Enhancing product concept image generation through semantic feature prompts and LoRA training.

Area of Science:

• Computer Science
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Current Generative Artificial Intelligence (GAI) models struggle with precise product conceptual image design.
• Limitations exist in controlling the output of large models for specific design applications.

Purpose of the Study:

• To propose an innovative strategy integrating fine-grained semantic feature decoding with Low-Rank Adaptation (LoRA) fine-tuning.
• To significantly improve text-to-image technology performance for product conceptual design.
• To provide a solution for controlled generation in large models for product design.

Main Methods:

• Utilized E-Prime software for semantic priming to extract key product design words.
• Employed DeepSeek prompt engineering to decode semantic features across mental, functional, and physical image dimensions.
• Applied Low-Rank Adaptation (LoRA) technique for independent dataset training based on expert-derived semantic feature prompts.

Main Results:

• Demonstrated the strategy's application in conceptual design using an intelligent pulse diagnostic instrument example.
• Conducted multi-dimensional assessments and comparative experiments to verify efficacy.
• Achieved optimal model configuration through LoRA fine-tuning based on semantic feature prompts.

Conclusions:

• The proposed strategy effectively enhances text-to-image generation for product conceptual design.
• The integration of semantic decoding and LoRA offers a viable solution for controlled GAI in design.
• Validated potential and efficacy through empirical testing and case study.