MiniMax H3 First Last Frame LoRA Trainer

URL to zip archive with video clips. Try to use at least 10 clips, although more is better. **Supported video formats:** .mp4, .mov, .avi, .mkv Note: The dataset must contain ONLY videos. Image datasets are not supported - images cannot serve as training media (some endpoints accept images as conditioning sidecars; see the endpoint's documentation). Mixed image/video archives are also rejected. The archive can also contain text files with captions. Each text file should have the same base name as the video clip it corresponds to.

The rank of the LoRA adaptation. Higher values increase capacity but use more memory.

Options: 8, 16, 32, 64, 128

The number of training steps.

Learning rate for optimization. Higher values can lead to faster training but may cause overfitting.

Number of frames per training sample. Must satisfy frames % 17 == 5 (e.g., 22, 39, 56, 73, 90, 107, 124).

Target frames per second for the video.

Resolution to use for training. Higher resolutions require more memory.

Options: low, medium

Aspect ratio to use for training. Matches the aspect ratios supported by the MiniMax H3 API.

Options: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16

A phrase that will trigger the LoRA style. Will be prepended to captions during training.

Probability of conditioning on the first frame only.

If true, videos will be automatically scaled to the target frame count and fps.

If true, videos above a certain duration threshold will be split into scenes.

The duration threshold in seconds. If a video is longer than this, it will be split into scenes.

When enabled, the trainer returns a downloadable archive of your preprocessed training data for manual inspection. Use this to verify that your videos, images, and captions were processed correctly before committing to a full training run.

Reject dataset-quality fallbacks during preprocessing, including captions that would use a blank prompt, missing sidecar keyframes, and videos without a readable audio track. By default these cases are reported in logs and use the documented blank-prompt, clip-frame, or silence fallback.

Probability of conditioning on the last frame only.

Probability of conditioning on BOTH the first and last frames (the API's first-to-last keyframe mode). The remaining probability mass trains unconditioned text-to-video-audio.