Qwen-Image-PE: two-stage training lifts text-to-image prompts by 7.81 points
Qwen · hf · 2026-10-27
The Qwen team introduces Qwen-Image-PE, which turns underspecified user requests into coherent visual specifications for text-to-image generation, explicitly learning fine-grained design dimensions.
Method highlights:
- Two-stage training: supervised fine-tuning with Caption-Based Reverse Distillation (CRD), using dense captions of curated design images as design supervision; then RL with Groupwise Design Reward (GDR) that jointly compares candidate images per request to refine lighting, composition, layout, and detail.
- Results: on Qwen-Image-Bench with Qwen-Image-2.1, the 397B version improves over original requests by 7.81 points, while the 9B version outperforms all evaluated PE baselines, with gains in visual design, aesthetics, and human preference.
- Generalization: PE effectiveness scales along training data, model size, and information density, and transfers to four additional T2I backbones including HunyuanImage and FLUX, even beating their native prompt enhancers.
More from Multimodal
- MPI's Alien Recombination method finds concept blends beyond human cognitive reach — profjamesevans · 2026-10-27
- Kling 4 generates 30-second drone-style footage, full prompt included — TheLocalLab · 2026-10-27
- Creator has Claude edit her iPhone Duo review inside Tella's agent video editor — evielync · 2026-10-27
- Curated list of 20 open-source motion libraries for building an AI motion studio — nikola_mr64990 · 2026-10-27
- Redditor builds AI fashion influencer explaining capsule wardrobes — can you tell? — Ai-Emiliaa344 · 2026-10-27
- MOSS Image Edit vs Krea 2 Edit: same prompt side-by-side shirt-swap test — Illustrious_Row_9971 · 2026-10-27