FloodDiffusionTiny Mobile ONNX
An unofficial mobile-oriented ONNX conversion of
AlayaLab/FloodDiffusionTiny,
prepared for local Android text-to-motion experiments, Layla mini-apps, VRM
animation, and offline motion-generation pipelines.
Attribution
The original FloodDiffusionTiny architecture, trained motion weights, research,
and source implementation belong to AlayaLab and the original FloodDiffusion
authors.
Original model:
The text encoder and tokenizer are derived from Google's UMT5-Base:
This repository is an unofficial community conversion. It is not affiliated
with or endorsed by AlayaLab or Google.
Conversion work
Converted and packaged by Sparkipy using Google Colab.
My contribution includes:
- converting the UMT5 text encoder to ONNX
- applying dynamic INT8 quantization to the UMT5 text encoder
- converting the FloodDiffusionTiny denoiser to fixed-slot ONNX
- converting the motion VAE decoder to ONNX
- packaging the tokenizer, scheduler, and model configuration
- creating conversion notebooks and validation reports
- preparing checksums and mobile integration documentation
- testing the pipeline with Android, Layla, WebGPU, and WASM
I do not claim authorship of the original model architecture, training data,
research, or trained source weights.
Included components
| Component | Format | Precision |
|---|
| UMT5 text encoder | ONNX | INT8 |
| FloodDiffusionTiny denoiser | ONNX | FP32 |
| Motion VAE decoder | ONNX | FP32 |
| UMT5 tokenizer | JSON and SentencePiece | Original tokenizer data |
| Scheduler and configuration | JSON | N/A |
Intended uses
This package is intended for:
- local Android text-to-motion research
- offline motion-generation experiments
- Layla mini-app development
- VRM and humanoid animation pipelines
- HumanML3D motion reconstruction
- downloadable neural-motion engines for applications
Applications may download this package after installation, verify its checksum,
extract the model files, and cache them for offline use.
Validation status
The individual ONNX components passed conversion and numerical validation
checks.
The Android pipeline has successfully loaded the models through ONNX Runtime
Web, executed using WebGPU and WASM, decoded HumanML3D motion features, and
recovered 22-joint motion.
Prompt-conditioned motion differentiation is still under active validation.
This release should therefore be considered experimental.
Limitations
- Only the UMT5 text encoder is currently INT8 quantized.
- The denoiser and VAE remain FP32.
- Performance depends on the available Android runtime backend.
- VRM retargeting, bone rotation solving, and foot locking are application-level steps.
- This is not a drop-in replacement for the original PyTorch repository.
- Hosted inference is not provided.
License
The original FloodDiffusionTiny repository and Google UMT5-Base identify their
licenses as Apache-2.0.
This conversion package is distributed under Apache License 2.0. Original
ownership and attribution remain with their respective creators.
