InstructIR: High-Quality Image Restoration Following Human Instructions **Authors:** Marcos V. Conde, Gregor Geigle, Radu Timofte **Affiliation:** Computer Vision Lab, CAIDAS & IFI, University of Würzburg; Sony PlayStation, FTG **GitHub:** https://github.com/mv-lab/InstructIR **Abstract:** Image restoration is a fundamental problem that aims to recover high-quality images from degraded observations. In this work, we present InstructIR, the first approach that uses human-written instructions to guide image restoration models. Given natural language prompts, InstructIR can recover high-quality images from their degraded counterparts, considering multiple degradation types. Our method achieves state-of-the-art results on several restoration tasks, including image denoising, deraining, deblurring, dehazing, and low-light image enhancement. InstructIR improves over previous all-in-one restoration methods by +1dB. Our dataset and results represent a novel benchmark for text-guided image restoration and enhancement. **Contributions:** - First approach to use human-written instructions to guide image restoration. - Achieves state-of-the-art performance on various image restoration tasks. - Generalizes to restoring images using arbitrary human-written instructions. - Single all-in-one model covers more tasks than many previous works. **Related Work:** - Image restoration methods focus on specific degradations or use general neural networks for diverse tasks. - All-in-one image restoration models tackle multiple degradation types and levels using a single deep blind restoration model. - Text-guided image manipulation methods use text prompts to guide image generation and editing. **Method:** - InstructIR consists of an image model (NAFNet) and a text encoder. - The text encoder maps user prompts to fixed-size vectors, enabling task-specific transformations within the model. - Task routing techniques are used to condition the features based on the text embedding. **Experimental Results:** - Extensive qualitative and quantitative results demonstrate the effectiveness of InstructIR on various image restoration tasks. - InstructIR outperforms previous methods in multiple tasks, achieving competitive results. - Ablation studies show the importance of text guidance and task routing. **Conclusion:** InstructIR is a powerful all-in-one model that uses human-written instructions to guide image restoration. It achieves state-of-the-art results and represents a novel benchmark for text-guided image restoration.InstructIR: High-Quality Image Restoration Following Human Instructions **Authors:** Marcos V. Conde, Gregor Geigle, Radu Timofte **Affiliation:** Computer Vision Lab, CAIDAS & IFI, University of Würzburg; Sony PlayStation, FTG **GitHub:** https://github.com/mv-lab/InstructIR **Abstract:** Image restoration is a fundamental problem that aims to recover high-quality images from degraded observations. In this work, we present InstructIR, the first approach that uses human-written instructions to guide image restoration models. Given natural language prompts, InstructIR can recover high-quality images from their degraded counterparts, considering multiple degradation types. Our method achieves state-of-the-art results on several restoration tasks, including image denoising, deraining, deblurring, dehazing, and low-light image enhancement. InstructIR improves over previous all-in-one restoration methods by +1dB. Our dataset and results represent a novel benchmark for text-guided image restoration and enhancement. **Contributions:** - First approach to use human-written instructions to guide image restoration. - Achieves state-of-the-art performance on various image restoration tasks. - Generalizes to restoring images using arbitrary human-written instructions. - Single all-in-one model covers more tasks than many previous works. **Related Work:** - Image restoration methods focus on specific degradations or use general neural networks for diverse tasks. - All-in-one image restoration models tackle multiple degradation types and levels using a single deep blind restoration model. - Text-guided image manipulation methods use text prompts to guide image generation and editing. **Method:** - InstructIR consists of an image model (NAFNet) and a text encoder. - The text encoder maps user prompts to fixed-size vectors, enabling task-specific transformations within the model. - Task routing techniques are used to condition the features based on the text embedding. **Experimental Results:** - Extensive qualitative and quantitative results demonstrate the effectiveness of InstructIR on various image restoration tasks. - InstructIR outperforms previous methods in multiple tasks, achieving competitive results. - Ablation studies show the importance of text guidance and task routing. **Conclusion:** InstructIR is a powerful all-in-one model that uses human-written instructions to guide image restoration. It achieves state-of-the-art results and represents a novel benchmark for text-guided image restoration.