We current UniGen-1.5, a unified multimodal massive language mannequin (MLLM) for superior picture understanding, technology and enhancing. Constructing upon UniGen, we comprehensively improve the mannequin structure and coaching pipeline to strengthen the picture understanding and technology capabilities whereas unlocking robust picture enhancing capacity. Particularly, we suggest a unified Reinforcement Studying (RL) technique that improves each picture technology and picture enhancing collectively by way of shared reward fashions. To additional improve picture enhancing efficiency, we suggest a lightweight Edit Instruction Alignment stage that considerably improves the enhancing instruction comprehension that’s important for the success of the RL coaching. Experimental outcomes present that UniGen-1.5 demonstrates aggressive understanding and technology efficiency. Particularly, UniGen-1.5 achieves 0.89 and 4.31 general scores on GenEval and ImgEdit that surpass the state-of-the-art fashions akin to BAGEL and reaching efficiency corresponding to proprietary fashions akin to GPT-Picture-1.
- † Institute of Reliable Embodied AI, Fudan College
- ‡ Mission lead
- § Corresponding authors
