Robust-U1
Robust-U1 is the official implementation of a unified multimodal language model that detects and repairs corrupted visual input before reasoning about it, aimed at robust visual understanding under real-world image degradation.
Robust-U1 is a unified multimodal large language model built to self-recover corrupted visual content and reason over the recovered image, released alongside the paper "Can MLLMs Self-Recover Corrupted Visual Content for Robust Understanding?" and accepted to ICML 2026.
The release includes code, pretrained models and an online demo. Model variants build on a BAGEL-7B-MoT base: a supervised fine-tuned checkpoint from stage one of training, and a reinforcement-learning checkpoint from stage two. Both a command-line demo, which takes a local model path and writes recovered images to an output directory, and a graphical demo are provided, alongside an online Hugging Face Spaces demo for trying the model without local setup.