oakmindai/minimax_h3_avatar_500

[image]

Watch the full 500-video showcase on YouTube

MiniMax H3 Avatar 500

An image-to-video dataset pairing reference avatar images with detailed generation prompts and generated avatar videos. This release contains 500 curated examples in both a browsable raw layout and a typed Hugging Face dataset.

Version 1.0 · Released August 14, 2026

Dataset contents

Each example contains:

• —A 1024 × 1024 reference avatar image
• —A detailed English generation prompt
• —A generated 640 × 480 MP4 video
• —A stable OakMind asset ID

The videos are 5.167 seconds long, contain 124 frames, and were generated using 24 inference steps.

Generation details

The generation records identify the MiniMax H3 Ref2VA model family and API version. The exact checkpoint filename was not stored in the per-sample metadata.

Creation and provenance

OakMind created this dataset as a fully synthetic image-to-video collection. The reference avatars were generated internally from structured, synthetic character descriptions using Tongyi-MAI/Z-Image-Turbo with eight inference steps at 1024 × 1024. They were not collected from external image or photo datasets and do not represent intentionally depicted real people.

Detailed English video prompts were created for each accepted avatar, preserving its visible identity and design while specifying motion, camera direction, environment, dialogue, sound, and music. Each reference-and-prompt pair was then processed through the MiniMax H3 reference-image-to-video pipeline described above.

Curation

Reference images were screened with automated quality checks before video generation. Accepted images were required to contain one full-body character with valid anatomy, two arms and two visible hands, visible feet, acceptable render quality, and a minimum quality score of 3. The release includes 500 records with a successfully completed generation and a complete reference image, prompt, and video triplet.

Repository structure

The repository provides a typed Parquet dataset for programmatic use and the original files for direct browsing.

data/
└── train-00000-of-00001.parquet

raw/
└── OA-AVA-XXXXXX/
    ├── reference.jpg
    ├── prompt.txt
    └── video.mp4

The Parquet shard powers the Hugging Face Dataset Viewer. The raw/ directory preserves each sample in a simple, portable folder layout.

Loading the dataset

from datasets import load_dataset

dataset = load_dataset("oakmindai/minimax_h3_avatar_500")
sample = dataset["train"][0]

print(sample["id"])
print(sample["text"])
image = sample["image"]
video = sample["video"]

Video decoding with Hugging Face Datasets requires compatible torchcodec and FFmpeg installations. The raw MP4 files can also be downloaded or streamed directly from the repository.

Intended uses

• —Image-conditioned avatar video generation research
• —Multimodal model training and evaluation
• —Prompt–output alignment analysis
• —Avatar consistency and motion-quality studies

Citation

If this dataset supports your work, please cite:

@dataset{oakmind2026minimaxh3avatar500,
  author    = {{OakMind AI}},
  title     = {MiniMax H3 Avatar 500},
  year      = {2026},
  month     = {8},
  version   = {1.0},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/oakmindai/minimax_h3_avatar_500}
}

Related Resources

Looking for game-development-ready 3D models and 3D-printable assets? Explore OakAtlas.

License

The generated videos are subject to the MiniMax H3 Community License Agreement.

MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved.

Users are responsible for reviewing and complying with the license, its territorial scope, acceptable-use requirements, and downstream-use restrictions before accessing or using this dataset. For dataset questions, contact hello@oakmind.ai.

Acknowledgments

OakMind assembled, curated, documented, and published this dataset. We acknowledge MiniMax for the MiniMax H3 model used to generate the videos and Tongyi-MAI for Z-Image-Turbo, used to generate the reference images.

Contact

Published by OakMind. Follow OakMind AI on X for new releases and updates. For questions and collaborations, contact hello@oakmind.ai.