DynamiCrafter — Quick Install with uv

This fork is configured for uv-managed environments. Core Python dependencies are declared in pyproject.toml and locked in uv.lock. PyTorch/XFormers are intentionally excluded and must be installed with the provided helper script to match your system (CUDA or CPU).

For usage guides, models, training, and examples, see the upstream documentation:

• Upstream repo: https://github.com/Doubiiu/DynamiCrafter

Prerequisites

• Windows 10/11
• NVIDIA GPU + recent driver for CUDA builds (optional)
• uv package manager installed (https://docs.astral.sh/uv/)

1) Create and sync the environment

From the repo root:

# Ensure the exact interpreter required by this project is available
uv python install 3.8.5

# Create/sync the project environment from pyproject.toml and uv.lock
uv sync --frozen

Notes:

• The project pins Python to 3.8.5 via pyproject.toml.
• uv sync --frozen uses uv.lock for fully reproducible versions.

2) Install PyTorch (CUDA or CPU)

Use the helper script to detect CUDA and install matching wheels. You can also force a tag:

# Auto-detect CUDA and install torch/torchvision/torchaudio with fallbacks
uv run --no-sync python scripts/install_torch.py

# Optional: force a specific wheel tag (cu118 | cu121 | cpu)
uv run --no-sync python scripts/install_torch.py --tag cu118

# Optional: clean reinstall or install MSVC runtime if needed on Windows
uv run --no-sync python scripts/install_torch.py --reinstall --install-vcredist

Tips:

• Use --no-sync with uv run so uv does not attempt to re-resolve/replace CUDA wheels.
• The script pins versions known to work on Python 3.8 by default and verifies CUDA availability after install.

3) Verify installation

uv run --no-sync python -c "import torch, torch.cuda as c; print('torch', torch.__version__, 'cuda', torch.version.cuda, 'available', c.is_available())"

If available is True and a CUDA version is shown, your GPU build is working. Otherwise, the CPU build is in use.

Next steps

• For inference, training, and demos, follow the instructions in the upstream README: https://github.com/Doubiiu/DynamiCrafter
• Launch commands (examples) will work when run via uv run --no-sync ... using this environment.

Run the Gradio demos by resolution

Image-to-Video (choose 256, 512, or 1024)

Launch the app for a specific resolution using the --res flag:

# 256x256
uv run --no-sync python gradio_app.py --res 256

# 320x512
uv run --no-sync python gradio_app.py --res 512

# 576x1024
uv run --no-sync python gradio_app.py --res 1024

Notes:

• On first run, the app will automatically download the corresponding pretrained weights from Hugging Face into checkpoints/ if they are missing.
• Keep using --no-sync so uv does not attempt to replace CUDA-enabled wheels.

Generative Frame Interpolation and Looping (320x512)

The interpolation and seamless looping demo runs at 320x512:

uv run --no-sync python gradio_app_interp_and_loop.py

Notes:

• This app will auto-download the DynamiCrafter512_interp weights to checkpoints/dynamicrafter_512_interp_v1/ on first run if needed.

Why use --no-sync with uv?

By default, uv run attempts to sync your environment to pyproject.toml/uv.lock. We intentionally do not pin torch in pyproject.toml because the correct wheel depends on your CUDA version. A normal sync can therefore pull a CPU-only torch transitively and overwrite your CUDA build.

--no-sync tells uv to run without modifying the environment, preserving the CUDA-enabled torch you installed via scripts/install_torch.py.

Alternative: use the venv’s Python directly

./.venv/Scripts/python gradio_app.py --res 1024
./.venv/Scripts/python gradio_app_interp_and_loop.py

Torch install/verify quick reference

# Reinstall CUDA-enabled torch wheels with automatic detection/fallbacks
uv run --no-sync python scripts/install_torch.py --reinstall

# Verify CUDA availability
uv run --no-sync python -c "import torch, torch.cuda as c; print('torch', torch.__version__, 'cuda', torch.version.cuda, 'available', c.is_available())"

If CUDA is not available, ensure your NVIDIA driver and MSVC 2015–2022 (x64) Redistributable are installed. You can add --install-vcredist to the installer script on Windows.

Checkpoint downloads (Hugging Face)

On first run, the apps auto-download checkpoints into checkpoints/. Downloads are large and may take time. The downloader is configured to resume and retry on transient errors.

If your network is flaky, enabling the Rust-based transfer tool often helps:

uv run --no-sync python -m pip install -U hf_transfer
setx HF_HUB_ENABLE_HF_TRANSFER 1
# Open a new terminal, then run the app again

Manual fallback (explicit CLI with resume):

uv run --no-sync huggingface-cli download Doubiiu/DynamiCrafter_512 model.ckpt `
  --local-dir checkpoints/dynamicrafter_512_v1 `
  --local-dir-use-symlinks False `
  --resume

uv run --no-sync python gradio_app.py --res 512