We run an AI product-photography service and generate studio packshots from a single uploaded product photo. The vast majority of generations are clean, but we’ve isolated a consistent failure mode that only appears under a specific combination of conditions, and we’d like help understanding the root cause.
Repro conditions (all three needed to trigger):
-
A geometrically intricate product — fine jewelry chains, thin openwork/script lettering, tiny bezels and connectors, i.e. lots of thin high-frequency detail.
-
High output resolution (2K / 4K).
-
A plain white or seamless background request.
When any one of these is relaxed (lower res, or a textured/scene background, or a simpler product), output quality is normal. It’s the intersection that breaks.
Observed failure types, in rough order of frequency:
-
Soft / mushy fine detail: the hero element renders correctly but thin structures (box-chain links, the join between a pendant and its bezel) come out smeared, as if the model stopped resolving high-frequency geometry and interpolated instead.
-
Partial render (“half a job”): part of the frame is finished to spec while another region is left unresolved — soft blobs, or a stray ghost artifact (e.g. a skin-toned smudge) where the model seems to have abandoned the region.
-
Full collapse: occasionally the product dissolves entirely into a blurred color field and is unrecognizable.
What it looks like: the model behaves as if it ran out of “effort”/compute budget on the parts of the image that needed the most work — and the large flat white background plus high pixel count seems to make that worse rather than easier, which is counterintuitive (we’d expect a plain background to be the easy case).
Questions for the team / community:
-
Is there a known interaction between high target resolution and large low-information (flat white) regions that can cause the model to under-allocate to the detailed subject?
-
Does increasing resolution increase the likelihood of premature
STOP/ incomplete-region behavior on detailed inputs specifically? (We’ve seen the related “complex prompt → no image /No image data found” reports, but ours returns an image — just an unfinished or degraded one.)
