Frames & timing

All times in the protocol are expressed as frame counts, not seconds.

This includes:

• segments[].duration_frames
• request.duration_frames
• constraint frames / frame
• Options.transition_frames

Clients that author in seconds convert with:

frames = round(seconds × fps)

Where fps is request.timing.fps if set, or otherwise Capabilities.models[].fps.

Capability fields like max_duration_seconds and native_clip_seconds remain in seconds because they describe model properties independent of any request.

request.timing

{
  "timing": { "fps": 24.0 }
}

When set:

• All frame counts and indices in the request (segments, constraints, duration_frames) are interpreted in this fps.
• The response glTF is resampled to this fps — slerp for rotations, linear interpolation for translations.

When omitted, frames are in the model's native fps and the response is at the model's native fps.

Resampling quality

Resampling is best-effort. Clients needing precise low-fps results SHOULD request the model's native fps and resample themselves.

Native vs requested fps

Capabilities tell you both:

{
  "id": "kimodo-soma-rp",
  "fps": 30.0,
  "native_clip_seconds": 10.0,
  "chunking": "stitched",
  "recommended_max_duration_seconds": 12.0
}

• fps — the model's native frame rate.
• native_clip_seconds — the duration the model was trained on. Requests beyond this are stitched per chunking.
• chunking: "stitched" — server may chunk longer requests internally. Visible seam quality is best-effort beyond recommended_max_duration_seconds.
• recommended_max_duration_seconds — the comfort zone. Clients SHOULD warn the user past this point.

Chunk boundaries (response)

When the server stitches, it lists the seam frames in the response so clients can recolor those frames in their dopesheet, run smoothing, or warn the user:

{
  "extensions": {
    "MMCP_motion": {
      "samples": [
        {
          "name": "sample_0",
          "num_frames": 360,
          "chunk_boundaries": [120, 240]
        }
      ]
    }
  }
}

chunk_boundaries is an empty array when the request fit in one chunk.