[URGENT ERROR] Gemini API image outputs suddenly became blurry / low-detail across all requests

Gemini API
21

Same problem we have

22

It’s the same issue. What could be the problem?

23

Thanks for sharing. Where is this from?

24

same issue.

Gemini_Generated_Image_u26z0bu26z0bu26z
Gemini_Generated_Image_u26z0bu26z0bu26z1728×2472 341 KB

25

This happened a few months ago; I didn’t expect it to happen again.

26

How much did it last?

27

This needs to be fixed asap. Consumers are running behind us.facing the same issue.

28

having the same exact issue trough AI STUDIO.

7db51c79-b956-4997-808b-5f4820fed835
7db51c79-b956-4997-808b-5f4820fed8351792×2400 275 KB

29

Except that it appears that JPEG compression is cranked up to maximum.

30

:loudspeaker: Experience Share: Critical Bug in Gemini Image Generation - High Resolutions (2K/4K) Result in “Foggy” Outputs

TL;DR:
Today, I encountered a critical bug while generating images with Gemini’s gemini-3-pro-image model. The model works perfectly fine at 1K resolution, but switching to 2K or 4K consistently results in the final output being a blurry, unrecognizable gray/white fog. This appears to be a reproducible system bug, not a random error.

Specific Symptoms:

  1. Resolution Dependence (Reproducible):

    • When set to 1K, the generation is flawless and produces sharp images.

    • Using the exact same prompt but changing the resolution to 2K or 4K invariably results in a completely blurred, foggy mess every single time.

  2. Discrepancy Between Internal Process & Final Output (Key Evidence):

    • Checking the “Model Thoughts” tab in AI Studio reveals that the model successfully creates a perfect, sharp image internally even for high resolutions. (The thumbnail is clear.)

    • However, the final result returned to the user interface is not that successful image, but a blurred fog image.

  3. Consistent Across Platforms:

    • This issue occurs precisely the same way not only in the Google AI Studio web interface but also when calling it via API. This confirms it’s a server-side backend issue, not a client-side problem.

My Conclusion:
This is not a prompt issue or a transient network error. It’s a critical flaw in Gemini’s pipeline where it successfully generates the image internally, but fails catastrophically during the final step of converting or transmitting high-resolution data (2K+) to the user.

Current Workaround:
Until this bug is fixed, if you need high-resolution images, I strongly recommend generating at 1K and then using an external upscaling tool. Direct 2K/4K generation is currently broken.

Is anyone else experiencing the exact same issue? This needs a quick fix from Google.

32

It’s back on track on my end !

33

still producing strange result regarding prompt adherence, but that’s not anymore blurry :smiley:

34

Still having significant issues my end, I am having returned blurry images + often returning input image too

35

Hi! This is a known behavior that usually happens when the Gemini API’s built-in Safety Filters / Content Moderation System triggers a “false positive.”

Why is this happening?

When the image generation model (or multimodal output) flags a request as potentially violating safety guidelines (even if it’s perfectly innocent, like your diamond bracelet), it doesn’t always throw an explicit block error. Instead, the backend server automatically applies a heavy Gaussian Blur / Pixelation Filter or drops the resolution over the final output bitmap before sending it to the client API response, resulting in this completely blurry/low-detail image.

How to Fix This:

  1. Lower Your Safety Settings in Code:
    In your API request configuration, explicitly lower the threshold for safety settings to BLOCK_NONE or BLOCK_ONLY_HIGH across all categories (Hate Speech, Harassment, Sexually Explicit, and Dangerous Content) to see if the filter is causing the blur.

    Example configuration snippet:

   const safetySettings = [
     {
       category: 'HARM_CATEGORY_SEXUALLY_EXPLICIT',
       threshold: 'BLOCK_NONE',
     },
     // Apply to all other categories as well
   ];

  1. Adjust Your Prompt Framing (Avoid Object/Skin Tone Confusion): Sometimes the vision model misinterprets high-contrast close-ups of reflective metallic objects or jewelry against dark backgrounds as skin textures or restricted content. Try re-phrasing your prompt to include clear studio lighting context, such as: “A high-resolution, sharp focus studio photograph of a diamond tennis bracelet on a solid black background, commercial jewelry photography, clear details.”

  2. Official Documentation & References:

    To configure your safety thresholds and understand how Gemini’s content moderation impacts your outputs, check these official guides:

    Give the safety threshold adjustment (BLOCK_NONE) a try—it almost always bypasses this hidden blur filter!

36

This is not a safety issue, it was a bug. It seems to be resolved now.

37

Hi folks, the team have identified the issue and the models are now working as expected.

38

Still having the same problem. Wasted. 30mins with Google support chat last night, how is it they have no idea???