Sparkipy/FloodDiffusionTiny-Mobile-ONNX

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:

https://huggingface.co/AlayaLab/FloodDiffusionTiny

The text encoder and tokenizer are derived from Google's UMT5-Base:

https://huggingface.co/google/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.