Accepted to ACCV 2026

ImIRImage-Instruction Tuning for
All-in-One Image Restoration

Süleyman AslanGörkay AydemirMısra YavuzYunus Bilge KurtNasrin RahimiAhmet Rasim EmirdağıBurak Can BinerM. Akın Yılmaz

Codeway AI Research

arXiv2609.25267 Source codeGitHub Open weightsHugging Face ProceedingsComing soon

ImIR output: rain removed from a moonlit forest — **One adapter, six tasks, no prompt.** The output of the task-agnostic ImIR is produced from the input image alone with no text prompt and no degradation label.

Degraded input: rain streaks over a moonlit forest — **One adapter, six tasks, no prompt.** The output of the task-agnostic ImIR is produced from the input image alone with no text prompt and no degradation label.

Abstract

Degradations vary widely across images, so a practical restoration system has to handle many degradation types with one model. A recent and effective recipe adapts a large pretrained image-editing model to restoration using a small low-rank adapter with a text prompt. We replace that prompt with an instruction derived from the degraded image itself. The image reaches the editor through two paths: its structure comes from the model’s VAE, and its semantic instruction comes from a lightweight token mapper that shifts the degraded image’s vision-language embedding toward the embedding a clean image would produce. Because the instruction is a continuous vector, scaling it yields a family of valid restorations for tasks whose target is not unique, such as low-light enhancement. We adapt one Qwen-Image-Edit model to six tasks with a single adapter trained in about three hours on one GPU. The image instruction outperforms text conditioning under a matched comparison, and it supports task agnostic restoration without a degradation label, which the text variant does not.

Method

The degraded image reaches the editor through two channels. Its VAE encoding supplies spatial structure, while a token mapper predicts a clean-image instruction from its vision-language embedding. We keep the backbone frozen and train one shared low-rank adapter.

ImIR framework: the degraded image enters Qwen2.5-VL and a trainable token mapper for instruction, and a VAE encoder for structure. Both condition the frozen diffusion transformer with a trainable LoRA adapter. The mapper uses a residual token MLP with FiLM and global image context. — **ImIR framework.** Qwen2.5-VL encodes the degraded image; the token mapper predicts the clean-image instruction that conditions the LoRA-adapted diffusion transformer. The mapper learns a residual correction from an identity initialization.

Instruction scaling

We scale the mapper’s residual shift to move along the degraded-to-clean direction. For tasks whose target is not unique, such as low-light enhancement, varying the scale produces a family of valid restorations.

ê_x = e_y + s · g_θ(e_y; γ, β)

Increasing the scale brightens the low-light image through a range of plausible exposures. Scaling beyond the default begins to over-brighten the output.

Instruction scales = 1

Original low-light vegetation scene — Degraded input

Low-light restoration at instruction scale 1 — ImIR s = 1

Quantitative results

We compare image and text conditioning on the same backbone. The image instruction improves PSNR on all six tasks. The task-agnostic variant performs on par with the task-aware model without a degradation label.

Task	Text LoRATask-aware	Text LoRATask-agnostic	Text LoRAVLM prompt	ImIRTask-agnostic	ImIRTask-aware	Gain over Text LoRATask-aware comparison
Low-light	16.3	15.6	16.5	21.2	21.3	+5.0 dB
Deraining	32.5	17.5	17.2	33.0	33.2	+0.7 dB
Dehazing	21.0	17.2	17.8	24.4	24.7	+3.7 dB
Deblurring	26.8	17.8	17.2	27.6	27.6	+0.8 dB
Denoising	34.4	19.0	18.6	35.6	35.7	+1.3 dB
JPEG	27.4	17.0	16.6	27.9	28.1	+0.7 dB

All variants use the same backbone and training data. Task-agnostic Text LoRA uses one neutral prompt; the VLM-prompt variant uses a per-image instruction generated by Qwen2.5-VL. Task-agnostic ImIR receives only the input image.

Qualitative comparison

Citation

BibTeX

@article{aslan2026imirimageinstructiontuningallinone,
  title={{ImIR}: Image-Instruction Tuning for All-in-One Image Restoration},
  author={Süleyman Aslan and Görkay Aydemir and Mısra Yavuz and Yunus Bilge Kurt
          and Nasrin Rahimi and Ahmet Rasim Emirdağı and Burak Can Biner
          and M. Akın Yılmaz},
  journal={arXiv preprint arXiv:2609.25267},
  year={2026},
  url={https://arxiv.org/abs/2609.25267}
}

ImIRImage-Instruction Tuning for All-in-One Image Restoration