Generative production
The pipeline behind most of Strategy's image and video work: custom brand models, multi-model video, and a lot of art direction before any training starts.
Project
- Role
- Art director, prompt engineer, model trainer
- Client
- Strategy
- Year
- 2025
Tools
- Midjourney
- FAL
- Hugging Face
- Adobe
Off-the-shelf models make beautiful images, but not your images. For generation to work for a brand, the brand has to live inside the model.
Brand models
I trained three LoRAs on the Strategy visual system: NeonPainter for high-contrast cyberpunk, MosaicPainter for the orange cube imagery, and MaxiPainter for the honey badger character. Each started in Midjourney, working through hundreds of variations to settle palette and lighting before building a dataset. A consistent dataset is what makes a model hold up. All three are hosted on Strategy’s Hugging Face account and used across the company.
For short-lived work I often skip training and use reference images with models like Nano Banana 2 Edit. Choosing between the two is now part of scoping each project.
Video
A finished shot rarely comes from one model. A typical sequence: a still from a brand model or FLUX 2 Pro, animated in Kling or Seedance, sometimes a Veo pass between two fixed frames, Topaz for upscaling, and music and voice from MiniMax or ElevenLabs.
The work has run on the Nasdaq tower in Times Square, in the Wall Street Journal, on Google’s homepage and across Strategy’s paid media. People now recognise it as Strategy without seeing a logo.