EchoMimic V3 Loader

If you've ever tried to install EchoMimic V3 the official way, you know the drill: clone the repo, pray through run_flash.sh, fight a CUDA build, and wonder why a talking-head demo needs a compiler toolchain. This loader is the "someone already suffered through that for you" route - it drags EchoMimic V3 (1.3B) into ComfyUI as a normal model-loading node, so the setup becomes "download three model folders" instead of "build the repo."

What it actually loads

EchoMimic V3 is Ant Group's audio-driven talking head, and it's genuinely one of the better open ones - the head motion and micro-expressions look natural rather than puppet-y. But it's not one big file. The pipeline needs a transformer (the EchoMimic-specific one that does the talking), a VAE, text/image encoders, and a wav2vec2-base-960h audio encoder. The loader walks ComfyUI/models/EchoMimicV3/, grabs all those pieces, and bundles them into a single pipe object.

One thing worth knowing from the source: if it can't find the EchoMimic transformer in the subfolder you point at, it falls back to the Wan2.1-Fun base directory for the VAE and encoders - and the code has a blunt warning that using the plain Wan transformer gives you bad results. So the folder layout matters more than almost anything else here.

The inputs that matter

• model_subfolder - the name of the folder under ComfyUI/models/EchoMimicV3/ that holds the main weights. The schema ships it defaulting to EchoMimicV3, while the README's layout names that folder Wan2.1-Fun-V1.1-1.3B-InP. Just set it to whichever folder you actually downloaded into.
• precision - bf16 is the default and the right call on a modern card.
• device - cuda. cpu and mps exist for the brave, but you won't enjoy them.

Output is one thing: pipe (ECHOMIMIC_PIPE), which wires straight into the EchoMimic V3 Sampler.

Install

The pack itself is the easy part:

cd ComfyUI/custom_nodes
git clone https://github.com/havvk/ComfyUI_AIIA.git
# restart ComfyUI

Or ComfyUI Manager → search "ComfyUI_AIIA" (publisher "hawk"). Then the part the README is blunt about: there is no auto-download for EchoMimic V3. Grab the three pieces yourself:

cd ComfyUI/models
hf download BadToBest/EchoMimicV3 --local-dir EchoMimicV3/EchoMimicV3
hf download alibaba-pai/Wan2.1-Fun-V1.1-1.3B-InP --local-dir EchoMimicV3/Wan2.1-Fun-V1.1-1.3B-InP
hf download facebook/wav2vec2-base-960h --local-dir EchoMimicV3/wav2vec2-base-960h

It'll also want diffusers>=0.30.1; the pack tries to pull that in, but if you see an import error, pip install -U diffusers fixes it.

Gotchas

• Wrong folder layout is the #1 failure mode, and it fails quietly - you'll get a talking head that drifts or looks wrong rather than an error. Verify the transformer/ folder exists in the subfolder you selected.
• Flash Attention 2 is auto-detected and gives a big speedup. It's a "strongly recommended" install in the README, not required.
• Expect this thing to want real VRAM. The README calls 24GB+ the comfortable zone. This is a 1.3B diffusion model driving video frames; nobody's running it happily on a 6GB card.