The Interesting Part of Qwen-Image-2.0-RL Is Not the Image Score

Qwen's new image paper is easy to read as another benchmark bump. Qwen-Image-2.0-RL takes the existing Qwen-Image-2.0 model, runs a reinforcement-learning pass on top, and reports better scores: 57.84 on Qwen-Image-Bench, up 2.61 points from the base model. Its text-to-image arena Elo moves from 1115 to 1193. Its image-editing arena Elo moves from 1256 to 1349. Those are the headline numbers. They are not the useful part. The useful part is the training story underneath them. The paper is a good reminder that "just optimize the reward" is a dangerously incomplete sentence, especially when the model is not an LLM and the output space is a whole image. The model got better, but not by one simple trick Qwen-Image-2.0-RL is a post-training pipeline for a diffusion image model. In plain English: the base model already knows how to generate and edit images. The RL stage tries to steer it toward outputs humans prefer, including better prompt following, better aesthetics, better portrait fidelity, and more reliable editing. The team builds task-specific reward models. For text-to-image, those rewards cover alignment, aesthetics, and portrait quality. For editing, they cover instruction following and face identity preservation. Then they train with a GRPO-style setup adapted for flow-matching diffusion models. If you only squint at that, it sounds like the same broad recipe people use for language models: generate candidates, score them, push the model toward the better ones. The paper is more interesting because it shows how fragile that story becomes once you touch the actual training loop. The CFG detail is the first real lesson Classifier-free guidance, usually shortened to CFG, is one of those diffusion-model knobs that users mostly experience as "make the image follow the prompt harder." Under the hood, it changes how the model samples. The Qwen team tested three ways to use it during RL. Using CFG during both rollout and training made the images collapse into incohere