Goddard/unimate-gguf

UniMate — GGUF Conversion

GGUF conversion of UniMate for running 3D motion generation locally.

This repository contains the UniMate motion generation model and its FLAN-T5 Base text encoder, converted to GGUF for use with a custom GGUF inference implementation.

Model Files

File	Description	Approximate Size
model-f60-v3-s150000-f32.gguf	UniMate motion model (unimate_uniml3d_f60_v3, step 150000), F32	283 MiB
flan-t5-base-encoder-f32.gguf	FLAN-T5 Base text encoder, F32, tokenizer embedded	420 MiB

Both files are required for inference. They are full-precision F32 conversions, not quantized.

Original Project

UniMate: One Unified Model to Animate Diverse Skeletons (SIGGRAPH Asia 2026), by Linzhan Mou, Jiahui Lei, Zhiyang Dou, Chenyue Cai, Chaoyue Song, Adam Finkelstein and Szymon Rusinkiewicz.

• Source code: https://github.com/Friedrich-M/UniMate
• Original checkpoints: https://huggingface.co/Linzhan/UniMate
• Paper: arXiv 2609.05415

The original architecture, training and research belong to the UniMate authors. This repository is an independent GGUF conversion, not an official release from the original authors.

Conversion Details

• Motion model: unimate_uniml3d_f60_v3, checkpoint step 150000
• Text encoder: google/flan-t5-base, encoder only
• Precision: F32
• Format: GGUF
• Tokenizer: embedded in the text encoder GGUF

The weights were not trained or fine-tuned further during the conversion.

Usage

These files need a compatible UniMate GGUF inference implementation. They will not run in llama.cpp or other standard GGUF tools, because the motion model is a custom flow-matching denoiser that standard tools don't implement.

They are being integrated into Project IE 4K (PIE4K) for local 3D animation workflows.

The model generates 60 frames (2 seconds) at 30 fps. Prompts work best when they describe only the motion and start with "An object", for example "An object throws a punch."

License and Attribution

• UniMate motion model: Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0), following the original checkpoint license. https://creativecommons.org/licenses/by-nc/4.0/
• FLAN-T5 Base encoder: derived from Google's google/flan-t5-base and distributed under its Apache 2.0 license.
• UniMate code: MIT, per the original repository.

The UniMate license covers only the authors' rights in the weights. Training data keeps its source terms. The v3 model was trained on data that includes Truebones ZOO, Mixamo and Objaverse-XL, each under its own terms. See the original checkpoint page for details.

This repository is intended for noncommercial use.

Citation

@article{mou2026unimate,
  title   = {UniMate: One Unified Model to Animate Diverse Skeletons},
  author  = {Mou, Linzhan and Lei, Jiahui and Dou, Zhiyang and Cai, Chenyue and Song, Chaoyue and Finkelstein, Adam and Rusinkiewicz, Szymon},
  journal = {arXiv preprint arXiv:2609.05415},
  year    = {2026}
}

Disclaimer

This is an unofficial conversion. It is not affiliated with or endorsed by the original UniMate authors or Google.