Generative production for brand and campaign work

Off-the-shelf models can produce beautiful images. They can’t produce your images. They don’t know your colour discipline, your lighting language, your relationship to realism. To make generation work in a brand context, the brand has to live inside the model.

Custom LoRAs, and the taste that comes first

I trained three LoRAs on the Strategy visual system. NeonPainter for high-contrast cyberpunk. MosaicPainter for the orange-cube-and-particle imagery that became Strategy’s signature. MaxiPainter for the honey-badger character work tied to the Bitcoin treasury narrative.

The hardest part of LoRA training isn’t the training. It’s the dataset. And the hardest part of the dataset is taste. Before any training run, I spend significant time in MidJourney developing the foundational style: working through hundreds of variations, narrowing the palette, locking the lighting language, defining what the brand actually looks like before anyone teaches a model to reproduce it. Without that upstream taste work, you train a model on inconsistent references and get an inconsistent model. With it, the LoRA generalises cleanly.

All three Strategy LoRAs are now officially hosted on Strategy’s Hugging Face account and are heavily used across the company.

Newer techniques

LoRA training isn’t always the right tool. Newer state-of-the-art models (Nano Banana 2 Edit in particular) can hold brand consistency through reference images alone, no training required. I’ve been experimenting with this approach for projects where the use case is short-lived or where the visual territory is still moving. When it works, it dramatically shortens the cycle from creative direction to on-brand output.

The choice between training a LoRA or using reference-based generation is now an active part of how I scope projects.

Model selection across the daily craft

Beyond the brand work, the daily craft is knowing which model to reach for.

FLUX 2 is the workhorse for ideation. Fast, predictable, four variations per prompt. FLUX 2 Pro for when the concept has settled and quality matters more than speed. Nano Banana 2 when the brief is closer to design than illustration: charts, infographics, layouts that need logic. Nano Banana 2 Edit for natural-language editing without masks.

Video as a multi-model discipline

Video is where the production complexity lives. A single shot rarely comes from a single model.

Kling V3 Pro is the cinematic workhorse. Native audio, multi-shot storyboarding, slow on purpose. The model is synthesising dozens of frames with temporal coherence, motion physics, and cinematic quality. Kling O3 when character consistency matters more than anything else. Seedance 1.5 Pro for native audio, emotional acting, and multilingual dialogue. Veo 3.1 for first-to-last-frame control where you define both endpoints. Grok Imagine for longer cinematic shots up to 15 seconds.

A typical pipeline

For a finished campaign video: a still through one of the brand-trained models or FLUX 2 Pro as start frame; Kling or Seedance turning that still into 5 to 10 seconds of motion; sometimes a Veo pass to control a transition between two specific frames; Topaz for the upscale; audio through MiniMax Music or ElevenLabs Music for score, MiniMax Speech for dialogue.

For spokesperson video, Kling Avatar lip-syncs a face image to an audio track, with audio usually generated through MiniMax TTS. End-to-end about eight minutes.

For 3D, Hunyuan 3D v3 generates downloadable GLB files with PBR materials. Useful for concepting form factors before commissioning real production assets.

Outcomes

The pipeline has produced work for the Nasdaq Times Square tower, Wall Street Journal placements, Google homepage launch creative, and across Strategy’s paid media calendar. Internally it produces tens of thousands of brand-consistent assets a year. Campaign cycles that used to take weeks of stock licensing and shoots now take days.

The aesthetic became distinctive enough that people inside and outside the company can identify Strategy imagery without seeing the logo. For a brand still establishing its new name, that’s the outcome that matters most.