Phase C shipped — fixing the extreme closeup shot
Following up on Phase A (design locking) and Phase B (model/outfit consistency).
Problem: The extreme closeup shot is supposed to be a pure product macro — jewelry alone on velvet. Without a hard constraint, the model would sometimes drift into generating skin, fingers, or wedding/mandap elements that had nothing to do with the original product photo.
Fix: Rewrote the extreme closeup prompt and added a productOnly flag threaded through the entire generation pipeline, with explicit negative constraints (no human, no props, no wedding elements). Tested across a batch of showcases — now consistently a clean catalog-style macro.
Next — Phase D: hardening the prompts across all three shots with stronger negative constraints.
jewelviz.com
I tried JewelViz after reading your post, and I really liked how straightforward the flow is from upload to generation.
One thing that stood out to me wasn't the generation quality itself—it was the moment before clicking "Generate." With credits involved, I found myself wishing for a bit more confidence about how much each setting would influence the final result. The more confident users feel before spending credits, the more willing they'll be to experiment.
Really appreciate this — you've put your finger on something I'd been sensing but hadn't articulated this clearly. The pre-generate confidence gap is a real problem, especially with credits on the line.
We already run a Gemini analysis step under the hood that reads the jewelry piece before generation. Thinking of extending that to surface 2-3 setting suggestions before you hit Generate, based on the actual piece — so it's not a blind choice.
One question, if you don't mind: are you a jeweller yourself, or were you testing this more as a curious user? Would help me understand whether the confidence gap matters more for business decisions (spending real budget) or just general experimentation.
Thanks! I'm not a jeweller — I was testing it as a curious user with a product/QA mindset.
I think the confidence gap matters in both cases, just for different reasons. For a business user it's about not wasting credits or time on the wrong settings. For someone like me, it's more about understanding what effect each option will have before committing to a generation. I really like the idea of surfacing a few AI-powered suggestions before generation. That feels much more like collaborating with the tool instead of guessing.