comfyui-mcp/packs/wan-transparent-img2vid-96gb/pack.yaml

Local-first, agent-native control plane for ComfyUI — MCP server + autonomous sidebar agent that drives your live graph in natural language on ANY LLM: Claude/ChatGPT/Gemini on your subscription (no API key), free local models via Ollama (fully offline),

name: wan-transparent-img2vid-96gb
display_name: WAN 2.2 Transparent Expressions — Image to Video (96GB fp16)
family: wan
kind: img2vid
description: >-
  WAN 2.2 image-to-video workflow that renders transparent-background animated
  character expressions (looping, pingpong WebP + GIF with an alpha channel) for
  use as SillyTavern expression sprites. Drives the FULL fp16 WAN 2.2 I2V A14B
  UNets via UNETLoader (26.6GB each, not GGUF — this is the 96GB tier; use
  wan-transparent-img2vid for the GGUF quants) with the 4-step Lightning I2V
  LoRAs, then cuts the subject
  out per-frame with the ComfyUI-RMBG BiRefNetRMBG node (Alpha output, the
  BiRefNet_toonout matting model) so the exported animation has a transparent
  background. Single-pipeline graph sliced from the wan-transparent monolith
  (one "WAN 2.2 IMAGE TO VIDEO" group, un-bypassed into a standalone graph).
vram: "96GB (full fp16 I2V A14B, 26.6GB per UNet, --highvram); OOMs on 24GB — use wan-transparent-img2vid for GGUF"
workflow: workflow.json
skill: wan-t2v-video
launch_args: []
sources:
  workflow: WAN2_2-IMG2VID_TRANSPARENT_WORKFLOW.json
  base_installer: WAN2_2-ULTRA-MODELS-NODES_INSTALL.bat
  mirror: https://huggingface.co/Aitrepreneur/FLX
  transparency_model: https://huggingface.co/1038lab/BiRefNet
  rmbg_node: https://github.com/1038lab/ComfyUI-RMBG
notes:
  - Run the generated installer from your ComfyUI root (the folder containing
    custom_nodes/ and models/).
  - Sliced + un-bypassed from the monolith's "WAN 2.2 IMAGE TO VIDEO" pipeline.
  - Transparency is node-based, not a LayerDiffuse-style model — the BiRefNetRMBG
    node (ComfyUI-RMBG) removes the background per frame and outputs RGBA using the
    BiRefNet_toonout model; VHS_VideoCombine then exports a lossless pingpong WebP
    and a GIF with the alpha preserved.
  - Base WAN 2.2 weights (VAE, umt5 text encoder, Lightning I2V LoRAs) come from
    the WAN 2.2 ULTRA installer mirror; the I2V high/low UNets are the full fp16
    Comfy-Org repackaged weights (diffusion_models/wan2.2_i2v_*_14B_fp16.safetensors,
    26.6GB each, loaded with UNETLoader). Launch ComfyUI with --highvram. For
    anything smaller than a 96GB card, install wan-transparent-img2vid instead.
  - >-
    Torch/CUDA fix — if a custom node (e.g. ComfyUI-RMBG) drags in a Torch
    version that breaks CUDA, reinstall the matching CUDA build, e.g.
    pip install --force-reinstall torch torchvision torchaudio
    --index-url https://download.pytorch.org/whl/cu121
post_install:
  - Restart ComfyUI (or use ComfyUI-Manager -> Install missing custom nodes to pull
    any node Python deps).
  - Load workflow.json. On first run, BiRefNetRMBG auto-downloads BiRefNet_toonout
    to models/RMBG/BiRefNet/ if the manifest did not already place it there.