Echo_LoadModel

Echo_LoadModel is the brain of the ComfyUI_EchoMimic pack. It's the node that decides which EchoMimic generation you're running, which way it's driven, and how the whole thing fits in your VRAM. Everything else in the chain - Echo_Predata and Echo_Sampler - is downstream of the choices you make here. It's also where most people's setup either succeeds quietly or fails loudly, because it's the node that actually has to find your weights.

EchoMimic is Ant Group's audio-driven portrait animation line (make a photo talk with an audio file), and this pack wraps V1, V2, and V3 behind one loader. The biggest choice is version:

• V3 - the current one, built on the Wan2.1-Fun 1.3B transformer. Unified multi-modal human animation, and the only line with the "flash" variant that runs 768×768 in ~12GB at 8 steps. This is the one you want for new work.
• V2 - semi-body animation (upper body, gestures) from late 2024.
• V1 - the original landmark-conditioned talking head.

Next is infer_mode, the driver:

• audio_drived / audio_drived_acc - audio-only: a photo plus a voice becomes a talking head. The _acc variant is the accelerated model (~6 steps, slight quality cost).
• pose_normal_dwpose / pose_normal_sapiens - pose-driven: a driving video's body motion moves your character. These two are equivalent in output; they differ in which pose extractor runs.
• pose_acc - the accelerated pose model.

The rest of the inputs are memory strategy and extras. lowvram (default true) and teacache_offload (default true) are the offloading switches - keep them on unless you have headroom and want maximum quality. use_mmgp is V3-only memory management with six choices from None to HighRAM_HighVRAM; the README's own tip is that setting lowram False disables mmgp's fp8 quantization and gives you better output quality if you can afford it. denoising is on by default and usually stays that way. The vae dropdown lists your installed VAEs (V3 wants Wan2.1_VAE, V1/V2 want sd-vae-ft-mse), and lora defaults to None - drop in a LightX2V speed LoRA here for the V3 10-step fast path.

It has two outputs and they split the work: model (MODEL_PIPE_E) goes to Echo_Sampler, and info (MODEL_INFO_E) goes to Echo_Predata, telling the preprocessor which version's face/audio pipeline to run.

The real gotcha is the model layout. This pack reads specific paths under ComfyUI/models/echo_mimic, and the V3 tree alone runs to six-plus files: the Wan2.1 transformer, Wan2.1_VAE, an fp8 UMT5 text encoder in models/clip, clip vision, wav2vec2, and the BadToBest/EchoMimicV3 transformer. Each version has its own subfolder layout - V1 wants denoising_unet.pth, face_locator.pth, motion_module.pth, reference_unet.pth, V2 wants a v2/ folder plus yolov8m.pt and a Sapiens pose model. The README's tree is explicit; follow it literally, and note that some models auto-download on first run while V3's retina-face model needs an internet connection to fetch retinaface.h5.

Install is the pack install, shared across all three nodes: ComfyUI Manager → search "EchoMimic", or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/smthemex/ComfyUI_EchoMimic.git
pip install -r requirements.txt

Restart, then check the version-specific dependencies: V1 needs pip install --no-deps facenet-pytorch; V3 wants retina-face==0.0.17, optionally mmgp, and tensorflow==2.15.0 (the README itself flags that one as dubious). If you hit an ffmpeg error, pip uninstall ffmpeg and pip install ffmpeg-python usually fixes it. And if a module is missing, the README's advice is blunt: install it. This is a capable but very hands-on pack - worth it for audio-driven talking heads, just budget an afternoon.