Research map / Vision-language and multimodal / Human and embodied modeling
AWRaCLe: All-Weather Image Restoration Using Visual In-Context Learning
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 AWRaCLe, an all-weather image restoration network that uses paired degraded/clean visual in-context examples, with CLIP-based degradation context extraction and cross-attention context fusion inside a Restormer decoder.
Problem
All-weather image restoration (deraining, desnowing, dehazing) with a single model, where existing methods learn degradation cues only from individual training images and do not exploit additional contextual guidance.
Approach
Feeds a degraded/clean context pair of the same degradation type as visual in-context prompt; extracts degradation-specific features from CLIP embeddings via self-attention (DCE blocks) and fuses them into each Restormer decoder level via cross-attention (CF blocks), with multi-level fusion.
Key results
- AWRaCLe reaches average 32.0 PSNR / 0.930 SSIM across SOTS, Rain100H, Rain100L and Snow100k, the best among compared methods (second best DiffUIR 30.3/0.903).
- On Rain100L, AWRaCLe achieves 35.7/0.966 versus Painter's 23.2/0.900 under the same in-context protocol.
- Ablation on SOTS: removing context pairing (unpaired) drops PSNR from 31.65 to 29.84; the full model with DCE, CF and multi-level fusion is best at 31.65/0.981.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AWRaCLe instead of the strongest reported baseline | AWIR methods · SOTS | PSNR: 31.0 -> 31.7 | Helped |
| use AWRaCLe instead of the strongest reported baseline | AWIR methods · Rain100H | PSNR: 26.5 -> 27.2 | Helped |
| use AWRaCLe instead of the strongest reported baseline | AWIR methods · Rain100L | PSNR: 31.8 -> 35.7 | Helped |
| use AWRaCLe instead of the strongest reported baseline | AWIR methods · Snow100k | PSNR: 33.4 -> 33.5 | Helped |
| add DCE (ablation) | AWRaCLe · SOTS | PSNR: 30.53 -> 31.65 | Helped |
| add CF (ablation) | AWRaCLe · SOTS | PSNR: 31.16 -> 31.65 | Helped |
| add multi-level fusion (ablation) | AWRaCLe · SOTS | PSNR: 30.12 -> 31.65 | Helped |
| use paired instead of unpaired context (ablation) | AWRaCLe · SOTS | PSNR: 29.84 -> 31.65 | Helped |
| provide incorrect rather than correct degradation context (ablation) | AWRaCLe · Rain100L | PSNR: 35.71 -> 25.93 | Hurt |
| provide incorrect rather than correct degradation context | Painter · Rain100L | PSNR: 23.19 -> 23.15 | No clear effect |
Surprising findings
- Painter's performance is essentially unchanged with incorrect context (23.15 vs 23.19 PSNR on Rain100L), suggesting it does not use the provided context.
- Unpaired context pairs (from different scenes) perform worse than paired ones, and the model with DCE and CF but unpaired context scores below the paired variant without DCE.
Limitations
- Requires knowing the degradation type of the query to select a matching context pair at inference.
- Context must be a paired degraded/clean image from the same scene, which may be hard to obtain in practice.
Experimental setup
Datasets: Snow100k Synthetic Rain Dataset (SRD) RESIDE OTS Rain100H Rain100L SOTS outdoor Foggy Cityscapes RainDS SnowCityscapes
Benchmarks: SOTS outdoor Rain100H Rain100L Snow100k
Metrics: PSNR SSIM
Baselines: WeatherDiff WGWS TSMC AirNet PromptIR Painter DA-CLIP DiffUIR DiffPlugin PromptGIP
Builds on: Restormer Painter CLIP PromptGIP
Closest papers by meaning
- Continuous Adverse Weather Removal via Degradation-Aware Distillation (CVPR 2025 · similarity 0.88)
- Towards Universal Rainy Image Restoration: Benchmark and Baseline (AAAI 2025 · similarity 0.88)
- Language-driven All-in-one Adverse Weather Removal (CVPR 2024 · similarity 0.87)
- Real-World Adverse Weather Image Restoration via Dual-Level Reinforcement Learning with High-Quality Cold Start (NeurIPS 2025 · similarity 0.87)
- Multi-axis Prompt and Multi-dimension Fusion Network for All-in-one Weather-degraded Image Restoration (AAAI 2025 · similarity 0.86)
- Genuine Knowledge from Practice: Diffusion Test-Time Adaptation for Video Adverse Weather Removal (CVPR 2024 · similarity 0.85)
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.