A new fine-tuning method breaks AI image prompts into checkable pieces and grades each one separately, instead of giving a single blurry score to the whole picture.
Researchers built CAST, a reinforcement-learning method for fine-tuning diffusion image generators such as FLUX.2-dev and Qwen-Image-2512. Instead of relying on a human-trained scoring model that returns nearly identical scores for most images, a problem the researchers call reward saturation, CAST splits a prompt into atomic, independently checkable claims, like an object, a count, an attribute, or a spatial relation, using what the authors call Causal Scene Graphs, then scores each atom on its own. It also determines automatically, from each model's own denoising trajectory, the exact step where objects and their layout get locked in, rather than relying on a manually set window for injecting exploration noise. Those per-atom scores then get projected back onto the image through attention maps, so the training signal points at the specific region that got something wrong.
This matters because current reinforcement learning methods for diffusion models have hit a wall: top open-source models already score so well on human-trained reward models that there is barely any difference between a good and a bad image to learn from. CAST's fix, breaking one mushy reward into dozens of checkable sub-rewards, delivers up to 3.07 times the improvement of a prior method, Flow-GRPO, on the hardest compositional prompts in the GenEval 2 benchmark, using roughly the same training budget.
It is a research paper, not a shipped feature. But it points at where the next real gain in image-model quality is likely to come from: not bigger models, but better bookkeeping for the mistakes they already make.