Volkswagen Group (10 brands, 6.6M vehicles delivered in the first nine months of 2025) partnered with the AWS Generative AI Innovation Center to tackle a very grown-up marketing problem: making thousands of brand-correct images without paying six figures for every photo shoot or waiting weeks for post-production and approvals. The output is an end-to-end system that generates photorealistic vehicle imagery and then automatically grades it—both for “is this the right car?” details and for “does this look like our brand?” nuance—using Amazon SageMaker AI for generation and Amazon Bedrock for evaluation.
On the creation side, the team found that off-the-shelf diffusion models could make “a Volkswagen-ish car,” but fell apart on the details that actually get marketing teams yelled at: grille textures, headlight geometry, wheel patterns by trim, and anything involving unreleased models. So they fine-tuned with DreamBooth using training data sourced from digital twins in NVIDIA Omniverse, then served a Flux.1-Dev model with a LoRA adapter on SageMaker. They also added automated prompt tuning with Amazon Nova Lite, because “silver VW in a forest” is not a prompt—it’s a cry for help when you need camera angles, lighting, and brand styling baked in.
The real magic is quality control at scale. Instead of relying on blunt image metrics, the pipeline segments vehicles into components (wheels, grille, headlights, logos, etc.) using Florence-2 on SageMaker, verifies segment labels with Nova Lite, and then has Claude 4.5 Sonnet (via Bedrock) score each component against reference images with a 1–5 rubric and explain what’s off (down to “internal headlight structure looks too detailed”). Separately, Claude evaluates brand guideline compliance—including regional gotchas like a UK localization that accidentally used a German-style plate and got dinged 2/5—because authenticity can be ruined by one tiny rectangle of metal.
Finally, they started customizing Nova Pro for brand evaluation without begging the marketing team for thousands of labels: they generated 1,000 “compliant” and 1,000 “noncompliant” prompts from the guidelines, created synthetic training pairs, and ran supervised fine-tuning on SageMaker—setting up a playbook that could be repeated across all ten brands.
“By combining our domain expertise with AWS, we built a generative AI platform that makes our marketing faster, smarter, and safer.” — Sebastian Angersbach, Head of IT Strategy & Innovation, Volkswagen Group Services.
Read more at AWS Blog.