Significant Image Quality & Fidelity Drop on Gemini 3 Pro (Inference/Sampling Issues)

There has been a severe degradation in output quality for Gemini 3 Pro image generation within the last 48 hours. The same high-level prompts that produced professional, crisp results previously now yield “mushy,” low-resolution outputs. This is not a prompting issue.

Key Issues Observed:

  • Loss of micro-detail: Skin textures and fine materials (like sand and jewelry) appear “airbrushed” or smudged compared to previous high-fidelity results.

  • Typography regression: Text rendering has lost its vector-like sharpness and now shows significant jagged edges and aliasing.

  • Lighting & dynamic range: Outputs feel “flat,” suggesting a reduction in the model’s sampling steps or the removal of a high-res refiner pass.

  • Consistency: This is occurring both via the Vertex AI API and the Gemini web interface, regardless of region or prompt complexity.

This appears to be linked to the ongoing Google Compute Engine (GCE) GPU reservation issues (specifically G2 instances). It seems the backend is currently using “lower-power” inference settings, reduced sampling steps or internal downscaling to maintain service availability during the capacity crunch.

Can the team confirm if “High Fidelity” or “Refiner” passes have been temporarily disabled? An ETA for when full-power inference will be restored to previous quality levels is needed.

Before (Reference for Expected Quality):

  • High-Frequency Detail: Sharp skin texture, visible pores, and realistic fine lines.

  • Material Accuracy: The luster on the pearls and the sub-surface scattering on the red beads are rendered with high physical accuracy.

  • Anatomical Precision: Crisp definition of the hands and collarbones with zero “smudging” or AI artifacts.

  • Resolution: Full-fidelity output with clean edges and professional-grade lighting depth.

After (Current Regression):

  • Texture “Smudging”: Significant loss of micro-details in skin and background; the image has an overly smoothed, “plastic” look consistent with reduced sampling steps.

  • Edge Aliasing: Visible “stair-stepping” and jagged edges on the fine jewelry chains and the “MAMA” charm.

  • Failed Text Rendering: The “LOVE” and “MAMA” text elements are blurry and lack the vector-sharpness seen in previous builds.

  • Upscaling Artifacts: The overall softness suggests the image is being generated at a lower internal resolution and then upscaled, leading to a loss of crispness in the jewelry’s metallic reflections.

System Details for Engineering Review:

  • Model: gemini-3-pro-preview

  • Endpoint Region: global

  • Date/Time of regression: Impact began approximately May 7, 2026.

Update on observed behavior:
“The quality drop is exclusive to the Vertex AI API. When using the exact same prompt and model (gemini-3-pro-preview) on the consumer Gemini app / AI Studio, the results remain high-fidelity and crisp. However, switching to the Vertex AI endpoint (Global) immediately introduces the texture ‘mushing’ and font aliasing. This confirms that the current G2 reservation outage is specifically throttling the enterprise Vertex AI infrastructure, likely by disabling high-res refiner passes to conserve capacity.”

The Vertex AI endpoint shows denoising artifacts and waxy skin, while the Gemini API retains micro-surface scattering and authentic pore definition.

Image generation by using Vertex AI endpoint;

The Vertex AI endpoint shows denoising artifacts and waxy skin, while the Gemini API retains micro-surface scattering and authentic pore definition.

Image generation by using Gemini API;

Same issue here. We’re dealing with mushy grass and vegetation since we do arch viz. But we’re struggling in the same way. Hope they fix it ASAP

Also facing the same issue, but just saw your post that it seems better as of this morning. So far I see the same on my side.

Finally kinda 10 hours ago the quality of the generated images started to show better results. Probably it was about the backend lower-power inference settings that i was guessing. But, thank you for engineers who fixed the issue.

Now this is the results that i am really happy to see these kind of generations BACK!

Best,