Over the last few days I noticed a major drop in image quality specifically inside Vertex AI Studio / Gemini API.
Previously, the same prompts produced sharp, realistic images with detailed skin texture and natural-looking faces. Now the outputs look blurry, over-smoothed, and plastic-like.
The biggest issue is facial skin quality:
waxy/plastic skin
loss of texture and detail
missing pores and natural skin texture
very soft blurry rendering even in 4K
Important:
I tested Nano Banana / gemini-3.1-flash-image-preview on other platforms and the quality there is noticeably better. The issue seems specific to Vertex AI Studio / Google Cloud.
Using:
the same prompts
the same settings
the same workflow
But image quality suddenly became much worse over the last few days.
Is anyone else experiencing this?
Did something change in the backend pipeline, compression, routing, or inference infrastructure?
+1, confirming the same regression on gemini-3.1-flash-image-preview and Pro via Vertex AI starting around 2026-05-07.
Our use case is virtual try-on for fashion / clothing, so face and
skin quality is load-bearing for us. What you describe matches
exactly:
Waxy, over-smoothed skin
Loss of pores and fine texture
Faces look “filtered” — like beauty mode is on by default
Even at imageSize: “4K” the output is softer than identical
prompts produced a week ago
What we held constant when comparing:
Identical prompt strings (no edits in the codebase)
Identical reference inputs (model photo + garment + studio bg)
Identical generation config;
Identical model ID: gemini-3.1-flash-image-preview
Same Vertex location (“global”)
Side-by-side comparison with our archived outputs from late
April / early May shows a clear, consistent quality drop across
many subjects — not a per-prompt fluke.
We’ve had to ship a fallback path to a different provider in the
meantime because stakeholders started flagging “plastic faces” in
review. We’d strongly prefer to keep Vertex as the primary if/when
this is rolled back.
Same here. We are seeing a consistent quality degradation with gemini-3-pro-image-preview for real-estate image enhancement workflows.
The main issue is not overall resolution. The model is returning true 4K native PNG outputs, but foliage, bushes, and grass often look noticeably softer / blurrier / more painterly than before, while architecture remains relatively sharp.
I guess some finish passes for upscaling are not working properly on google’s end. Hope they solve it quick..
Hello I’m facing the same problem, I made a architectural rendering process, the users get the screen capture using the software, and send to nano banana 2 ou NB pro, and the results looks likes blurrier. When you zoom up the wood floor, or any part of the image, looks like softer
Confirming on our side as well. Outputs are back to the expected quality. Thanks.
Out of curiosity, can you share what was actually going on?
Would help us know what to watch for if something similar shows up again.
