WeAgent-MMGenEdit: A Full-Stack Recipe for Multimodal Agentic Image Generation and Editing

To address the unreliability of existing multimodal agents in image generation and editing tasks due to the lack of verification from external knowledge, the WeAgent team proposed the WeAgent-MMGenEdit full-stack solution. This solution integrates the WeAgent-Harness multimodal runtime, a scalable data construction pipeline, and the bilingual benchmark WeBench-MMGenEdit, and employs a two-step training process based on supervised fine-tuning (SFT) and reinforcement learning (RL). Through this approach, researchers constructed a dataset containing 23,000 supervised trajectories and 14,700 reinforcement learning tasks, and trained an agent policy model with a total of 30 billion parameters and 3 billion active parameters. Experiments show that this model outperforms models of the same size in performance, approaching the level of agents with 100 billion parameters.