Recent advances in image generation models have expanded their applications beyond aesthetic imagery toward practical visual content creation. However, existing benchmarks mainly focus on natural image synthesis and fail to systematically evaluate models under the structured and multi-constraint requirements of real-world commercial design tasks. In this work, we introduce BizGenEval, a systematic benchmark for commercial visual content generation. The benchmark spans five representative document types: slides, charts, webpages, posters, and scientific figures, and evaluates four key capability dimensions: text rendering, layout control, attribute binding, and knowledge-based reasoning, forming 20 diverse evaluation tasks. BizGenEval contains 400 carefully curated prompts and 8000 human-verified checklist questions to rigorously assess whether generated images satisfy complex visual and semantic constraints. We conduct large-scale benchmarking on 26 popular image generation systems, including state-of-the-art commercial APIs and leading open-source models. The results reveal substantial capability gaps between current generative models and the requirements of professional visual content creation. We hope BizGenEval serves as a standardized benchmark for real-world commercial visual content generation.
Figures & tables
Figure 1: Real-world samples of BizGenEval from 5 content domains and 4 capability dimensions. For each capability dimension, different subtasks are highlighted with blue boxes. Each example covers distinct aspects, with the relevant regions highlighted in red and the corresponding simplified questions listed below.
Figure 2: Overview of the construction and evaluation pipeline. The system converts real world references and domain knowledge into structured prompts, and evaluates generated images using rigorous checklists to provide a comprehensive analysis.
Figure 3: Dataset Statistics of BizGenEval . (a) Prompt token length distribution by evaluation dimension, with scatter points color-coded by document type. (b) Hierarchical subcategory distribution across evaluation dimensions. The inner ring shows question proportions per dimension; the outer ring shows subcategory breakdowns. (c) Checklist Questions keyword clouds per evaluation dimension.
Model
Slides
Webpage
Poster
Chart
Sci. Figure
Average
Nano-Banana-Pro [ 12 ]
82.2 / 94.8
77.5 / 96.5
76.5 / 94.8
73.0 / 92.2
74.2 / 90.0
76.7 / 93.7
Nano-Banana-2.0 [ 13 ]
73.8 / 95.8
71.2 / 94.5
67.5 / 91.2
60.2 / 89.2
69.5 / 92.0
68.5 / 92.5
Seedream-5.0 [ 30 ]
54.5 / 80.8
47.0 / 80.8
50.7 / 77.0
46.0 / 76.2
45.8 / 81.5
48.8 / 79.2
GPT-Image-1.5 [ 27 ]
40.8 / 89.2
41.0 / 86.0
42.0 / 83.5
28.2 / 76.5
27.8 / 72.8
35.9 / 81.6
Seedream-4.5 [ 29 ]
33.8 / 71.0
36.2 / 75.5
35.5 / 72.2
18.0 / 47.8
27.3 / 64.8
30.1 / 66.2
Wan2.6-T2I [ 39 ]
27.1 / 56.5
25.5 / 67.0
27.5 / 62.5
17.2 / 48.8
12.5 / 58.8
21.9 / 58.7
Table 1: Performance of Image Generative Models Across Five Commercial Content Domains. Closed-source and open-source models are marked with purple and blue , respectively. Each cell reports scores on the hard/ easy testsets.
Figure 4: Qualitative evaluation of different commercial image generation models across five content domains. Columns represent content domains, and rows show model outputs. Evaluation questions are listed in the top row, with correct and incorrect regions highlighted in blue and red boxes.
Model
Layout
Attribute
Text
Knowledge
Average
Nano-Banana-Pro [ 12 ]
72.2 / 91.2
65.6 / 92.2
86.4 / 95.0
82.6 / 96.2
76.7 / 93.7
Nano-Banana-2.0 [ 13 ]
68.4 / 91.0
57.4 / 91.6
83.4 / 94.6
64.6 / 93.0
68.5 / 92.5
Seedream-5.0 [ 30 ]
67.6 / 89.0
42.4 / 77.2
43.4 / 75.6
41.8 / 75.2
48.8 / 79.2
GPT-Image-1.5 [ 27 ]
51.6 / 84.8
25.8 / 75.2
40.4 / 82.8
26.0 / 83.6
35.9 / 81.6
Seedream-4.5 [ 29 ]
35.4 / 71.6
22.4 / 62.8
41.4 / 72.4
21.4 / 58.2
30.1 / 66.2
Wan2.6-T2I [ 39 ]
46.4 / 80.6
16.6 / 60.6
12.6 / 52.6
12.2 / 41.0
21.9 / 58.7
Table 2: Performance of Image Generative Models Across Four Capability Dimensions.
Figure 5: Qualitative evaluation of different commercial image generation models across four capability dimensions.
Model
GenEval [ 10 ]
OneIG-EN [ 2 ]
BizGenEval
HunyuanImage-3.0 [ 1 ]
0.72
-
13.0 / 40.1
GPT-Image-1.0 [ 26 ]
0.84
0.533
11.2 / 52.4
Z-Image [ 38 ]
0.84
0.546
8.2 / 43.8
Qwen-Image [ 42 ]
0.87
0.539
2.8 / 23.8
LongCat-Image [ 35 ]
0.87
-
0.7 / 12.0
Table 3: Cross-benchmark comparison.
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Example Task: Knowledge-based Reasoning. Assessment of a model’s ability to incorporate domain knowledge into visual generation.
Figure 7: Example Task: Attribute Binding. Evaluation of precise associations between specific visual elements and their described properties.
Figure 8: Example Task: Layout Control. Measurement of spatial arrangement and structural organization within professional design contexts.
Figure 9: Example Task: Text Rendering. Focus on the model’s capacity for legible, accurate, and typographically coherent text generation.