ImIR: Task-Agnostic Image Restoration

ImIR: Image-Instruction Tuning for All-in-One Image Restoration
ACCV 2026

Süleyman Aslan · Görkay Aydemir · Mısra Yavuz · Yunus Bilge Kurt
Nasrin Rahimi · Ahmet Rasim Emirdağı · Burak Can Biner · M. Akın Yılmaz
Codeway AI Research

The output of task-agnostic ImIR is produced from the input image alone, with no text prompt and no degradation label. A single adapter supports low-light enhancement, deraining, dehazing, deblurring, denoising, and JPEG artifact removal.

Model

suleymanaslan/imir is the fixed 1MP task-agnostic ImIR model. This repository provides its image-instruction mapper, LoRA adapter, and inference configuration. The ImIR inference package loads these components together with Qwen-Image-Edit-2511. The backbone is downloaded separately and cached automatically.

Setting	Value
Backbone	Qwen/Qwen-Image-Edit-2511
Backbone revision	6f3ccc0b56e431dc6a0c2b2039706d7d26f22cb9
Restoration canvas	Target area of 1,048,576 pixels
Resize policy	target_area
Sampling steps	30
Instruction scale	1.0
Output resolution	Original input resolution by default

The input is resized toward the target area, allowing both upscaling and downscaling. Aspect ratio is preserved up to alignment of the canvas dimensions to multiples of 16. The canvas is not necessarily square. The restored image is resized back to the original input resolution by default.

Files

File	Contents
imir_config.json	Mapper architecture, backbone revision, and inference defaults.
mapper.safetensors	Image-instruction mapper weights.
pytorch_lora_weights.safetensors	Restoration LoRA weights.

All three files are required. Configuration is loaded directly from the checkpoint, with no separate configuration directory.

Usage

Use the ImIR inference package to load the mapper and LoRA together. Loading only the LoRA into a standard Diffusers pipeline does not reproduce ImIR inference.

Use Python 3.10 or newer and a CUDA-enabled PyTorch build. Install the package from the ImIR inference repository:

git clone https://github.com/suleymanaslan/imir.git
cd imir
python -m pip install -e .

Command line

Run from the inference repository to use its included deraining sample:

imir --checkpoint suleymanaslan/imir \
  --input examples/deraining.png \
  --output outputs/deraining.png \
  --cpu-offload \
  --seed 0

This writes the restored image and a JSON file containing inference settings and image sizes. CPU offloading moves model components between CPU and GPU and requires a CUDA-capable GPU and sufficient system RAM.

For another image, replace --input with its path. Use --steps or --scale to override the checkpoint defaults, or --keep-generation-size to retain the restoration canvas resolution. No task label or text prompt is required.

Python

from pathlib import Path
from imir import ImIRPipeline

pipe = ImIRPipeline.from_pretrained(
    "suleymanaslan/imir",
    cpu_offload=True,
)
result = pipe("input.png", seed=0)

Path("outputs").mkdir(exist_ok=True)
result.image.save("outputs/restored.png")
print(result.metadata)

The checkpoint and its pinned backbone revision are downloaded on first use. You can also pass a local directory containing the three checkpoint files. For a reproducible model version, pass revision="<model-commit-sha>" to from_pretrained, or --revision <model-commit-sha> to the CLI.

Acknowledgments

ImIR uses Qwen-Image-Edit and Hugging Face Diffusers. See NOTICE for attribution.

License

The ImIR mapper and LoRA weights are released under the MIT License. The separately downloaded backbone and inference dependencies retain their own licenses. See NOTICE for image attribution.

Citation

@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}
}