Skip to content

Latest commit

 

History

History
61 lines (44 loc) · 2.53 KB

README.md

File metadata and controls

61 lines (44 loc) · 2.53 KB

Model Checkpoints Download Guide

This project requires several model weights to function. By default, app.py and gradio_app.py support automatic downloading from Hugging Face. However, for offline environments or faster startup, we recommend manual downloading.

1. Motion Generation Models (T2M Models)

These are the main DiT-based motion generation models. You can choose between the Standard or Lite version.

Model Name Hugging Face Repo ID
HY-Motion-1.0 tencent/HY-Motion-1.0/HY-Motion-1.0
HY-Motion-1.0-Lite tencent/HY-Motion-1.0/HY-Motion-1.0-Lite

Manual Download Command:

# Example for Standard version
huggingface-cli download tencent/HY-Motion-1.0 --include "HY-Motion-1.0/*" --local-dir ckpts/tencent

# Example for Lite version
huggingface-cli download tencent/HY-Motion-1.0 --include "HY-Motion-1.0-Lite/*" --local-dir ckpts/tencent

2. Text Encoders

The model uses CLIP and Qwen as text encoders. According to hymotion/network/text_encoders/text_encoder.py, the paths depend on the USE_HF_MODELS environment variable.

  • If USE_HF_MODELS=1 (Default): The code will fetch models directly from Hugging Face using Repo IDs: openai/clip-vit-large-patch14 and Qwen/Qwen3-8B.
  • If USE_HF_MODELS=0: The code expects the weights to be located in the ckpts/ directory.

Manual Download Commands:

# CLIP Large
huggingface-cli download openai/clip-vit-large-patch14 --local-dir ckpts/clip-vit-large-patch14/

# Qwen Text Encoder
huggingface-cli download Qwen/Qwen3-8B --local-dir ckpts/Qwen3-8B

3. Prompt Rewriter & Duration Predictor (Optional)

If you intend to use the text rewriting and duration estimation module (via vllm as configured in start_space.sh), you will need this model:

Manual Download Command:

huggingface-cli download Text2MotionPrompter/Text2MotionPrompter --local-dir ckpts/Text2MotionPrompter

Recommended Directory Structure

If you prefer local loading (recommended for stability), your ckpts/ directory should look like this:

ckpts/
├── tencent/
│   ├── HY-Motion-1.0/         # Contains config.yml and latest.ckpt
│   └── HY-Motion-1.0-Lite/    # Optional
├── clip-vit-large-patch14/     # CLIP weights
├── Qwen3-8B/                   # Qwen text encoder weights
└── Text2MotionPrompter/        # vLLM Rewriter weights (Optional)

Note: Ensure you set USE_HF_MODELS=0 in your environment variables if you want the application to read from the local ckpts/ folder instead of the internet.