The headline is about Qwen-Image-2.0-RL and its real innovation being the post-training methodology, not the image score.

Short summary

Qwen-Image-2.0-RL's real innovation isn't the benchmark gains (+78 Elo for text-to-image) but the post-training methodology: selective CFG use prevents training collapse, timestep-focused optimization blocks reward hacking, and on-policy distillation merges task-specific teachers into one model. The approach teaches how reward shaping works in diffusion models differently than LLMs. Most valuable insight: production constraints (one deployable model, no policy router) drive training architecture choices.

• •Selective CFG and timestep sampling prevent reward hacking and model collapse in diffusion post-training
• •On-policy distillation simplifies deployment: merges specialized teachers into one model with no runtime policy router
• •Gains of +78 Elo meaningful but secondary to the training recipe, which reveals how diffusion post-training differs from LLM post-training

Generated with AI, which can make mistakes.