A research team built an image-generating AI that marks its own work, then trains on the correction - no bigger teacher model needed.
The method, called UniEvo-VL, takes one multimodal model and splits it into two roles during training. A "student" version only sees the original prompt, while a "teacher" version sees the same prompt plus a critique of an earlier attempt. Training pulls the student's image-generation process toward what the teacher would have produced, using the model's own self-critique as the extra signal. Applied to the open-source Qwen-image-2512 model, the approach lifted its GenEval benchmark score from 0.747 to 0.808, and a related text-accuracy measure (GenEval2 Soft-TIFA) from 32.97 to 35.53.
That's notable because the usual way to make a model better at judging and fixing its own output is to bolt on a bigger, separate teacher model, which costs more compute and more engineering. Here, one model supervises itself by comparing its plain answer against its own annotated one. The researchers also tested swapping in a stronger outside critic, GPT5.6-Luna, and found performance improved further, suggesting the ceiling for this kind of self-improvement rises with the quality of the critic, not just more training data.
The catch: gains weren't uniform, and text rendering inside generated images improved unevenly - a reminder that self-taught models tend to sharpen skills they already half-have, not invent new ones.