Edit2Restore:Few-Shot Image Restoration via Parameter-Efficient Adaptation of Pre-trained Editing Models

TL;DR

The paper tackles the data-intensive nature of traditional image restoration by reusing large-scale pre-trained text-conditioned image editing models. It introduces a parameter-efficient fine-tuning strategy using LoRA adapters on the FLUX.1 Kontext model to perform denoising, deraining, and dehazing with only $16$ to $128$ paired examples per task. A single unified LoRA adapter conditioned on text prompts handles multiple degradations, and restoration quality is guided by rectified flow matching to emphasize perceptual fidelity over pixel-wise metrics. Ablations show that joint text encoder fine-tuning and unified multi-task adapters often match or exceed task-specific counterparts while dramatically reducing storage, demonstrating a practical pathway for few-shot restoration with editing priors.

Abstract

Image restoration has traditionally required training specialized models on thousands of paired examples per degradation type. We challenge this paradigm by demonstrating that powerful pre-trained text-conditioned image editing models can be efficiently adapted for multiple restoration tasks through parameter-efficient fine-tuning with remarkably few examples. Our approach fine-tunes LoRA adapters on FLUX.1 Kontext, a state-of-the-art 12B parameter flow matching model for image-to-image translation, using only 16-128 paired images per task, guided by simple text prompts that specify the restoration operation. Unlike existing methods that train specialized restoration networks from scratch with thousands of samples, we leverage the rich visual priors already encoded in large-scale pre-trained editing models, dramatically reducing data requirements while maintaining high perceptual quality. A single unified LoRA adapter, conditioned on task-specific text prompts, effectively handles multiple degradations including denoising, deraining, and dehazing. Through comprehensive ablation studies, we analyze: (i) the impact of training set size on restoration quality, (ii) trade-offs between task-specific versus unified multi-task adapters, (iii) the role of text encoder fine-tuning, and (iv) zero-shot baseline performance. While our method prioritizes perceptual quality over pixel-perfect reconstruction metrics like PSNR/SSIM, our results demonstrate that pre-trained image editing models, when properly adapted, offer a compelling and data-efficient alternative to traditional image restoration approaches, opening new avenues for few-shot, prompt-guided image enhancement. The code to reproduce our results are available at: https://github.com/makinyilmaz/Edit2Restore

Figures (3)

• Figure 1: Traditional approaches require training specific models with thousands of paired examples for each restoration task. Recent diffusion-based multi-task image restoration models leverage strong priors but still demand extensive paired data to train from scratch. Our approach achieves few-shot efficiency by adapting pre-trained image editing models with only 16-128 examples per task through parameter-efficient fine-tuning.
• Figure 2: Qualitative comparison of image restoration results.
• Figure 3: Qualitative comparison between task-specific and unified multi-task adapters across different restoration tasks. Both configurations produce visually similar results, demonstrating that a single unified adapter can effectively handle multiple degradation types through text prompt conditioning while maintaining quality comparable to specialized task-specific adapters. The unified model offers practical advantages in terms of storage and deployment simplicity without sacrificing restoration quality.