Phase A shipped — fixing design drift in AI jewelry generationFollowing up on the audit post from last week.

JewelViz generates 3 shots per showcase (closeup, mid-length, extreme closeup), each as an independent AI call. With no shared reference beyond the raw image, jewelry details would drift between shots — different stone counts, shifted motifs, sometimes an extra bangle appearing that wasn't in the original photo.
What I shipped: A Gemini 2.5 Flash analysis step that runs once per generation — reads the uploaded photo, outputs a structured description (metal, stones, motifs, items present), and feeds that same locked description into all 3 generation calls as a hard constraint.
Honest results from testing so far: Necklaces and pendants — clear improvement, design stays much more consistent across shots. Complex asymmetric pieces like multi-tier earrings — still seeing drift in some cases (missing tiers, occasionally an extra accessory added). Haven't run enough volume yet to put a real number on the improvement — didn't want to post inflated stats before I have actual data.
Next — Phase B: reference chaining, where each shot uses the previous shot's output as an additional reference, not just the original photo. Should help most with the earring/complex-jewelry cases.
jewelviz.com