUR x2 (5K ablation) | PSNR: 22.433 -> 22.986 | Improved |
| add visual degradation tokens (with attribute alignment) (ablation) | DeLang-SR · RELLISUR x2 (5K ablation) | PSNR: 22.398 -> 22.986 | Improved |
| add attribute alignment (ablation) | DeLang-SR · RELLISUR x2 (5K ablation) | PSNR: 22.516 -> 22.986 | Improved |
| add attribute alignment (ablation) | DeLang-SR · RELLISUR x2 (5K ablation) | LPIPS: 0.2456 -> 0.2482 | No clear effect |
| add descriptor conditioning in update-strength predictor (ablation) | DeLang-SR · RELLISUR x2 (5K ablation) | PSNR / ColorErr: PSNR 22.754 -> 22.986; ColorErr 0.1023 -> 0.1120 | Mixed |
| add degradation-control modules to S3Diff-based base model (ablation) | DeLang-SR · RELLISUR x2 (5K ablation) | PSNR: 21.358 -> 22.986 | Improved |
| use LQ-only degradation intent instead of generic or oracle intent (ablation) | DeLang-SR · RELLISUR | quality | No clear effect |
| compare against DOD on no-reference quality | DeLang-SR · RELLISUR x2 | CLIP-IQA: 0.4224 (DOD) vs 0.4188 | Mixed |
| use SD3-based one-step flow update vs lightweight SR backbones | DeLang-SR · Inference time | seconds: 0.088 (DOD) -> 0.448 | Hurt |
| use DeLang-SR vs best baseline at x4 | DeLang-SR · RELLISUR x4 | PSNR: 20.276 (Restormer) -> 21.259 | Improved |
Experimental setup
Builds on: flow matching
Closest papers by meaning
- When Latents Forget Pixels: Restoring Fidelity in Diffusion Transformer Super-Resolution (NeurIPS 2026 · similarity 0.86)
- DORNet: A Degradation Oriented and Regularized Network for Blind Depth Super-Resolution (CVPR 2025 · similarity 0.86)
- PRISM: Prior Rectification and Uncertainty-Aware Structure Modeling for Diffusion-Based Text Image Super-Resolution (NeurIPS 2026 · similarity 0.86)
- Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution (CVPR 2024 · similarity 0.86)
- TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution (CVPR 2025 · similarity 0.86)
- LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution (NeurIPS 2026 · 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.
