Comfy-Org/ComfyUI · error · ValueError

Gemini API blocked the request. Reasons: {blocked_reasons}

Error message

Gemini API blocked the request. Reasons: {blocked_reasons}

What it means

Raised when the Gemini response DID contain candidates, but every one finished with finishReason == 'IMAGE_PROHIBITED_CONTENT' and no usable parts were extracted. This is the per-candidate safety outcome for image generation: the prompt was accepted, the model ran, but the requested image content was classified as prohibited and suppressed.

Source

Thrown at comfy_api_nodes/nodes_gemini.py:194

        )
    parts = []
    blocked_reasons = []
    for candidate in response.candidates:
        if candidate.finishReason and candidate.finishReason.upper() == "IMAGE_PROHIBITED_CONTENT":
            blocked_reasons.append(candidate.finishReason)
            continue
        if candidate.content is None or candidate.content.parts is None:
            continue
        for part in candidate.content.parts:
            if part_type == "text" and part.text:
                parts.append(part)
            elif part.inlineData and _mime_matches(part.inlineData.mimeType, part_type):
                parts.append(part)
            elif part.fileData and _mime_matches(part.fileData.mimeType, part_type):
                parts.append(part)

    if not parts and blocked_reasons:
        raise ValueError(f"Gemini API blocked the request. Reasons: {blocked_reasons}")

    return parts


def get_text_from_response(response: GeminiGenerateContentResponse) -> str:
    """
    Extract and concatenate all text parts from the response.

    Args:
        response: The API response from Gemini.

    Returns:
        Combined text from all text parts in the response.
    """
    parts = get_parts_by_type(response, "text")
    return "\n".join([part.text for part in parts])

View on GitHub (pinned to 1c6d8d45b3)

Solutions

Soften or abstract the prohibited elements of the prompt (avoid real people, explicit violence, brand characters).
Reframe as illustration styles or fictional subjects rather than realistic depictions.
Retry with minor wording changes — the output filter has non-deterministic thresholds.
Combine with IMAGE+TEXT modality to see any model-side explanation.

Defensive patterns

Strategy: try-catch

Try / catch

try:
    parts = extract_parts(response, "image")
except ValueError as e:
    if "IMAGE_PROHIBITED_CONTENT" in str(e):
        # output-level filter: soften the prompt; do not retry identically
        raise ProhibitedImage(str(e)) from e
    raise

Prevention

Avoid real-person likenesses, explicit violence, and trademarked characters in image prompts.
Reframe prohibited subjects as stylized/fictional equivalents.
Expect non-deterministic thresholds — minor rewording can pass.

When it happens

Trigger: One or more candidates with finishReason IMAGE_PROHIBITED_CONTENT are collected into blocked_reasons, the parts list ends up empty, and the extractor raises with the list of finish reasons.

Common situations: Image prompts about real public figures, violent or sexual imagery, trademarked characters, or borderline content that passes prompt-level filters but fails output-level ones; imagedepict requests the safety layer flags post-generation.

Related errors

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/000edb20641c5e59. Report an issue: GitHub.