comfyui-mcp/packs/wan-pusa-extend/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-pusa-extend
display_name: WAN 2.2 Video Extend — Pusa (temporal)
family: wan
kind: video
description: >-
  Temporally EXTEND / continue an existing clip with Pusa 2.2 on the
  ComfyUI-WanVideoWrapper stack. The loaded clip is VAE-encoded
  (WanVideoEncode) and injected as the FIRST latents; the flowmatch_pusa
  scheduler + WanVideoAddPusaNoise hold those conditioning frames near-clean
  while generating the tail, so the existing motion, subject, camera and color
  carry across the seam (no I2V freeze-then-restart). WAN 2.2 T2V A14B HIGH/LOW
  fp8 + Pusa V1 hi/lo LoRAs + lightx2v distill (6-step, cfg 1, two-pass). Ships
  with a ColorMatchV2 seam-regrade so the generated tail doesn't saturate, and
  attention_mode=sdpa so it runs without SageAttention/triton.
vram: >-
  24GB+ (dual A14B fp8 + Pusa LoRAs, block swap on, 81 frames @832x480).
  12-16GB: heavier WanVideoBlockSwap + tiled VAE encode/decode + 49 frames.
workflow: workflow.json
skill: video-extend
sources:
  wanvideowrapper: https://github.com/kijai/ComfyUI-WanVideoWrapper
  reference_graph: ComfyUI-WanVideoWrapper/example_workflows/wanvideo_2_2_14B_Pusa_extension_example_01.json
  pusa_loras: https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Pusa
  pusa_upstream: https://huggingface.co/RaphaelLiu/PusaV1
notes:
  - The T2V A14B UNets, UMT5 encoder and WAN VAE are SHARED with wan-t2v-video —
    if you have that pack you already have them (apply_manifest skips existing
    files). Only the two Pusa V1 LoRAs (~1.9 GB) are Pusa-specific.
  - Set VHS_LoadVideo to your own clip (the bundled graph points at a placeholder).
  - "fps seam — if your clip was frame-interpolated (e.g. RIFE'd to 32/50 fps),
    set VHS_LoadVideo.force_rate=16 so the conditioning cadence matches WAN-native;
    otherwise the encoded past runs at 2-3x pace and the seam speed-jumps. Best:
    extend the pre-interpolation 16fps master, then upscale/interpolate the
    combined result."
  - "Runs on attention_mode=sdpa with no torch.compile — no SageAttention/triton
    needed. To accelerate, install them and switch the WanVideoModelLoaders to
    sageattn + re-attach WanVideoTorchCompileSettings (see triton-sageattention)."
  - "merge_loras=false on the WanVideoLoraSelectMulti nodes — REQUIRED on fp8
    models; merge_loras=true hard-crashes ComfyUI during LoRA load."
  - "Seam color/saturation drift is fixed by a ColorMatchV2 (hm-mkl-hm) that
    re-grades the generated frames to the source clip's palette, plus base
    precision bf16 (not fp16_fast)."
  - "Action latitude vs. fidelity: dial CreateScheduleFloatList end_value (the
    noise_multipliers ramp top). ~0.2 clings tightly to the source; ~0.35 frees a
    new action (e.g. turn + smile); too high drifts identity/scene."
  - "Chain for longer video: feed the tail of the NEW output back into
    WanVideoEncode and re-run; color-match each hop to the previous to stop
    drift compounding."
post_install:
  - Restart ComfyUI (or ComfyUI-Manager -> Install missing custom nodes).
  - Load workflow.json, point VHS_LoadVideo at your clip, run.