FloodDiffusion Downloads

This repository contains the datasets, dependencies, and pretrained models for FloodDiffusion: Tailored Diffusion Forcing for Streaming Motion Generation.

Code repository: GitHub

Repository Structure

The files in this repository are organized to match the directory structure required by FloodDiffusion.

1. Model Checkpoints

The outputs.zip and outputs_tiny.zip archives contains the pretrained model weights.

• Target: Unzip into your project root. It should create an outputs/ folder.

outputs/
├── vae_1d_z4_step=300000.ckpt          # VAE model (1D, z_dim=4)
├── 20251106_063218_ldf/
│   └── step_step=50000.ckpt            # LDF model checkpoint (HumanML3D)
├── 20251107_021814_ldf_stream/
│   └── step_step=240000.ckpt           # LDF streaming model checkpoint (BABEL)
├── 20251217_023720_ldf_tiny/
│   └── step_step=60000.ckpt            # LDF tiny model checkpoint
└── 20251219_01492_ldf_tiny_stream/
    └── step_step=200000.ckpt           # LDF tiny streaming model checkpoint

2. Datasets

Due to the large number of files, datasets are provided as ZIP archives.

• HumanML3D.zip: Contains the HumanML3D dataset (extracted features and texts).
  • Target: Unzip into raw_data/. It should create raw_data/HumanML3D/ containing new_joint_vecs, texts, etc.
• BABEL_streamed.zip: Contains the BABEL dataset processed for streaming generation.
  • Target: Unzip into raw_data/. It should create raw_data/BABEL_streamed/.

3. Dependencies (deps.zip)

• deps.zip: Contains necessary dependencies like the T5 text encoder, evaluation models (T2M), and GloVe embeddings.
  • Target: Unzip into your project root. It should create a deps/ folder.

deps/
├── t2m/                     # Text-to-Motion evaluation models
├── glove/                   # GloVe word embeddings
└── t5_umt5-xxl-enc-bf16/    # T5 text encoder

How to Download & Setup

We recommend using the python script below to automatically download and place files in the correct structure.

Python Script (Recommended)

Save this as download_assets.py in your FloodDiffusion project root:

from huggingface_hub import hf_hub_download
import zipfile
import os

REPO_ID = "AlayaLab/FloodDiffusionDownloads"

def download_extract_zip(filename, target_dir="."):
    print(f"Downloading {filename}...")
    path = hf_hub_download(repo_id=REPO_ID, filename=filename, repo_type="model")
    print(f"Extracting {filename} to {target_dir}...")
    with zipfile.ZipFile(path, 'r') as zip_ref:
        zip_ref.extractall(target_dir)

# 1. Download and extract Dependencies (creates ./deps/)
download_extract_zip("deps.zip", ".")

# 2. Download and extract Datasets (creates ./raw_data/HumanML3D and ./raw_data/BABEL_streamed)
os.makedirs("raw_data", exist_ok=True)
download_extract_zip("HumanML3D.zip", "raw_data")
download_extract_zip("BABEL_streamed.zip", "raw_data")

# 3. Download Models (creates ./outputs/)
download_extract_zip("outputs.zip", ".")
download_extract_zip("outputs_tiny.zip", ".")

print("Done! Your project is ready.")

Data License & Acknowledgements

This repository provides pre-processed motion features (263-dim) to facilitate the reproduction of FloodDiffusion.

• HumanML3D: The motion features are derived from the HumanML3D pipeline, originally built upon AMASS and HumanAct12.
• BABEL: The streaming motion features are derived from the BABEL dataset, which also builds upon AMASS.

Important Note: We only distribute the extracted motion features and text annotations, which is standard practice in the research community. We do not distribute the raw AMASS data (SMPL parameters/meshes). If you require the raw motion data or plan to use it for commercial purposes, you must register and agree to the licenses on the AMASS website.