Learning visual design is essential for prompt enhancement (PE) to turn underspecified user requests into coherent visual specifications for text-to-image (T2I) generation. Existing methods preserve user intent and improve overall visual appeal, with less explicit emphasis on fine-grained design dimensions. To address this gap, we introduce Qwen-Image-PE, which learns visual design through two-stage training. Supervised fine-tuning with Caption-Based Reverse Distillation (CRD) uses dense captions of well-designed images as design supervision. Reinforcement learning with Groupwise Design Reward (GDR) jointly compares all candidate images for the same request to refine lighting, composition, layout, and detail. On Qwen-Image-Bench with Qwen-Image-2.1, Qwen-Image-PE-397B improves over original requests by 7.81 points, while Qwen-Image-PE-9B outperforms all evaluated PE baselines. Further evaluations show gains in visual design, aesthetics, image quality, and human preference. We find that PE scales along three axes: training data, model size, and information density, with sufficient information density as a prerequisite for effective enhancement. Moreover, we demonstrate that Qwen-Image-PE generalizes to four additional T2I backbones, even surpassing native prompt enhancers on HunyuanImage-2.1 and FLUX 3 Image Klein. We release Qwen-Image-PE-9B at https://huggingface.co/Qwen/Qwen-Image-2.1-PE-T2I.
Figures & tables
Figure 2: Overview of Qwen-Image-PE. (a) Supervised fine-tuning with Caption-Based Reverse Distillation (CRD) trains the enhancer to imitate design choices described in dense captions of well-designed images. (b) Reinforcement learning with Groupwise Design Reward (GDR) refines lighting, composition, layout, and detail by jointly scoring images for the same request, using the full rollout group.
Figure 3: Motivating Caption-Based Reverse Distillation. (a) Dense captions of well-designed images guide Qwen-Image-2.1 to reproduce key visual design choices in lighting, composition, layout, and detail. (b) In the t-SNE projection, Qwen-Image-PE-397B prompt embeddings overlap more with embeddings of sampled T2I training captions than forward-based prompt embeddings do.
Model
Evaluation Dimension
Overall
Quality
Aesthetics
Alignment
Real-world
Creative
Fidelity
Generation
GPT Image 2
58.65
67.53
65.85
57.38
75.23
64.69
Nano Banana 2.0
54.77
61.08
62.40
54.28
67.05
59.82
GPT Image 1.5
55.14
60.88
61.72
53.95
66.35
59.65
Nano Banana Pro
55.67
60.26
61.25
54.07
66.23
59.45
Table 1: Results on Qwen-Image-Bench. Prompt enhancers share a frozen Qwen-Image-2.1 backbone. Bold marks the best score among methods using this backbone. † marks our internal model. Qwen-Image-PE-397B leads across all five dimensions, while Qwen-Image-PE-9B surpasses all PE baselines.
Model
QIB Design-related Capabilities
UniPercept
HPSv3 ↑
PickScore ↑
Lighting
Composition
Layout
Detail
IAA
IQA
ISTA
w/o Prompt Enhancer
Original Request
64.12
54.68
54.88
55.15
66.28
72.21
43.71
6.3773
22.2699
Closed-Source API-Based Prompt Enhancers
Qwen-3.8-Max
71.02
62.68
60.34
56.73
70.17
74.34
46.64
8.5632
23.2987
Gemini-3.1-Pro
70.81
61.78
60.16
57.23
70.38
74.05
45.49
9.4122
23.5107
Table 2: Evaluation of visual design and aesthetics. Bold and underlined scores indicate the best and second-best results per column, respectively. † marks our internal model. Qwen-Image-PE-397B ranks first on eight of nine metrics, while Qwen-Image-PE-9B outperforms trained PE baselines on all metrics.
Figure 5: Qualitative comparisons across four design dimensions. In these examples, prompts from Qwen-Image-PE-397B yield more expressive lighting, more cohesive compositions, better-organized layouts, and richer details than original requests or prompts from baseline enhancers.
Figure 6: Benchmark performance during RL training. The curves show benchmark evaluation scores, not training rewards. Qwen-Image-PE-397B improves across design dimensions on Qwen-Image-Bench, with the most consistent gains in Lighting. Dashed lines indicate linear trends.
Method
Quality
Aesthetics
Alignment
Real-world
Creative
Overall
Fidelity
Generation
(a) Supervised Fine-Tuning
T2I backbone: Qwen-Image-2.1
Original Request
55.09
54.47
54.59
51.38
54.47
54.36
+ Forward-Based Supervision
55.98
60.34
61.77
54.00
67.70
59.86
+ Caption-Based Reverse Distillation
57.10
63.34
62.33
55.82
70.12
61.52
Table 3: Ablation of the two training stages. For Qwen-Image-PE-397B, Caption-Based Reverse Distillation outperforms Forward-Based Supervision across all five dimensions. Groupwise Design Reward further improves Overall over both the SFT initialization and Pointwise Design Reward.
Method
JS ↓
FD ↓
MAUVE ↑
Unigram
Bigram
Qwen-3.8-Max
0.229
0.539
0.0709
0.032
Qwen-Image-PE-9B
0.111
0.347
0.0221
0.221
Qwen-Image-PE-397B
0.105
0.329
0.0207
0.414
Table 4: Prompt–caption proximity. Lower JS and FD and higher MAUVE indicate closer distributions.
Figure 7: Scaling properties of PE. PE scales with (a) training data, (b) model size, and (c) information density. Sufficient information density is a prerequisite for effective prompt enhancement.
Method
Quality
Aesthetics
Alignment
Real-world
Creative
Overall
Fidelity
Generation
T2I backbone: FLUX.2-dev
Original Request
50.84
53.21
51.97
43.84
47.59
50.52
+ Qwen-Image-PE-9B
51.77
57.38
54.21
45.51
51.71
53.05
+ Qwen-Image-PE-397B
52.17
58.16
55.85
47.00
55.03
54.38
T2I backbone: Krea 2 Turbo
Table 5: Generalization behavior of PE. Both PE variants outperform original request on four T2I backbones, surpassing native enhancers on HunyuanImage-2.1 and FLUX 3 Image Klein.