Qwen-Image-2.0-RL Technical Report Details Gains From Reinforcement Learning

Alibaba's Qwen team applied reinforcement learning post-training to Qwen-Image-2.0. Combining on-policy distillation with reward models for alignment, aesthetics, and portrait fidelity, it reached an Arena Elo of 1,193 for text-to-image generation, up 78, and 1,349 for image editing, up 93. Its overall Qwen-Image-Bench score rose 2.61 points to 57.84. The model is closed source.

Visit original source