AWS re:Invent 2025 - Volkswagen Group's custom GenAI models for global brand consistency (SPS313)

Volkswagen's Custom GenAI Models for Global Brand Consistency

Overview

Volkswagen Group, one of the world's largest automotive manufacturers, partnered with AWS to develop a custom generative AI (GenAI) platform to streamline their global content production and ensure brand consistency across 10 distinct vehicle brands and 150+ markets.

Challenges

• Volkswagen's massive scale: 6.6 million vehicles delivered in 9 months, with over 1 million being electrified
• Need for fast, high-quality content to capture brand loyalty in emerging markets
• Requirement for hyper-localized, data-driven marketing interventions to address challenges in established markets
• Maintaining brand identity and DNA across 10 distinct vehicle brands
• Slow, risky, and manual content production process that was unable to keep up with demand

The Solution

Volkswagen and AWS built a custom GenAI platform with two core capabilities:

1. Image Generation: To empower creative teams and accelerate content production
2. Image Evaluation: To automate compliance and brand checks at scale

Image Generation

• Leveraged fine-tuning of diffusion models on Volkswagen's proprietary image data to ensure confidentiality and brand accuracy
• Used "Dream Boost" technique to maintain flexibility and prevent overfitting
• Employed parameter-efficient fine-tuning methods like Low-Rank Adaptation (LoRA) to reduce the number of trainable parameters

Prompt Optimization

• Integrated a large language model (Amazon Nova Light) to automatically enhance user prompts with style modifiers, technical details, and composition guidance
• This removed the skill barrier for marketers and ensured consistent, high-quality output

3D Digital Twins

• Leveraged Volkswagen's partnership with Solid Meta and Univas to generate images from 3D CAD data and digital twins
• This provided infinite variations without the need for physical photo shoots, including pre-production vehicles

Image Evaluation

• Used image segmentation to break down images into individual components for detailed verification
• Leveraged vision-language models (e.g., Florence, SAM-Free) to compare generated images against multiple reference images
• Developed a brand compliance evaluation system using Amazon Bedrock and large language models to assess adherence to Volkswagen's brand guidelines

Technical Details

• Fine-tuning techniques: Dream Boost, Low-Rank Adaptation (LoRA)
• Vision-language models: Florence, SAM-Free
• Large language models: Amazon Nova Light, Claude Sonnet 4.5, Opus 4.5
• Image segmentation: Open-source Florence models hosted on Amazon SageMaker
• Brand compliance evaluation: Amazon Bedrock with custom recipes for supervised fine-tuning

Business Impact

• Massive time savings in content production and evaluation
• Increased confidence in brand compliance through automated checks
• Integrated multiple projects into a shared GenAI platform
• Significantly reduced time-to-market for marketing campaigns

Examples

• Generated images of Volkswagen vehicles in various environments (e.g., autumn, desert, city) to enable personalized, region-specific marketing
• Automatically detected minor discrepancies, such as incorrect wheel designs or license plates, to ensure complete brand accuracy
• Translated Volkswagen's brand guidelines into machine-readable rules to enable automated brand compliance evaluation

Conclusion

Volkswagen's custom GenAI platform, built in partnership with AWS, has transformed their global content production process. By leveraging advanced techniques in image generation, prompt optimization, and automated evaluation, Volkswagen has achieved significant improvements in speed, quality, and brand consistency across their 10 vehicle brands and 150+ markets.