UniGen-1.5: Enhancing Image Generation and Editing through Reward Unification in Reinforcement Learning

TL;DR

UniGen-1.5 advances unified multimodal learning by integrating image understanding, generation, and editing within a single model and optimizing both generation and editing through a unified reinforcement learning framework with shared rewards. A lightweight Edit Instruction Alignment stage enhances editing instruction comprehension, enabling more reliable RL signals. Empirical results show competitive GenEval and ImgEdit performance, surpassing several open models and approaching proprietary baselines, while maintaining strong understanding capabilities. The work presents a scalable baseline for unified MLLMs with practical implications for controllable visual generation and editing in real-world applications.

Abstract

We present UniGen-1.5, a unified multimodal large language model (MLLM) for advanced image understanding, generation and editing. Building upon UniGen, we comprehensively enhance the model architecture and training pipeline to strengthen the image understanding and generation capabilities while unlocking strong image editing ability. Especially, we propose a unified Reinforcement Learning (RL) strategy that improves both image generation and image editing jointly via shared reward models. To further enhance image editing performance, we propose a light Edit Instruction Alignment stage that significantly improves the editing instruction comprehension that is essential for the success of the RL training. Experimental results show that UniGen-1.5 demonstrates competitive understanding and generation performance. Specifically, UniGen-1.5 achieves 0.89 and 4.31 overall scores on GenEval and ImgEdit that surpass the state-of-the-art models such as BAGEL and reaching performance comparable to proprietary models such as GPT-Image-1.

Figures (6)

• Figure 1: Examples of images generated by UniGen-1.5.
• Figure 2: The architecture of UniGen-1.5 jointly optimized for (a) image understanding, (b) text-to-image generation and (c) image editing. See more details in \ref{['sec:method_arch']}.
• Figure 3: Illustration of Edit Instruction Alignment in the entire training pipeline of UniGen-1.5.
• Figure 4: Left: The pipeline of GRPO training in UniGen-1.5. We utilize shared reward models for both text-to-image generation and image editing. For the former, we directly input the generated image with the text prompt to obtain rewards. For the latter, we get reward signals by measuring the alignment between the edited image description and the generated image. Right: The pipeline of edited image description estimation. We leverage powerful external MLLMs and LLMs to generate the description of desired edited images.
• Figure 5: Examples generated by UniGen-1.5, highlighting the contribution of GRPO training.