Qwen-Image-PE: Learning Visual Design for Text-to-Image Prompt Enhancement

Qwen-Image-PE. Learning Visual Design for Text-to-Image Prompt Enhancement. By Zekai Zhang and twenty-eight others. Published on the twenty-fourth of October. No code repository is linked to it.

Users ask text-to-image models for things in loose, underspecified terms. Prompt enhancement is the step that turns such a request into a coherent visual specification, and the paper argues that doing this well means learning visual design. Existing prompt enhancers preserve what the user intended and improve overall visual appeal, but they put less explicit weight on the fine-grained dimensions of design.

Qwen-Image-PE learns visual design through two stages of training, each with its own source of design signal. The first borrows from images that are already well designed. The second compares candidate images against one another. The ambition is an enhancer that does more than lengthen a prompt or make it prettier in general terms. It should make specific design decisions about how the image is lit, framed, arranged and finished, while still keeping to what the user asked for.

Stage one is supervised fine-tuning with Caption-Based Reverse Distillation, or CRD. Dense captions of well-designed images serve as the design supervision. Stage two is reinforcement learning with a Groupwise Design Reward, or GDR. For a given request, all candidate images are compared jointly, and the reward refines lighting, composition, layout and detail. The two stages divide the labour. The first gives the enhancer a picture of what good design looks like in words. The second sharpens it by judging the candidates for one request as a group, rather than one image at a time.

On Qwen-Image-Bench, using Qwen-Image-2.1 as the generator, the largest enhancer, Qwen-Image-PE-397B, improves on the original requests by seven point eight one points. The small Qwen-Image-PE-9B outperforms every prompt enhancement baseline evaluated. Further evaluations report gains in visual design, aesthetics, image quality and human preference.

The paper also finds that prompt enhancement scales along three axes: training data, model size, and information density. Sufficient information density is a prerequisite for effective enhancement. And the enhancer generalises to four additional text-to-image backbones, even surpassing the native prompt enhancers of HunyuanImage-2.1 and FLUX 3 Image Klein.

Design is treated here as something an enhancer can be taught, from captions of good images and from side-by-side judgments of candidates, rather than left implicit in a request for something beautiful. The nine billion parameter model is released on Hugging Face.