Summary

When Gemini's image-generation endpoint returns finishReason: IMAGE_SAFETY / IMAGE_PROHIBITED_CONTENT, LiteLLM discards the finishReason and returns an empty ImageResponse with no exception. Downstream code can't tell "safety refusal" from "unrelated failure".

The chat-completion path was fixed by #20524 (which closed #20357), but only patched vertex_and_google_ai_studio_gemini.py. The image-gen transformation files were never updated.

Affected (verified 2026-05-27)

• main and v1.86.1
• Files: litellm/llms/gemini/image_generation/transformation.py, litellm/llms/vertex_ai/image_generation/vertex_gemini_transformation.py

Both iterate candidates[].content.parts[] for inlineData, ignoring finishReason. Grep for finish|safety|prohibit|content_filter|_hidden_params in either file returns zero matches.

Reproduction

response = await litellm.aimage_generation(
    model="gemini/gemini-2.5-flash-image",
    prompt="turn this person into Yoda",
)
# response.data == []
# no exception

Underlying Gemini response contains candidates[0].finishReason: "IMAGE_PROHIBITED_CONTENT" — lost.

Suggested fix

Per #21348 Criterion 2 — preserve raw finishReason in _hidden_params or provider_specific_fields, and/or raise ContentPolicyViolationError when finishReason in {IMAGE_SAFETY, IMAGE_PROHIBITED_CONTENT, SAFETY, IMAGE_RECITATION}, matching the chat-completion behavior.

Related

• #20357, #20524 (chat-only fix), #21348