# WAN MultiTalk — audio-driven talking-avatar (MeiGen-AI MultiTalk on WAN 2.1 14B I2V).
#
# Pipeline (kijai WanVideoWrapper): LoadImage (portrait) + LoadAudio → AudioSeparation
# /AudioCrop → DownloadAndLoadWav2VecModel + MultiTalkWav2VecEmbeds (speech embeds) →
# WanVideoModelLoader (WAN 2.1 14B I2V GGUF) + MultiTalkModelLoader + WanVideoLoraSelect
# (lightx2v step-distill) → WanVideoImageToVideoMultiTalk → WanVideoSampler →
# WanVideoDecode → VHS_VideoCombine. 480p, ~10s.
#
# All 6 model URLs below are HEAD-verified live (packs:check-urls, 6/6 OK). Wav2Vec
# is auto-downloaded by DownloadAndLoadWav2VecModel (no manifest entry needed).
# The GGUF downloads under city96's name and is saved as the workflow's local name.
# The workflow's AudioSeparation/AudioCrop nodes come from the explicitly declared
# audio-separation-nodes-comfyui pack below; they are not WanVideoWrapper nodes.
# The bundled WanVideoWrapper revision rejects scaled_fp8 UMT5 checkpoints, so use
# the compatible FP16 pairing instead. NOT yet render-verified — do
# a live render on a local ComfyUI before shipping.

pip:
  - imageio-ffmpeg   # VHS_VideoCombine needs an ffmpeg for the mux

custom_nodes:
  - https://github.com/kijai/ComfyUI-WanVideoWrapper.git   # WanVideo*/MultiTalk*/Wav2Vec/T5 nodes
  - https://github.com/kijai/ComfyUI-KJNodes.git           # ImageResizeKJv2
  - https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite.git  # VHS_VideoCombine
  - https://github.com/christian-byrne/audio-separation-nodes-comfyui.git  # AudioSeparation + AudioCrop

models:
  # --- WAN 2.1 14B Image-to-Video 480p, GGUF Q8 (main transformer) ---
  # VERIFY: city96 hosts these; the repo filename differs from the workflow's
  # local name (rename on download or update the LoadModel widget to match).
  - url: https://huggingface.co/city96/Wan2.1-I2V-14B-480P-gguf/resolve/main/wan2.1-i2v-14b-480p-Q8_0.gguf
    local_path: diffusion_models/Wan2.1_14b_Image_to_Video_480p_GGUF_Q8.gguf

  # --- MeiGen-AI MultiTalk weights (14B fp8_e4m3fn; kijai's available quant) ---
  - url: https://huggingface.co/Kijai/WanVideo_comfy/resolve/main/WanVideo_2_1_Multitalk_14B_fp8_e4m3fn.safetensors
    local_path: diffusion_models/WanVideo_2_1_Multitalk_14B_fp8_e4m3fn.safetensors

  # --- umt5-xxl text encoder (fp16; compatible with LoadWanVideoT5TextEncoder) ---
  - url: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/resolve/main/split_files/text_encoders/umt5_xxl_fp16.safetensors
    local_path: text_encoders/umt5_xxl_fp16.safetensors

  # --- WAN 2.1 VAE (bf16) ---
  - url: https://huggingface.co/Kijai/WanVideo_comfy/resolve/main/Wan2_1_VAE_bf16.safetensors
    local_path: vae/Wan2_1_VAE_bf16.safetensors

  # --- CLIP-Vision H (image conditioning) ---
  - url: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/resolve/main/split_files/clip_vision/clip_vision_h.safetensors
    local_path: clip_vision/clip_vision_h.safetensors

  # --- lightx2v CFG step-distill LoRA (rank64, keeps it to a few steps) ---
  - url: https://huggingface.co/lightx2v/Wan2.1-I2V-14B-480P-StepDistill-CfgDistill-Lightx2v/resolve/main/loras/Wan21_I2V_14B_lightx2v_cfg_step_distill_lora_rank64.safetensors
    local_path: loras/Wan21_I2V_14B_lightx2v_cfg_step_distill_lora_rank64.safetensors