Video tuning

This page provides prerequisites and detailed instructions for fine-tuning Gemini on video data using supervised learning.

Use cases

Fine-tuning lets you adapt base Gemini models for specialized tasks. Here are some video use cases:

  • Automated video summarization: Tuning LLMs to generate concise and coherent summaries of long videos, capturing the main themes, events, and narratives. This is useful for content discovery, archiving, and quick reviews.

  • Detailed event recognition and localization: Fine-tuning allows LLMs to identify and pinpoint specific actions, events, or objects within a video timeline with greater accuracy. For example, identifying all instances of a particular product in a marketing video or a specific action in sports footage.

  • Content moderation: Specialized tuning can improve an LLM's ability to detect sensitive, inappropriate, or policy-violating content within videos, going beyond simple object detection to understand context and nuance.

  • Video captioning and subtitling: While already a common application, tuning can improve the accuracy, fluency, and context-awareness of automatically generated captions and subtitles, including descriptions of nonverbal cues.

Limitations

  • Maximum video file size: 100MB. This may not be sufficient for large video files. Some recommended workarounds are as follows:
    • If there are very few large files, drop those files from including those in the JSONL files.
    • If there are many large files in your dataset and cannot be ignored, reduce visual resolution of the files. This may hurt performance.
    • Chunk the videos to limit the files size to 100MB and use the chunked videos for tuning. Make sure to change any timestamp annotations corresponding to the original video to the new (chunked) video timeline.
  • Maximum video length per example: 5 minutes with MEDIA_RESOLUTION_MEDIUM and 20 minutes with MEDIA_RESOLUTION_LOW.
  • Dropped examples: If an example contains video that is longer than the supported maximum length, that example is dropped from the dataset. Dropped examples are not billed or used for training. If more than 10% of the dataset is dropped, the job will fail with an error message before the start of training.
  • Mixing different media resolutions isn't supported: The value of mediaResolution for each example in the entire training dataset must be consistent. All lines in the JSONL files used for training and validation should have the same value of mediaResolution.

Dataset format

The fileUri field specifies the location of your dataset. It can be the URI for a file in a Cloud Storage bucket, or it can be a publicly available HTTP or HTTPS URL.

The mediaResolution field is used to specify the token count per frame for the input videos, as one of the following values:

  • MEDIA_RESOLUTION_LOW: 64 tokens per frame
  • MEDIA_RESOLUTION_MEDIUM: 256 tokens per frame

Model tuning with MEDIA_RESOLUTION_LOW is roughly 4 times faster than the ones tuned with MEDIA_RESOLUTION_MEDIUM with minimal performance improvement.

When a video segment is used for training and validation, the video segment is in the videoMetadata field. During tuning, this data point is decoded to contain information from the segment extracted from the specified video file, starting from timestamp startOffset (the start offset, in seconds) until endOffset.

To see the generic format example, see Dataset example for Gemini.

The following sections present video dataset format examples.

JSON schema example for cases where the full video is used for training and validation

This schema is added as a single line in the JSONL file.

{
  "contents": [
    {
      "role": "user",
      "parts": [
        {
          "fileData": {
            "fileUri": "gs://<path to the mp4 video file>",
            "mimeType": "video/mp4"
          },
        },
        {
          "text": "
          You are a video analysis expert. Detect which animal appears in the
          video.The video can only have one of the following animals: dog, cat,
          rabbit.\n Output Format:\n Generate output in the following JSON
          format:\n
          [{\n
            \"animal_name\": \"<CATEGORY>\",\n
          }]\n"
        }
      ]
    },
    {
      "role": "model",
      "parts": [
        {
          "text": "```json\n[{\"animal_name\": \"dog\"}]\n```"
        }
       ]
    },
  ],
  "generationConfig": {
    "mediaResolution": "MEDIA_RESOLUTION_LOW"
  }
}

JSON schema example for cases where a video segment is used for training and validation

This schema is added as a single line in the JSONL file.

{
  "contents": [
    {
      "role": "user",
      "parts": [
        {
          "fileData": {
            "fileUri": "gs://<path to the mp4 video file>",
            "mimeType": "video/mp4"
          },
         "videoMetadata": {
            "startOffset": "5s",
            "endOffset": "25s"
          }
        },
        {
          "text": "
          You are a video analysis expert. Detect which animal appears in the
          video.The video can only have one of the following animals: dog, cat,
          rabbit.\n Output Format:\n Generate output in the following JSON
          format:\n
          [{\n
            \"animal_name\": \"<CATEGORY>\",\n
          }]\n"
       }
      ]
   },
   {
      "role": "model",
      "parts": [
        {
          "text": "```json\n[{\"animal_name\": \"dog\"}]\n```"
        }
      ]
    },
  ],
"generationConfig": {
     "mediaResolution": "MEDIA_RESOLUTION_LOW"
  }
}

What's next