We’re running a production image generation platform and facing two consistent, reproducible quality issues with gemini-3.1-flash-image-preview:
Issue 1 — Generation Drift: After 5–10 iterative re-edits (each using the previous output as the new base image), image quality degrades significantly — loss of sharpness, color distortion, and structural artifacts compound with each pass. Is this a known model limitation? Are there any API parameters to mitigate quality cascading across iterative edits?
Issue 2 — Time-of-Day Quality Variance: The same prompt + same input produces noticeably worse outputs during afternoon/evening IST (peak US hours). During off-peak hours quality is significantly better. This seems like a server load issue on shared preview infrastructure. Is there a timeline for this to stabilize? Is there a region with more consistent capacity?
We are on a paid billing account. Happy to share prompt examples and output comparisons. Other developers seem to report the same — wanted to make this visible officially.