Any generation request with `video_metadata` results in a 500 error

[kfajdsl]

Environment details

• Programming language: Python (also tested with REST API)
• OS: macOS
• Language runtime version: 3.12.9
• Package version: 1.16.1

Steps to reproduce

1. Run the following script:

from google import genai
from google.genai import types
from pathlib import Path
import os

client = genai.Client(api_key=os.environ['GEMINI_API_KEY'])

video_path = Path("vid.mp4")
video_bytes = video_path.read_bytes()

response = client.models.generate_content(
    model='models/gemini-2.5-flash-preview-05-20',
    contents=types.Content(
        parts=[
            types.Part(
                inline_data=types.Blob(
                    data=video_bytes,
                    mime_type='video/mp4',
                ),
                video_metadata=types.VideoMetadata(
                    fps=5
                )
            ),
            types.Part(text='Please summarize the video in 3 sentences.')
        ]
    )
)

This also happens with start_offset and end_offset.

Also, note, the docs page seems wrong. It has the following:

            types.Part(
                inline_data=types.Blob(
                    data=video_bytes,
                    mime_type='video/mp4',
                    video_metadata=types.VideoMetadata(fps=5)
            )),

This fails with a Pydantic error when run, and according to the API reference and other examples, video_metadata is a member of Part, not Blob.

[Madhuvod]

i think this is a sdk issue no. i have been getting the same errror - 400 INVALID_ARGUMENT. {'error': {'code': 400, 'message': '* GenerateContentRequest.contents[1].parts: contents.parts must not be empty.\n', 'status': 'INVALID_ARGUMENT'}}
MOS Log: Error type: ClientError ; for the code: gemini_api_response = gemini_client.models.generate_content( model=GEMINI_MODEL_NAME, contents=[ AUDIO_ONLY_SYSTEM_PROMPT, types.Part.from_bytes(data=audio_data, mime_type=mime_type) ], config={ "response_mime_type": "application/json", "response_schema": GeminiMOSResponse, } ) ; but the same thing works for google-genai 1.15.0

[kfajdsl]

@Madhuvod I believe my particular issue is an api issue since I have the same issue when calling the REST API directly rather than using the Python SDK.

[Jumaron]

Same Issue, looks like an API Issue.

[kiransair]

Hi @kfajdsl, Thanks for reporting this issue. While trying to reproduce the error it seems the issue is with gemini-2.5-flash-preview-05-20 but works fine with gemini-2.0-flash. Will bring this issue to the engineering team's attention. Please refer to this gist. Thank You

[hkt74]

Hi @kfajdsl

Your code snippet works for me with model models/gemini-2.5-flash-preview-05-20, (and I agree the docs page is incorrect)

I tried with this sample video

Perhaps the issue might be with the video processing itself? Have you had a chance to try it with a different video?

[Jumaron]

Hi @kfajdsl

Your code snippet works for me with model models/gemini-2.5-flash-preview-05-20, (and I agree the docs page is incorrect)

I tried with this sample video

Perhaps the issue might be with the video processing itself? Have you had a chance to try it with a different video?

Super confusing, the video works for me with gemini-2.5-flash-preview-05-20 and flash 2.0.
But my own video gives the 500 internal error, flash 2.0 works.

[Jumaron]

I might have found the issue,
it seems like the new API only accepts videos with an valid audiotrack, atleast adding an audiotrack to my video solved it for me.
Having no audio track will result in that error.

[Etragas]

Wow thank you @Jumaron I added empty audio to my video and it worked!

[janasangeetha]

Hi @kfajdsl
Looks like the issue is resolved! Please let us know if you are still facing issue.
Thank you!

[khayyamkhan]

I have a bit of a dilemna:

Current setup (Vertex AI SDK):

• Using vertexai.generative_models.GenerativeModel
• Limited to 1fps video analysis (no VideoMetadata fps parameter support)
• Works fine for basic analysis but lacks timing precision

Desired setup (Google GenAI SDK):

• Using google.generativeai library with API keys
• Has VideoMetadata class with fps parameter for 24fps analysis
• Would give me much better timing precision for shot detection

The problem is IP protection. I work with unreleased content from major studios (think pre-release studio work etc.) and need to ensure maximum data protection.

My understanding was that:

• Vertex AI = enterprise-grade, content not used for training, better IP protection
• GenAI SDK with API keys = less secure, potentially used for training?

But I'm honestly not sure if this is accurate or just assumptions I've picked up. The documentation isn't super clear on the security differences between these two approaches.

Specific questions:

1. Are there actual security/IP protection differences between the Vertex AI SDK and GenAI SDK?
2. Does vertexai.generative_models support VideoMetadata fps feature in a way that I'm missing?

The VideoMetadata fps feature would be incredibly valuable for my work, but I can't risk client IP. If both libraries offer the same enterprise protections, I'd love to switch. If not, I'll stick with Vertex AI.

Any clarity you can provide would be hugely appreciated!