Work / TRUTH · Rarely Heard Voices
An AI production studio for retail, where the product must not change
TRUTH turns image generation into a retail production process. Starting from the product imagery a retailer already has, it creates a reusable cast of AI models, tests which methods reproduce a product faithfully, produces on-model imagery, puts it in motion and publishes it to every channel. I designed the product end to end and built it: brand, explainer site, pitch, the app's UI and its front-end, on a FastAPI and Gemini backend.
Keep scrolling: the explainer site scrolls with you
01 · The ideaThe model can change. The product must not.
"Creative freedom exists around the product, not at the expense of it. That is why we call it TRUTH."
AI can make an attractive fashion image easily. Retail needs something harder: an image in which the product remains the product, in colour, print, fabric, cut, fit and construction, every time and at scale. Most AI imagery tools are a series of disconnected experiments. TRUTH is a controlled production line with five stages.
- Synthesiser. Creates consistent, reusable AI models, "Synths", that become part of a retailer's own cast and reappear across products, campaigns and channels.
- Fidelity Lab. Tests which combinations of models, instructions and settings reproduce a product faithfully, so successful approaches become repeatable recipes rather than lucky one-offs.
- Studio. A flat product plus a Synth or human model, poses, environments and art direction, producing finished on-model imagery with the product's characteristics preserved.
- Motion. Approved stills become moving content, with product and fabric information informing how the garment should behave.
- Publish. One approved source, transformed to the crops, formats and specifications of every ecommerce, marketplace and social destination.
02 · The productOne workspace, every stage
The app is a React single-page application on a FastAPI backend that calls Gemini's image models. Each module is a real route with shareable links; the sidebar, library and asset model are shared across all of them. These are the live Urban Outfitters instance at uo.rarelyheardvoices.ai, captured signed in.
03 · Product truthOne product. Six angles. Nothing drifts.
The same ivory V-neck and wide-leg trouser, generated front to back: colour, cut, fabric and fit held constant while only the pose changes.
One synth, restyled across looks. The same reusable model carried across a range: new styling and colourways without booking a shoot again.
Fidelity, before and after. The layered guarantee in practice: a tighter prompt, best-of-N generation, attribute QA on a pro model, a deterministic colour gate, then human sign-off. Real validation on retailer flats found that language-model QA alone misses gross colour drift, which is why the colour gate is deterministic.
04 · MotionFrom one product, a film. No second shoot.
The same ivory suite shown across six studio angles, now moving through a location that never had to be booked.
05 · OutcomeA product, not a demo
TRUTH runs as a multi-tenant platform with a live Urban Outfitters instance, an in-app help assistant that explains the product from real copy and cannot act, tenant reporting, and auto-deploys on every push. It has been pitched to Next as a managed image-production service, with a decision brief on AI imagery versus photography, a fidelity validation run on real retailer flats, and a pre-mortem of the risks.