Prompting Rain Off: Evolving Compact Dual Prompts for Continual De-Raining

Liu, Minghao; Yang, Wenhan; Liu, Jiaying

doi:10.1109/TIP.2026.3689428

IEEE Transactions on Image Processing, vol. 35, pp. 5266-5280, 2026

Prompting Rain Off Evolving Compact Dual Prompts for Continual De-Raining

Minghao Liu, Wenhan Yang, and Jiaying Liu

Peking University · Pengcheng Laboratory

DOI IEEE Xplore GitHub

EcoDPL architecture with image prompts, feature prompts, P-Fuser, and Grad-Tuner — EcoDPL injects compact prompts at both image and feature levels, then evolves them through adaptive fusion and gradient-aware compression.

Abstract

Continual De-Raining Without Rehearsal

Single-image rain removal has advanced rapidly, but most methods still assume a static data distribution. In real scenes, rain patterns change with season, geography, camera, and weather intensity, so a de-raining model must learn new degradations without catastrophically forgetting old ones.

We propose Evolving COmpact Dual Prompt Learning (EcoDPL), a rehearsal-free continual learning framework for low-level vision. EcoDPL maintains two prompt pools at the image and feature levels, inserts them into images and embedding tokens, and uses P-Fuser to generate adaptive attention-aware prompt weights for each input.

To keep knowledge compact, Grad-Tuner compresses learned knowledge into fewer prompts through a gradually tuned dictionary learning strategy, freeing space for future tasks while preserving useful de-raining behavior. Across synthetic and real-world benchmarks, EcoDPL consistently improves continual rain removal and remains effective under stationary data.

Highlights

What EcoDPL Adds

Prompt-Based Continual De-Raining

A prompt-learning continual model tailored for low-level rain removal, using fixed additional parameters rather than expanding a full model for every task.

Dual-Level Prompt Pools

Image-level prompts capture local restoration cues, while feature-level prompts guide higher-level representation transfer across evolving rain types.

P-Fuser

An adaptive weight generation module assigns attention maps to prompts, allowing different inputs to receive flexible prompt combinations.

Grad-Tuner

A dictionary-learning compression strategy packs useful knowledge into fewer prompts, leaving capacity for new tasks in continual learning.

Method

Evolving Compact Dual Prompts

Comparison between traditional continual learning and EcoDPL — Compared with regularization-style continual learning, EcoDPL stores task-specific knowledge in compact learnable prompts and selectively updates them for each instance.

Training pipeline of EcoDPL for continual rain removal — During each task, EcoDPL updates prompt pools and then compresses them with Grad-Tuner.

Details of the P-Fuser prompt fusion module — P-Fuser produces input-aware prompt weights instead of relying on a fixed distance rule.

Results

Continual and Real-World Rain Removal

Rain800 → Rain100H 25.18 / 0.8441

Average PSNR / SSIM with EcoDPL.

Prompt Capacity Compact

Grad-Tuner compresses old knowledge before learning new tasks.

Backbone Friendly 3 Models

Validated with TransWeather, Restormer, and PromptIR backbones.

Visual comparison on Rain800 after continual training — On synthetic rain streak removal, EcoDPL recovers cleaner details after sequential training than classical continual learning baselines.

Real rain visual comparison after sequential Rain800 and Rain100H training — Real rainy scenes show the practical benefit of adaptive prompt fusion.

Real-world visual comparison with EcoDPL — EcoDPL reduces rain artifacts while preserving structure and texture details.

Table II continual rain removal PSNR and SSIM comparison — Continual performance on Rain800 and Rain100H.

Table III comparison across de-raining backbones — EcoDPL remains effective across different de-raining backbones.

Table IV non-reference metrics on real-world datasets — Non-reference real-world evaluation on IIVIPC-DQA and DNN-SIRR.

Resources

Project Materials

Repository StaryMoon/Prompting-Rain-Off Project Page starymoon.github.io/Prompting-Rain-Off/ Paper DOI 10.1109/TIP.2026.3689428 IEEE Xplore Document 11511395

Citation

Cite This Work

@article{liu2026prompting,
  title={Prompting Rain Off: Evolving Compact Dual Prompts for Continual De-Raining},
	  author={Liu, Minghao and Yang, Wenhan and Liu, Jiaying},
	  journal={IEEE Transactions on Image Processing},
	  volume={35},
	  pages={5266--5280},
	  year={2026},
	  doi={10.1109/TIP.2026.3689428},
	  publisher={IEEE}
	}