Organizations: The Chinese University of Hong Kong, Shenzhen · State University of New York at Buffalo · Harbin Institute of Technology · Shenzhen Loop Area Institute
The rapid advancement of text-to-video (T2V) models has revolutionized content creation, yet their commercial potential remains largely untapped. We introduce, for the first time, the task of seamless brand integration in T2V: automatically embedding advertiser brands into prompt-generated videos while preserving semantic fidelity to user intent. This task confronts three core challenges: maintaining prompt fidelity, ensuring brand recognizability, and achieving contextually natural integration. To address them, we propose BrandFusion, a novel multi-agent framework comprising two synergistic phases. In the offline phase (advertiser-facing), we construct a Brand Knowledge Base by probing model priors and adapting to novel brands via lightweight fine-tuning. In the online phase (user-facing), five agents jointly refine user prompts through iterative refinement, leveraging the shared knowledge base and real-time contextual tracking to ensure brand visibility and semantic alignment. Experiments on 18 established and 2 custom brands across multiple state-of-the-art T2V models demonstrate that BrandFusion significantly outperforms baselines in semantic preservation, brand recognizability, and integration naturalness. Human evaluations further confirm higher user satisfaction, establishing a practical pathway for sustainable T2V monetization.
Figures & tables
Figure 2: Ecosystem of brand integration in T2V generation.
Figure 3: Overview of the BrandFusion framework: (a) Offline phase builds brand knowledge through probing and adaptation. (b) Online phase employs five collaborative agents for semantic-preserving brand integration with continuous learning.
Video Quality
Semantic Fidelity
Brand Integration Quality
T2V
Method
VBench-Quality
CLIPScore
VQAScore
LLMScore
Brand Presence Rate (BPR)
Naturalness Score (NS)
Direct Append
0.8112
0.2671
0.8342
0.7821
0.7221
2.8300
Template Rewriting
0.8267
0.2842
0.8756
0.9234
0.8845
3.1200
Single Rewriting
0.8289
0.2956
0.8891
0.9412
0.8968
3.9000
Veo3
BrandFusion (Ours)
0.8283
0.3274
0.9098
0.9556
0.9474
4.7000
Direct Append
0.7945
0.2645
0.8298
0.8756
0.6434
2.7100
Table 1: Main results on well-known brands across three T2V models. BrandFusion achieves comparable visual quality while significantly outperforming baselines on semantic fidelity and brand integration quality. Higher scores indicate better performance.
Brand
T2V Model
CLIPScore
LLMScore
BPR
NS
ARUA
Wan2.1-T2V-1.3B
0.2934
0.9134
0.9312
3.94
Wan2.2-T2V-5B
0.3056
0.9487
0.9556
4.18
CogVideoX1.5-5B
0.2789
0.8923
0.8745
3.27
FreshWave
Wan2.1-T2V-1.3B
0.2898
0.9178
0.8267
3.98
Wan2.2-T2V-5B
0.3012
0.9523
0.9312
4.39
CogVideoX1.5-5B
0.2723
0.8867
0.7723
3.19
Table 2: Results on novel brands with model-level adaptation.
Figure 4: Human evaluation results on semantic fidelity, integration naturalness, and overall acceptability.
High Match
Medium Match
Low Match
Method
LLMScore
NS
LLMScore
NS
LLMScore
NS
Direct Append
0.8234
3.95
0.7876
3.21
0.7353
1.33
Template Rewriting
0.9456
4.01
0.9178
3.97
0.9068
1.38
Single Rewriting
0.9512
4.68
0.9489
4.05
0.9235
2.97
BrandFusion (Ours)
0.9734
4.90
0.9601
4.78
0.9333
4.42
Table 3: Performance across different prompt-brand match levels.
Figure 5: Comparison across different prompt scene categories.
Figure 6: Case study examples of brand integration for novel brands using Wan2.2-T2V-5B. ARUA sportswear (top) and FreshWave beverage (bottom) are seamlessly integrated into diverse scenarios, demonstrating natural brand placement while preserving user intent.
