Why the spiral hinge defeats generative imagery — measured, not argued.
A controlled production test on frame 10300570, run before answering the RFI. Every output shown, including every rejected one. Transparency about what did not survive our quality gate is part of the method.
The measurement
No current generative model resolves a specific mechanical geometry at twelve pixels. It falls back on the statistical average of all eyewear hinges — and the average has a screw. This is not a prompt problem; it is a scale problem, and it applies to every provider in this process equally.
The evidence — target vs. generated
The test — six generations, six rejections
Frame 10300570, both required angles, three generations each. Five of the six are shown below. The sixth failed before the hinge even became the question: it broke the locked model identity and collapsed the framing into an eye macro — rejected on sight, and further proof that consistency must be engineered, not hoped for. References supplied per shot: frontal packshot, three-quarter packshot, dedicated joint macro, plus a locked model identity matching your brief (age ~38, natural skin, no beauty smoothing). The mechanism's geometry was additionally described explicitly in text. The verdict standard is binary — "close enough" is a rejection.
Extension — a second construction type
To test whether the finding is specific to the steel frame, we ran the same brief on frame 10084715 — a translucent acetate construction whose hinge, embedded visibly in the material, spans roughly 200 pixels on-model instead of twelve. The result confirms the thesis from the other direction: give the mechanism an order of magnitude more pixels, and the generation gets an order of magnitude closer.
This is the scale law in action, and it cuts both ways: it explains precisely why the twelve-pixel spiral cannot be generated, and it confirms that the failure is a property of detail size, not of any particular frame. The quality gate stays binary either way — close is a rejection.
What held
- Skin and age register. Real texture, fine lines, freckles — no artificial smoothing, exactly the 35–40 aesthetic your brief specifies.
- Identity consistency. One model, recognisably the same person across all six generations.
- Background discipline. Flat light grey, no gradient, no visible seam — lockable to one exact value across the catalogue.
- Wearing physics. Nose-pad contact, temple pressure, cast shadows, lens reflection — credible throughout.
- No invented branding. No fabricated logos, wordmarks or engravings anywhere.
Conclusion
The person, the light and the scene are a solved problem. The product is not — so the product must never be left to a generative step.
Our production approach composites the frame's actual geometry from your own data into the generated wearing context, so the mechanism's pixels originate with you, not with a model. What this requires is source material worthy of the detail: camera-original masters instead of web derivatives, one sharp joint macro per frame family — and ideally CAD data, which removes the angle dependency entirely and is the prerequisite for the 360° views and virtual try-on in your project overview.
Why we show rejected work. A provider who only shows successes is asking you to trust a highlight reel. We would rather you see the failure mode precisely, because the provider who can name it is the one who has engineered around it. Ask every vendor in this process for this same measurement.