Qwen-Image-2.0-RL Improves Visual Quality via GRPO-based RLHF

Qwen-Image-2.0-RL uses a post-training pipeline combining RLHF and on-policy distillation to enhance diffusion model instruction-following and aesthetics. The framework utilizes a scalable GRPO-based reinforcement learning approach with composite reward models covering alignment, portrait fidelity, and identity preservation.