MiniMax-H3-Self-Generated-Dataset

# MiniMax-H3 Self-Generated Dataset A synthetic **text-to-video-with-audio (T2VA)** dataset of 6,430 clips generated with [MiniMax-H3](https://modelscope.cn/models/MiniMax/MiniMax-H3) through [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio). Prompts are sourced from [GokuScraper/seedance-2-prompts-datasets](https://huggingface.co/datasets/GokuScraper/seedance-2-prompts-datasets) and rewritten to fit the 5-second clip duration. ## Dataset at a glance | | | |---|---| | Samples | 6,430 | | Frames per video | 124 | | Frame rate | 24 | | Inference steps | 30 | | Resolutions | 768P | | Generator | MiniMax-H3 (FL2VA weights, T2VA mode) | ## Structure ``` . ├── metadata.jsonl └── videos/ ├── 000001.mp4 ├── 000002.mp4 └── ... 006430.mp4 ``` Each line of `metadata.jsonl` describes one clip: ```json { "prompt": "integrated_multimodal_description: [Shot 1] Live-action, cinematic ...\n\noverall_soundscape: ...\n\nnon_diegetic_music: N/A", "video_path": "videos/000001.mp4", "height": 768, "width": 1344, "num_frames": 124, "num_inference_steps": 30 } ``` ### Fields | Field | Type | Description | |---|---|---| | `prompt` | str | Structured H3 prompt | | `video_path` | str | Path to the mp4, relative to the dataset root | | `height` / `width` | int | Generated resolution in pixels | | `num_frames` | int | Always 124 | | `num_inference_steps` | int | Always 30 | ## Resolutions Each clip keeps the aspect ratio of its source reference and is scaled to fill a budget of `768 × 1344 = 1,032,192` pixels, with both dimensions rounded to a multiple of 32. | Resolution (W×H) | Clips | |---|---| | 1344 × 768 | 4,651 | | 768 × 1344 | 954 | | 864 × 1184 | 195 | | 992 × 1024 | 153 | | 1184 × 864 | 128 | | 1536 × 672 | 125 | | other (21 more) | 224 |