We’re using the Gemini API (both 2.5 flash and pro) for video understanding and have noticed some issues when sending videos (via Files API or base64 encoded) with ultrawide aspect ratios (e.g. 16:3). The model appears to centre and crop the frames to a 16:9 aspect ratio. This results in the contents on either side of the frame being omitted.
I can’t find any guidance on why this happens or best practices when it comes to aspect ratios and sending video content. Is this the expected behaviour and if so would there be any documentation for this?
If I was to send the single 16:3 frame as an image there is no issues with understanding the full content so it appears to be video related and how the frames are extracted and sent on the API side?
You can replicate the issue in AI Studio also - uploading a 16:3 video and asking it whats on the left or right (assuming there is something of note) it will not be able to answer accurately.
Could you please share your code along with some sample images that you are using? This will help us reproduce the issue on our end and analyze it more effectively.
Thank you for sharing the details. We will thoroughly review and analyze this issue internally and get back to you as soon as possible with our findings.