What causes "Gemini response contained no images.{detail}" in invoke-ai/InvokeAI?

Read the 'Response status'/'Response text' suffix in the message — it names the finish reason or the model's textual excuse.

invoke-ai/InvokeAI · error · ExternalProviderRequestError

Gemini response contained no images.{detail}

Error message

Gemini response contained no images.{detail}

What it means

ExternalProviderRequestError raised after the provider parsed all candidates and found zero decodable inline images. The message is enriched with a detail suffix: finish messages/finish reasons from candidates (e.g. 'Finish reason: SAFETY') or, if none, the model's text parts (truncated to 500 chars), so the developer can see why Gemini produced text-only or empty output despite a 200 response.

Source

Thrown at invokeai/app/services/external_generation/providers/gemini.py:172

                        raise ExternalProviderRequestError(
                            f"Gemini returned fileUri instead of inline image data: {file_uri}"
                        )
                text = part.get("text")
                if isinstance(text, str):
                    text_parts.append(text)

        if not images:
            self._logger.error("Gemini response contained no images: %s", data)
            detail = ""
            if finish_messages:
                combined = " ".join(message.strip() for message in finish_messages if message.strip())
                if combined:
                    detail = f" Response status: {combined[:500]}"
            elif text_parts:
                combined = " ".join(text_parts).strip()
                if combined:
                    detail = f" Response text: {combined[:500]}"
            raise ExternalProviderRequestError(f"Gemini response contained no images.{detail}")

        return ExternalGenerationResult(
            images=images,
            seed_used=request.seed,
            provider_metadata={"model": request.model.provider_model_id},
        )


def _iter_response_parts(candidate: dict[str, object]) -> list[dict[str, object]]:
    content = candidate.get("content")
    if isinstance(content, dict):
        content_parts = content.get("parts")
        if isinstance(content_parts, list):
            return [part for part in content_parts if isinstance(part, dict)]
    contents = candidate.get("contents")
    if isinstance(contents, list):
        parts: list[dict[str, object]] = []
        for item in contents:

View on GitHub (pinned to 0b6a024f2f)

Solutions

Read the 'Response status'/'Response text' suffix in the message — it names the finish reason or the model's textual excuse.
If SAFETY: rephrase the prompt or remove flagged reference/init images.
If text-only: strengthen the instruction to request an image (the provider's system instruction already tries this) and avoid prompts that read as questions.
Retry — transient model degradation often resolves on a second attempt.
Try a different Gemini image model id for the workload.

Example fix

// before: retrying blindly on empty output
for _ in range(3):
    try:
        return provider.generate(request)
    except ExternalProviderRequestError:
        pass
// after: inspect finish reason and rephrase once if safety/text-only
try:
    return provider.generate(request)
except ExternalProviderRequestError as e:
    if "SAFETY" in str(e):
        request = replace_prompt(request, sanitized_prompt)
        return provider.generate(request)
    raise

Defensive patterns

Strategy: try-catch

Try / catch

try:
    result = provider.generate(request)
except ExternalProviderRequestError as e:
    detail = str(e)
    if "SAFETY" in detail or "RECITATION" in detail:
        notify_user("Gemini blocked this request; rephrase the prompt.")
    elif "Response text:" in detail:
        retry_once_with_stronger_image_instruction(request)  # model answered in prose
    else:
        retry_once(request)  # transient empty response

Prevention

Parse the detail suffix — it tells you whether to rephrase (safety/text) or just retry (transient).
Write prompts that clearly request an image, not questions or explanations.
Prefer init/reference images that are policy-compliant to avoid finishReason SAFETY.
Cap retries at 1-2 to avoid burning quota on prompts the model will never fulfill.

When it happens

Trigger: All candidates contained only text parts (model answered with prose instead of an image), parts were skipped as non-dict, inline_data.data was empty, or candidates carried finishReason like SAFETY/RECITATION with no image payload.

Common situations: Prompt interpreted as a question rather than an image request; safety-filtered content with finishReason SAFETY; model overload returning degraded text-only answers; prompt asking for text output explicitly; invalid imageConfig causing silent fallback.

Related errors

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/e1099a942e00c61d. Report an issue: GitHub.