The Artist is NOT Present

The project was carried out at a Stockholm-based creative agency that produces advertising for major commercial clients. The agency has been experimenting with generative AI since the early days of the wave and is now moving from experimentation into a structured, production-grade AI capability that can hold up to real client briefs.

Generative AI has reached a point where producing one beautiful image is essentially solved. Making the same person, character, or product appear consistently across an entire campaign, in different scenes, lighting, and formats, is not. For an agency that produces work for paying clients, that gap is the difference between using AI as a sketching tool and trusting it for production. Today several methods to achieve consistency exist, spanning from edit models to LoRA training. The goal was to find a reproducible process that could leverage the current best practices.

The project is a proof of concept for generating a defined character consistently across multiple ad scenarios. The core method stacks two techniques: a LoRA trained on reference images of the character locks identity, while a separate edit model handles wardrobe, location, and scene variation on top. Neither approach holds up alone. LoRA on its own is rigid, edit models on their own drift in hair and small facial details. Combined, they produce something usable. The pipeline lives in a node-based workflow tool so another person at the agency can pick it up, swap inputs, and run it without rebuilding from scratch.