vast.ai ComfyUI provisioning manifest: Wan 2.2 I2V 14B models + video custom nodes

wan22-i2v.yaml

# ComfyUI provisioning: Wan 2.2 Image-to-Video 14B
#
# Consumed by the vast.ai provisioner via the PROVISIONING_MANIFEST env var.
# Every phase is content-hash checked and downloads skip existing files, so a
# re-run or a retry after an interrupted pull costs nothing.
#
# Models are the stock Comfy-Org repackaged Wan 2.2 I2V set (35.42 GB).
# Custom nodes are the video/general toolkit needed by that workflow.
#
# Requires ~120 GB disk: 15 GB base image, ~3 GB nodes, 35.4 GB models,
# ~14 GB transient download space, rest for output.

version: 1

settings:
  venv: "/venv/main"
  concurrency:
    # Downloads land in a tmpdir then move. Same filesystem, so the move is a
    # rename, but N parallel pulls hold N files at once. 2 keeps the transient
    # peak near 14 GB instead of 27 GB while staying reasonably fast.
    hf_downloads: 2

apt_packages:
  - libmagic1

pip_packages:
  - tool: uv
    packages:
      - av
      - librosa
      - soundfile
      - resampy
      - pydub
      - ultralytics
      - supervision

git_repos:
  # --- Video / Wan-specific ---
  - url: https://github.com/kijai/ComfyUI-WanVideoWrapper
    dest: "${WORKSPACE:-/workspace}/ComfyUI/custom_nodes/ComfyUI-WanVideoWrapper"
    post_commands:
      - "uv pip install --no-cache --python /venv/main/bin/python -r requirements.txt"

  - url: https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite
    dest: "${WORKSPACE:-/workspace}/ComfyUI/custom_nodes/ComfyUI-VideoHelperSuite"
    post_commands:
      - "uv pip install --no-cache --python /venv/main/bin/python -r requirements.txt"

  - url: https://github.com/kijai/ComfyUI-KJNodes
    dest: "${WORKSPACE:-/workspace}/ComfyUI/custom_nodes/ComfyUI-KJNodes"
    post_commands:
      - "uv pip install --no-cache --python /venv/main/bin/python -r requirements.txt"

  - url: https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler
    dest: "${WORKSPACE:-/workspace}/ComfyUI/custom_nodes/ComfyUI-SeedVR2_VideoUpscaler"
    post_commands:
      - "uv pip install --no-cache --python /venv/main/bin/python -r requirements.txt"

  # --- General toolkit ---
  - url: https://github.com/Fannovel16/comfyui_controlnet_aux
    dest: "${WORKSPACE:-/workspace}/ComfyUI/custom_nodes/comfyui_controlnet_aux"
    post_commands:
      - "uv pip install --no-cache --python /venv/main/bin/python -r requirements.txt"

  - url: https://github.com/ltdrdata/ComfyUI-Impact-Pack
    dest: "${WORKSPACE:-/workspace}/ComfyUI/custom_nodes/ComfyUI-Impact-Pack"
    post_commands:
      - "uv pip install --no-cache --python /venv/main/bin/python -r requirements.txt"

  - url: https://github.com/cubiq/ComfyUI_IPAdapter_plus
    dest: "${WORKSPACE:-/workspace}/ComfyUI/custom_nodes/ComfyUI_IPAdapter_plus"

  - url: https://github.com/filliptm/ComfyUI_Fill-Nodes
    dest: "${WORKSPACE:-/workspace}/ComfyUI/custom_nodes/ComfyUI_Fill-Nodes"
    post_commands:
      - "uv pip install --no-cache --python /venv/main/bin/python -r requirements.txt"

  - url: https://github.com/WASasquatch/was-node-suite-comfyui
    dest: "${WORKSPACE:-/workspace}/ComfyUI/custom_nodes/was-node-suite-comfyui"
    post_commands:
      - "uv pip install --no-cache --python /venv/main/bin/python -r requirements.txt"

downloads:
  # --- VAE (242 MB) ---
  - url: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/vae/wan_2.1_vae.safetensors
    dest: "${WORKSPACE:-/workspace}/ComfyUI/models/vae/wan_2.1_vae.safetensors"

  # --- LightX2V 4-step LoRAs (1.14 GB each) ---
  - url: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/loras/wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors
    dest: "${WORKSPACE:-/workspace}/ComfyUI/models/loras/wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors"

  - url: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/loras/wan2.2_i2v_lightx2v_4steps_lora_v1_low_noise.safetensors
    dest: "${WORKSPACE:-/workspace}/ComfyUI/models/loras/wan2.2_i2v_lightx2v_4steps_lora_v1_low_noise.safetensors"

  # --- Text encoder (6.27 GB) ---
  - url: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors
    dest: "${WORKSPACE:-/workspace}/ComfyUI/models/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors"

  # --- Diffusion models (13.31 GB each) ---
  - url: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/diffusion_models/wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors
    dest: "${WORKSPACE:-/workspace}/ComfyUI/models/diffusion_models/wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors"

  - url: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/diffusion_models/wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors
    dest: "${WORKSPACE:-/workspace}/ComfyUI/models/diffusion_models/wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors"