Skip to content

Latest commit

 

History

History
61 lines (44 loc) · 2.03 KB

README.md

File metadata and controls

61 lines (44 loc) · 2.03 KB
Raw

Tora diffusers-version

This path contains the diffusers-version of Tora. It is independent from the original Tora code which based on SwissArmyTransformer.

Installation

Please make sure your Python version is between 3.10 and 3.12, inclusive of both 3.10 and 3.12.

# Install Pytorch (we use Pytorch 2.4.0) and torchvision following the official instructions: https://pytorch.org/get-started/previous-versions/. For example:
conda create -n tora python==3.10
conda activate tora
conda install pytorch==2.4.0 torchvision==0.19.0 pytorch-cuda=12.1 -c pytorch -c nvidia

# Install requirements
cd diffusers-version
pip install -r requirements.txt

Model Weights

  • HuggingFace

# This can be faster
pip install "huggingface_hub[hf_transfer]"
HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download Le0jc/Tora_T2V_diffusers --local-dir ckpts/Tora_T2V_diffusers

or

# use git
git lfs install
git clone https://huggingface.co/Le0jc/Tora_T2V_diffusers

  • ModelScope

pip install modelscope
modelscope download --model Alibaba_Research_Intelligence_Computing/Tora_T2V_diffusers --local_dir ckpts/Tora_T2V_diffusers

or

git lfs install
git clone https://www.modelscope.cn/Alibaba_Research_Intelligence_Computing/Tora_T2V_diffusers.git

Inference

cd diffusers-version
python inference.py --prompt "A squirrel gathering nuts." --model_path ckpts/Tora_T2V_diffusers --output_path ./output.mp4 --generate_type t2v --point_path ../sat/trajs/pause.txt --enable_model_cpu_offload --enable_slicing --enable_tiling --enable_sageattention --enable_compile
  • If your VRAM is still not enough, you can replace "--enable_model_cpu_offload" to "--enable_sequential_cpu_offload" and try again. This can reduce the VRAM usage to about 5 GiB. Note that sequential_cpu_offload is much slower.
  • If you have enough VRAM, you can disable cpu offload, VAE slicing and tiling, to speed up the inference.
  • Note that --enable_compile will speed up inference at the cost of slowing down the first inference step.