Creative Product Design under Brand Image Constraints

Takeshi Onishi, Hiroya Makino, Takahiro Yamaguchi, Hiroyuki SakaiarXiv 2026 · arXiv

TL;DR

Researchers from Toyota Central R&D Labs have developed a multi-agent framework that allows AI to generate creative product redesigns without losing the "soul" of the brand. By explicitly researching a brand's identity and using it to guide an iterative self-enhancement loop, the system achieves a superior balance between visual novelty and brand recognizability compared to standard text-to-image models.

What Problem Does This Paper Address?

In industrial design, creativity isn't a free-for-all. A new car design must be innovative, but it also must be instantly recognizable as a "Toyota" or a "Porsche." While modern text-to-image (T2I) models like Flux or Stable Diffusion are incredibly creative, they lack a mechanism to respect these implicit, product-specific constraints.

Existing methods for "creative" AI often suffer from identity drift: as the model tries to make a design more unique, it gradually strips away the defining features of the brand. This paper tackles the challenge of "brand-aware redesign," where the goal is to maximize creativity while ensuring the product remains identifiable to consumers.

Methods

The authors propose a multi-agent framework that breaks the design process into three distinct modules, mimicking a professional design studio workflow:

  1. Survey Module: A "Marketer" agent uses web search (via OpenAI's Deep Research) to gather data on the product's history, market environment, and target customers. This is summarized into a Brand Image Survey Report, turning implicit brand values into explicit text.
  2. Creative Direction Module: A "Brand Manager" agent takes the survey report and a set of random "inspiration images" to derive a specific design strategy. It selects the direction that best appeals to the target customer while preserving the brand's essence.
  3. Self-Enhancement Module: This is an iterative loop between a Designer agent (which writes prompts for the T2I model, Flux.1-dev) and a Reviewer agent. The Reviewer critiques the generated images based on the specific "Creative Direction" and brand constraints, forcing the Designer to refine the prompt over 15 iterations.
Figure 1: Overview of the proposed framework.

Datasets

The framework was evaluated on a newly constructed benchmark of 50 redesign tasks across five diverse categories:

  • Automotive: (e.g., Toyota Prius, Tesla Model 3) - High structural and brand constraints.
  • Motorcycle: (e.g., Honda Super Cub) - High structural constraints.
  • Beverage: (e.g., Coca-Cola, Red Bull) - Constraints focused on graphics and packaging.
  • Character Series: (e.g., Pokémon, Star Wars) - Constraints based on fictional world consistency.
  • Automobile Manufacturer Logos: (e.g., BMW, Audi) - Highly compact semantic constraints.

Key Insights and Ideas

  • Textualizing the Implicit: The core "aha" moment is that brand image is too complex to be captured by a simple prompt or a fixed latent embedding. By dynamically generating a textual report, the system makes the brand's "vibe" actionable for the AI reviewer.
  • The Identity Drift Trap: Without brand constraints, iterative refinement (Self-Enhancement) actually hurts brand identity. As shown in the study, a Pokémon character might become more "creative" but eventually turns into a generic monster that no longer looks like a Pokémon.
  • Constraint-Driven Creativity: The framework proves that constraints can actually guide creativity. In highly restricted categories like cars, the "Creative Direction" module helps the AI find a "plausible" path for innovation that doesn't break the brand rules.
Figure 4: Creativity and Identity Scores during the self- enhancement loop in character_series. Solid lines represent the proposed framework, and dashed lines rep- resent Self-Enhancement. The images at the top show gener- ated samples for Pokémon at the corresponding iterations.

Key Results

The authors used a VLM-as-Judge protocol (using Gemini 2.5 Pro) to score images on Creativity (Novelty + Usefulness) and Identity (how accurately the VLM could guess the brand).

  • Superior Trade-off: The proposed framework achieved a significantly better balance than baselines. While unconstrained Self-Enhancement had the highest creativity, its identity scores plummeted. The proposed method maintained high identity scores (comparable to naive T2I) while significantly boosting creativity.
  • Category Wins: The gains were most pronounced in the Automotive and Motorcycle categories, where the design space is narrow. In these areas, all other baselines struggled to be creative without making the product unrecognizable.
Figure 2: Creativity Score and Identity Score with ex- ample outputs from the proposed framework and Self- Enhancement.

Limitations and Future Work

  • VLM vs. Human Judgment: The study relies on AI models to judge creativity and identity. While VLMs are becoming more aligned with humans, they may not capture the subtle nuances of professional brand aesthetics or consumer perception.
  • Narrow Constraint Scope: The framework currently only focuses on brand image. Real-world industrial design must also satisfy "hard" constraints like aerodynamics, safety regulations, and manufacturing feasibility, which this system does not yet verify.
  • Future Directions: The authors suggest extending the framework to include these engineering constraints, potentially creating a more holistic AI assistant for professional designers.

Summary updated

Abstract

Industrial design requires creative visual exploration while satisfying constraints that are often product-specific and implicit. Existing text-to-image (T2I) approaches can generate creative designs through concept blending or multi-agent self-enhancement, but they provide limited mechanisms for ensuring that generated designs remain consistent with such constraints. We focus on brand image as a representative design constraint and propose a framework for brand-aware redesign generation. Given a product name, the framework constructs a brand image survey report that summarizes product and market information, derives a creative direction from the report, and integrates this direction into a self-enhancement loop in which generated images are critiqued and prompts are revised. We evaluate the framework on 50 redesign tasks across five product categories, comparing it with Naive T2I, IP-Adapter/IT-Blender, and unconstrained self-enhancement under the same T2I backbone. By a VLM-as-Judge protocol to measure creativity and brand identifiability, the proposed framework achieves a more favorable trade-off between the two objectives, with particularly clear gains in highly constrained categories such as automotive and motorcycle, where the plausible design space is relatively limited compared with that of other product categories. These results suggest that dynamically eliciting and textualizing brand image can guide creative product design generation under brand image constraints.
BibTeX
@misc{onishi2026creative,
  author = {Takeshi Onishi and Hiroya Makino and Takahiro Yamaguchi and Hiroyuki Sakai},
  title = {{Creative Product Design under Brand Image Constraints}},
  year = {2026},
  eprint = {2611.17762},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.17762},
}
Figure