Qwen-Image-PE Teaches Neural Networks Design: Prompt Enhancer Adds +7.81 Points

The prompt "cozy winter café" for an image generator is almost nothing. Lighting, composition, frame layout, and details are left entirely to the model, and the result often ends up mediocre. The Qwen team — 29 authors led by Zekai Zhang and Chenfei Wu — has introduced Qwen-Image-PE, a prompt enhancer that fleshes out such queries into a coherent design specification. On the internal Qwen-Image-Bench benchmark, using the Qwen-Image-2.1 generator, the Qwen-Image-PE-397B version improves the score of original queries by 7.81 points. The compact Qwen-Image-PE-9B outperforms all evaluated baseline enhancers, and it's this version the authors have made publicly available.

Reverse

Distillation: Learning Design from Descriptions of Finished Images

Most existing enhancers preserve the user's intent and make the resulting image generally more appealing. Subtle design parameters, however, are rarely singled out as a distinct objective. Qwen-Image-PE closes this gap through two-stage training.

The first stage is supervised fine-tuning using the Caption-Based Reverse Distillation (CRD) method. Well-crafted images and their dense, detailed descriptions serve as the reference material. The model learns to reconstruct, from a short query, text as rich as what would describe a successful shot. Design knowledge comes from real work, not from abstract instructions like "make it look nice."

Group Reward Instead of Single-Image Scoring

The second stage is reinforcement learning using the Groupwise Design Reward (GDR). For a single query, several variants are generated, and all resulting images are evaluated jointly, in comparison with each other. The reward targets four parameters: lighting, composition, layout, and detail. Comparing within a group produces a clearer signal than scoring each image in isolation. The model sees how a successful variant differs from its neighbors generated from the same query.

Additional checks showed gains in visual design, aesthetics, image quality, and human preference.

Three Axes of Scaling and Transfer to Other Generators

The authors claim the enhancer's quality improves along three axes: volume of training data, model size, and prompt information density. Information density is described as a necessary condition. If the expanded query isn't sufficiently rich in detail, the other two levers won't help.

Another finding concerns transferability. Qwen-Image-PE generalizes to four additional T2I architectures. On HunyuanImage-2.1 and FLUX 3 Image Klein, it outperforms those models' own native enhancers.

Key facts:

• Model: Qwen-Image-PE, a prompt enhancer for text-to-image generation.
• Authors: 29 researchers, arXiv paper 2610.33472.
• Stage 1: SFT with Caption-Based Reverse Distillation (CRD) on dense descriptions of quality images.
• Stage 2: RL with Groupwise Design Reward (GDR), jointly comparing all candidates for a single query.
• Target parameters: lighting, composition, layout, detail.
• Qwen-Image-PE-397B: +7.81 points over original queries on Qwen-Image-Bench with Qwen-Image-2.1.
• Qwen-Image-PE-9B: outperforms all evaluated PE baselines, released publicly.
• Scaling axes: data, model size, information density.
• Transfer: four additional T2I backends, including HunyuanImage-2.1 and FLUX 3 Image Klein, where the model outperformed native enhancers.
• Release: huggingface.co/Qwen/Qwen-Image-2.1-PE-T2I.

How to Apply This in Practice

For teams building products on image generation, the key news is the 9B version. It's a model you can realistically keep alongside your generator as a separate layer. It takes a raw user query and returns an expanded prompt, while the generator itself stays unchanged. Since the enhancer transfers to other backends, you can test it with whatever model you're already using and compare it against that model's built-in solution on your own queries.

The finding about information density is also worth considering. If you're writing prompts manually or tuning your own enhancer, the length of the expanded description and the number of concrete design decisions it contains affect results more than you'd expect. It's best to verify this on your own scenarios: the authors' figures come from their own Qwen-Image-Bench benchmark using the Qwen-Image-2.1 generator.