NO_IMAGE — Gemini API returned no image (Nano Banana finishReason)

Published 2026-09-23

finishReason Retry depends on cause

The exact error

NO_IMAGE

You asked a Gemini image model for a picture and the candidate ended without one. NO_IMAGE is Google’s way of saying the model failed to produce the image it was supposed to, as opposed to a filter deciding it should not.

What it means

The Gemini API reference defines NO_IMAGE in the FinishReason enum as: the model was expected to generate an image, but none was generated. Google’s API errors page lists the Interactions API equivalent, no_image, under Generation error codes, the group for structural problems with the model’s output. It is not in the Generation blocked codes group, where image_safety, image_prohibited_content and image_other sit. That placement is the key distinction: officially, NO_IMAGE is a generation failure, not a policy block.

Google’s Vertex page on Gemini image generation limitations describes the behaviors that lead to a missing image, without naming a finish reason for each:

The model might only create text and no image if the prompt is ambiguous. Google’s advice is to ask for images clearly.
The model might stop generating before it is finished. Google says to try again or use a different prompt.
If a prompt is potentially unsafe, the model might reply that it cannot create unsafe images, and in that case the FinishReason is STOP.

That last point matters for agent code: a text-only answer with STOP can be a refusal, and it will not show up as a block code.

Merged variant: "finishReason": "OTHER" with empty content. In forum thread 111867, a user calling gemini-3-pro-image-preview reports occasional responses where the candidate’s content is {} and finishReason is OTHER. The usage metadata showed prompt and thinking tokens but no output. Other posters report the same, including one using the Python SDK with an image alongside the prompt, who says the request works in AI Studio. No cause was posted. The API reference defines OTHER only as an unknown reason.

How to confirm it’s this

candidates[0].finishReason is NO_IMAGE, and the candidate has no image part. Check whether it has text, and log it.
For the merged variant: finishReason is OTHER and content is empty.
promptFeedback.blockReason is not set. If it is, the prompt was blocked.
Python SDK: python-genai issue #2024 reports that when the API returned a finish reason the SDK did not recognize, such as IMAGE_SAFETY or NO_IMAGE, the SDK emitted a UserWarning (“is not a valid FinishReason”) and then hung on finish_reason access. The reporter wrapped property access in a thread with a timeout. The issue was closed as stale. If your worker freezes after that warning, this is a likely cause.

Fix

Ask for an image explicitly. Google’s limitations page says ambiguous prompts can yield text only. Make sure your request config asks for image output (the reporter in issue #2024 set response_modalities=["IMAGE"]), and phrase the prompt as a request for an image.
Check the parts, not only the finish reason. Read the text part when there is no image. It may explain what happened, or be a refusal.
Guard your SDK calls. Put a timeout around response handling so an unrecognized finish reason cannot stall an agent worker.
Log the prompt, code path, modelVersion and usage metadata for failures. In thread 111867, the people trying to reproduce asked for the prompt and the code snippet.

Retry or not

It depends on what came back.

NO_IMAGE or an empty OTHER with no text: a limited retry is reasonable. Google’s limitations page says that when the model stops before it is finished, you can try again, and the thread 111867 reports describe the empty OTHER result as occasional. Retry once or twice with backoff, then fail clearly.
Text only, STOP: retrying the same ambiguous prompt is unlikely to change much. Google’s advice is to change the request to ask for images clearly.
Text that declines on safety grounds: treat it like a block and do not loop.

When it’s not this

IMAGE_SAFETY or IMAGE_PROHIBITED_CONTENT: explicit blocks. Do not retry automatically.
IMAGE_OTHER: a block for unspecified reasons, which users describe as intermittent.
An error about response modalities: if the request itself fails instead of returning a candidate, the problem is the request setup, not generation. See the response modalities page below.

Error text last verified against the sources below on 2026-09-23.