Qwen-Image-PE: Learning Visual Design for Text-to-Image Prompt Enhancement
Abstract Overview
This paper presents Qwen-Image-PE, a prompt enhancer for text-to-image generation that is trained to specify fine-grained visual design choices such as lighting, composition, layout, and detail. The method uses a two-stage pipeline: supervised fine-tuning with Caption-Based Reverse Distillation (CRD), which learns from dense captions of well-designed images, followed by reinforcement learning with Groupwise Design Reward (GDR), which scores groups of candidate images for the same request. The authors evaluate the method mainly on Qwen-Image-Bench with a frozen Qwen-Image-2.1 backbone and also study scaling with training data, model size, and information density. They further test transfer to additional text-to-image backbones to examine whether the learned prompt enhancement behavior generalizes beyond the training generator.
Novelty
The paper's main novelty is to frame prompt enhancement as learning visual design explicitly, rather than only preserving intent or optimizing generic preference. It introduces CRD to derive prompt targets from dense image captions and GDR to refine prompts through groupwise visual comparison across candidates for the same request.
Results
On Qwen-Image-Bench with Qwen-Image-2.1, Qwen-Image-PE-397B reaches an Overall score of 62.17, improving over original requests by 7.81 points, while Qwen-Image-PE-9B outperforms the evaluated prompt-enhancement baselines. The method also improves design-related metrics, UniPercept quality measures, and human preference, and both variants outperform original requests across four additional text-to-image backbones, exceeding native prompt enhancers on HunyuanImage-2.1 and FLUX 3 Image Klein.
Key Points
- Qwen-Image-PE is trained in two stages: caption-based supervised learning from well-designed images and reinforcement learning with groupwise design rewards.
- The reported gains are strongest on visual design-related outcomes, with improvements in lighting, composition, detail, aesthetics, image quality, and human preference on the benchmarked backbone.
- The scaling analysis suggests that performance increases with more data, larger models, and higher information density, and that sufficient information density is necessary for effective prompt enhancement.