M3: High-fidelity Text-to-Image Generation via Multi-Modal, Multi-Agent and Multi-Round Visual Reasoning
Published in arXiv preprint, 2026
Recommended citation: Bangji Yang, Ruihan Guo, Jiajun Fan, Chaoran Cheng, Ge Liu. "M3: High-fidelity Text-to-Image Generation via Multi-Modal, Multi-Agent and Multi-Round Visual Reasoning." arXiv preprint arXiv:2602.06166, 2026. https://arxiv.org/abs/2602.06166
M3 is a training-free, multi-agent loop for compositional text-to-image generation: a Planner turns the prompt into a verifiable checklist, Checker, Refiner and Editor agents fix one constraint at a time, and a Verifier keeps every round an improvement. Applied to open-source models, Qwen-Image with M3 reaches state of the art on OneIG-EN, ahead of commercial systems such as Imagen 4 and Seedream 3.0.