ImIR: Image-Instruction Tuning for All-in-One Image Restoration
The paper replaces text prompts with image-derived instructions for one-model restoration.
ImIR feeds the degraded image through a VAE for structure and through a lightweight token mapper for a semantic restoration instruction. That continuous instruction can be scaled, giving multiple plausible outputs for tasks like low-light enhancement. The authors say one Qwen-Image-Edit model was adapted to six restoration tasks with a single adapter in about three hours on one GPU. In matched tests, the image instruction beat text conditioning and worked without a degradation label. HF Daily Papers' note
ImIR feeds the degraded image through a VAE for structure and through a lightweight token mapper for a semantic restoration instruction. That continuous instruction can be scaled, giving multiple plausible outputs for tasks like low-light enhancement. The authors say one Qwen-Image-Edit model was adapted to six restoration tasks with a single adapter in about three hours on one GPU. In matched tests, the image instruction beat text conditioning and worked without a degradation label. HF Daily Papers' note
score 4