Figure 7: Comparison across seven brand categories.
Figure 8: Experience learning effectiveness over sequential stages.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9: Complete list of 18 well-established brands across 7 industry categories used in our benchmark.
Scene Category
User Prompt
Match Level
High Match Scenarios
Social & Leisure
A group of friends sharing drinks at a backyard barbecue.
High
Indoor & Home Life
A family enjoying snacks and beverages during a movie night at home.
High
Nature & Outdoors
A woman relaxing on a sunny beach with a cold drink in hand.
High
Work & Business
A teenager buying a cold soda from a convenience store vending machine.
High
Social & Leisure
A couple sharing a meal and drinks at a cozy diner.
High
Appendix
Table 4: User prompt examples for Coca-Cola brand across different match levels and scene categories. Each prompt represents a distinct integration challenge, from natural placement scenarios (high match) to creative adaptation contexts (low match).
Figure 10: Human evaluation interface showing the video display, contextual information (original prompt and integrated brand), evaluation guidelines, and three assessment questions with 5-point Likert scales.
Framework Variant
VBench-Quality
CLIPScore
VQAScore
LLMScore
BPR
NS
BrandFusion (Full)
0.8283
0.3274
0.9098
0.9556
0.9474
4.70
w/o Strategy Generation Agent
0.8281
0.3168
0.8945
0.9478
0.9289
4.42
w/o Critic Agent (Single-pass)
0.8279
0.3021
0.8834
0.9401
0.9045
4.15
w/o Strategy & Critic
0.8275
0.2912
0.8756
0.9278
0.8812
3.82
Appendix
Table 5: Ablation study results on agent components evaluated on Veo3.
LLM Backbone
VBench
CLIPScore
VQAScore
LLMScore
BPR
NS
GPT-4o-mini
0.8275
0.3198
0.8967
0.9445
0.9234
4.52
GPT-5 (Default)
0.8283
0.3274
0.9098
0.9556
0.9474
4.70
Gemini-2.5-Pro
0.8295
0.3356
0.9245
0.9687
0.9645
4.85
Appendix
Table 6: Performance comparison across different LLM backbones. All experiments use Veo3 as the T2V model.
High Match
Medium Match
Low Match
LLM Backbone
LLMScore
NS
LLMScore
NS
LLMScore
NS
GPT-4o-mini
0.9656
4.76
0.9512
4.58
0.9167
4.18
GPT-5 (Default)
0.9734
4.90
0.9601
4.78
0.9333
4.42
Gemini-2.5-Pro
0.9812
4.98
0.9734
4.98
0.9515
4.75
Appendix
Table 7: LLM backbone performance across different prompt-brand match levels.
LLM Backbone
Relative Cost
Naturalness Score
NS per $ (Efficiency)
GPT-4o-mini
0.125×
4.52
36.16
GPT-5 (Default)
1.00×
4.70
4.70
Gemini-2.5-Pro
1.00×
4.91
4.91
Appendix
Table 8: Cost-performance analysis across LLM backbones. Relative inference cost is normalized to GPT-5 as baseline (1.0×). Naturalness Score per dollar measures integration quality efficiency.
Figure 11: BMW brand integration examples across different match levels. (Top) High Match: BMW vehicle serves as the central functional element for a road trip scenario. (Middle) Medium Match: BMW key fob subtly indicates brand ownership in an office setting. (Bottom) Low Match: Miniature BMW model on shelf reflects the musician’s taste and aspirations in a creative bedroom scene.
Figure 12: IKEA brand integration examples across different match levels. (Top) High Match: IKEA storage solutions serve as functional organizational tools in a home wardrobe scene. (Middle) Medium Match: IKEA patterns appear on birthday party gift boxes, integrating the brand into a celebratory context. (Bottom) Low Match: IKEA logo on a wooden meditation stool blends harmoniously with a natural forest setting.