Qwen-Image-PE teaches an image model design: prompt enhancer adds 7.81 points

The prompt "cozy café in winter" tells an image generator almost nothing. Lighting, composition, layout, and detail are left to the model, and the result is often mediocre. A Qwen team of 29 authors led by Zekai Zhang and Chenfei Wu has introduced Qwen-Image-PE, a prompt enhancer that expands such requests into a coherent design specification. On the in-house Qwen-Image-Bench benchmark, with Qwen-Image-2.1 as the generator, Qwen-Image-PE-397B improves the score of the original prompts by 7.81 points. The compact Qwen-Image-PE-9B outperforms all the baseline enhancers evaluated, and it is the one the authors have released openly.

Reverse distillation: teaching design from descriptions of finished images

Most existing enhancers preserve the user's intent and make the image more appealing overall. Fine-grained design parameters are rarely treated as a separate target. Qwen-Image-PE closes that gap with two-stage training.

The first stage is supervised fine-tuning with a method called Caption-Based Reverse Distillation (CRD). It starts with well-made images and their dense, detailed captions. These captions serve as the reference: given a short prompt, the model learns to reconstruct the same rich text that would describe a successful frame. The design knowledge comes from real work, not from abstract instructions like "make it look good."

Group reward instead of scoring one image at a time

The second stage is reinforcement learning with a Groupwise Design Reward (GDR). For a single prompt, the system generates several variants and scores all the resulting images together, in comparison with one another. The reward targets four parameters: lighting, composition, layout, and detail. Comparing within a group gives a sharper signal than scoring each frame on its own. The model sees how the successful variant differs from its neighbors for the same prompt.

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

Three scaling axes and transfer to other generators

The authors say enhancer quality grows along three axes: the amount of training data, model size, and the information density of the prompt. They describe information density as a necessary condition. If the expanded prompt is not rich enough in detail, the other two levers will not help.

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

How to use it in practice

For teams building products on image generation, the 9B version is the main news. It is realistic to run a model of that size next to the generator as a separate layer. It takes a raw user prompt and returns an expanded one, while the generator stays the same. Because the enhancer transfers to other backends, you can test it on the model you already use and compare it with that model's built-in solution on your own prompts.

The finding on information density is also worth noting. If you write prompts by hand or tune your own enhancer, the length of the expanded description and the number of concrete design decisions in it affect the result more than you might expect. It is best to verify this on your own use cases: the authors' figures come from their Qwen-Image-Bench benchmark and the Qwen-Image-2.1 generator.