Visual content creation tasks demand a nuanced understanding of design conventions and creative workflows-capabilities challenging for general models, while workflow-based agents lack specialized knowledge for autonomous creative planning. To overcome these challenges, we propose VisionCreator, a native visual-generation agentic model that unifies Understanding, Thinking, Planning, and Creation (UTPC) capabilities within an end-to-end learnable framework. Our work introduces four key contributions: (i) VisGenData-4k and its construction methodology using metacognition-based VisionAgent to generate high-quality creation trajectories with explicit UTPC structures; (ii) The VisionCreator agentic model, optimized through Progressive Specialization Training (PST) and Virtual Reinforcement Learning (VRL) within a high-fidelity simulated environment, enabling stable and efficient acquisition of UTPC capabilities for complex creation tasks; (iii) VisGenBench, a comprehensive benchmark featuring 1.2k test samples across diverse scenarios for standardized evaluation of multi-step visual creation capabilities; (iv) Remarkably, our VisionCreator-8B/32B models demonstrate superior performance over larger closed-source models across multiple evaluation dimensions. Overall, this work provides a foundation for future research in visual-generation agentic systems.
Figures & tables
Figure 1: Human Evaluation results on VisGenBench-Image and VisGenBench-Video.
Figure 4: VisionAgent framework for dataset generation.
Figure 5: VisGenData-4k dataset statistics.
Figure 6: Our Native VisionCreator framework.
Figure 7: Comparison of the real environment and our virtual VisGenEnv environment, with an example of using a video generation tool.
Figure 8: LtrReward Components.
Type
Content
Content
Object
Scene
Style
Variety
Visual
Video
Video
Creative
Match
Consistency
Consistency
Consistency
Amount
Duration
Storyboard
Image Tasks
50
50
50
50
50
50
100
–
–
Video Tasks
50
50
50
50
50
50
100
200
200
Total
100
100
100
100
100
100
200
200
200
Table 1: Test dataset composition of VisGenBench, with 400 image tasks and 800 video tasks.
Method
Creative
Match
Object
Scene
Style
Variety
Amount
Duration
Storyboard
S-Rate
O-Score
GPT-5
0.683
0.641
0.593
0.579
0.638
0.232
0.620
0.263
0.660
0.863
0.577
Gemini2.5-Pro
0.777
0.802
0.625
0.602
0.573
0.345
0.540
0.376
0.700
0.933
0.627
Qwen3-VL-8B-Tk
0.104
0.078
0.100
0.065
0.109
0.014
0.160
0.034
0.040
0.142
0.085
VisionCreator-8B
0.651
0.661
0.645
0.638
0.595
0.211
0.480
0.429
0.580
0.925
0.581
Table 2: Comparisons on VisGenBench by VLM Evaluation. S-Rate: Success Rate, O-Score: Overall Score. The best and second-best results are highlighted.
Model
Success Rate
Human Evaluation
Overall Score
Image
Video
Image
Video
GPT-5
95.95%
93.00%
3.52
3.25
3.19
Gemini2.5-Pro
91.00%
84.00%
3.53
3.35
3.01
Qwen3-VL-32B-Thinking
97.00%
93.00%
3.47
3.23
3.18
Qwen3-VL-32B-RL
91.00%
87.00%
3.51
3.40
3.07
Qwen3-VL-32B-SFT
96.00%
94.00%
3.53
3.37
3.27
Table 3: Comparisons on VisGenBench by Human Evaluation. All models use the new version system prompt, which differs from Tab. 2 . Overall Score = (Success Rate of Image × Human Evaluation of Image + Success Rate of Video × Human Evaluation of Video) / 2. The performance comparisons of all detailed human evaluation dimensions are shown in Fig. 1 .
Method
Creative
Match
Object
Scene
Style
Variety
Amount
Duration
Storyboard
S-Rate
O-Score
RL1
0.534
0.817
0.694
0.547
0.579
0.249
1.000
0.397
0.625
0.904
0.634
RL2
0.579
0.808
0.677
0.479
0.558
0.265
0.800
0.478
0.875
0.942
0.644
RL3
0.671
0.674
0.621
0.622
0.555
0.217
0.800
0.513
0.750
0.885
0.631
RL4
0.573
0.794
0.672
0.696
0.569
0.150
1.000
0.534
0.625
0.925
0.654
v1
0.000
0.050
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.019
0.007
v2
0.230
0.334
0.382
0.339
0.396
0.163
0.600
0.134
0.500
0.490
0.357
Table 4: Ablation study with VisionCreator-8B on VisGenBench-104 comparing different training strategies. VisGenBench-104 is a sampled subset of VisGenBench. Model configurations: RL1 : PST + Result+Format reward); RL2 : PST + Plan×(Result+Format) reward; RL3 : Qwen3-VL + Plan×(Result+Format) reward; RL4 : PST + Plan×Fine reward; v1 : 3×VisGenData-4k; v2 : 3×VisGenData-4k + General-1M; v3 : 20×VisGenData-4k + General-1M; v4 : PST + 3×VisGenData-4k + General-1%; v5 : PST + 3×VisGenData-4k + General-5%; v6 : PST + 3×VisGenData-4k + General-10%; v7 : PST + 3×VisGenData-4k + General-20%.
Figure 9: Visualization comparisons of consistency.
Model
Semantic
Style
Emotion
Subject
Design
Visual
Text
Creativity
Overall
Matching
Matching
Matching
Consistency
Integrity
Integrity
Quality
GPT-5
3.4883
3.6214
2.9656
3.6024
3.4408
3.4218
2.7565
3.4408
3.3458
Gemini2.5-Pro
3.3943
3.5399
2.8119
3.4034
3.2214
3.2669
2.7300
3.3215
3.2123
Qwen3-VL-32B-Tk
3.4435
3.7248
2.9876
3.5890
3.4047
3.4823
2.8130
3.4726
3.3659
Qwen3-VL-32B-SFT
3.3504
3.8016
2.8896
3.7632
3.4368
3.5040
2.8224
3.5232
3.3888
VisionCreator-32B
3.6432
3.8412
3.1581
3.7620
3.4452
3.6531
2.8809
3.5739
3.4947
Table 5: Human Evaluation of Detailed Dimensions on VisGenBench-Image (Score = Success Rate × Human Evaluation Score)
Category
Name
Quantity
NLP
DeepSeek-R1-Distill-110k Liu (2025)
110k
LONGCOT-Refine-500K PowerInfer Team (2025)
500k
alpaca-gpt4-data Peng et al. (2023a) ; Peng et al. (2023b)
100k
Multimodal
M3IT Li et al. (2023)
1592k
Tool Calling
function-calling-chatml Glaive AI (2023)
112k
xlam-function-calling-60k Zhang et al. (2024)
60k
Table 7: General-purpose Datasets.
Video Generation Tasks
Image Generation Tasks
No.
Task Type
No.
Task Type
Product marketing videos
Product images
Public service advertisements
Detail pages
Corporate promotion videos
Key Visual (KV)
Brand story videos
Landing pages / H5 graphics
Event promotion videos
Complete brand visual identity
Table 8: Task Distribution in VisGenData-4k Dataset