HighSync Model Loader

HighSync Model Loader is where all the expensive work in this pack actually happens. The lip-sync nodes themselves are fairly quick diffusion passes; this node is the one that reads a few gigabytes of weights off disk, builds the whole model stack, and moves it onto your GPU. You run it once at the top of the graph, and it hands you a single highsync_model object that both lip-sync nodes expect as their first input.

Under the hood it assembles everything HighSync needs: the SD VAE (AutoencoderKL), the reference UNet, the EchoUNet3D denoising UNet from the two .pth checkpoints, a zero-SNR DDIM scheduler built from the model's config (v-prediction, trailing timestep spacing), and the Whisper tiny audio processor that turns your audio into conditioning features. It targets your main GPU via ComfyUI's own model_management, so it plays nice with the rest of your graph's memory accounting.

Two inputs, both easy to leave alone:

• models_dir - the folder holding the downloaded weights. Default is D:/ComfyUI/models/highsync; wire this from the HighSync Download Models node's output if you've run that, or type in your own path.
• weight_dtype - fp16 (default), bf16, or fp32. fp16 is the right call here: these are SD1.5-class weights, not 14B video models, so precision barely matters and fp16 keeps VRAM comfortable. Reach for bf16 if your card happens to prefer it; only go fp32 if you're chasing some numerical weirdness, since it roughly doubles the footprint.

One output: highsync_model, type HIGHSYNC_MODEL. That's not a real tensor - it's a bundle (config, pipeline, audio processor, device) that only the lip-sync nodes understand, so don't expect to inspect it. Just drag it into HighSync LipSync or LipSync Advanced.

Install

Same story as the rest of the pack - clone into custom_nodes, install the requirements into ComfyUI's Python env, restart:

cd ComfyUI/custom_nodes
git clone https://github.com/magicwang1111/ComfyUI-HighSync-Wrapper
cd ComfyUI-HighSync-Wrapper
pip install -r requirements.txt

Or search "ComfyUI-HighSync-Wrapper" in ComfyUI Manager and let it handle it. Restart afterwards either way.

Troubleshooting

• First load is slow. Reading both UNets, casting to fp16, and uploading to VRAM takes a while. The good news: results are cached per (models_dir, weight_dtype, device), so re-running the workflow is fast. The slow load is a one-time tax per session.
• "install requirements.txt first" error. The loader imports diffusers and omegaconf and will refuse to run without them. That's the pip step you skipped, not a bug.
• Missing-file error on load. The loader runs the same validation as the download node; if it fails, run HighSync Download Models with download=True (or drop the files in by hand) and point models_dir at the right folder.