Hunyuan-Foley Dependencies Loader

Nobody gets excited about the dependencies loader, which is exactly why it's the node that quietly decides whether your first foley run works. HunyuanVideo-Foley isn't one model - it's a diffusion transformer plus a small fleet of support models that do very different jobs. This node loads them all, bundles them into a single HUNYUAN_DEPS object, and hands it to the sampler so you never wire five things up by hand.

Here's what's actually inside that one wire, from the pack's source:

• DAC-VAE - the audio codec. It decodes the latent into a real waveform at the end. This is the vae_name dropdown.
• Synchformer - the video-audio sync analyzer. It watches the clip at 25fps and figures out where the sound events should land in time, which is what stops a door slam from drifting a half-second late. This is the synchformer_name dropdown.
• SigLIP2 - a vision-language model that reads the content of the video at 8fps, so the model knows it's looking at footsteps on ice rather than a vacuum cleaner.
• CLAP - a contrastive text-audio model that encodes your prompt and negative prompt.

That division of labor is worth internalizing because it explains why the model works without you doing anything clever: one branch answers "what is happening," the other answers "when is it happening," and the transformer turns both into sound.

The node has just two inputs, both dropdowns filtered from the foley folder - vae_name only lists files with "vae" in the name, synchformer_name only files with "synch." So name your downloads accordingly. The two local files you need, from huggingface.co/phazei/HunyuanVideo-Foley, are vae_128d_48k_fp16.safetensors (~743MB) and synchformer_state_dict_fp16.safetensors (~475MB), both dropped in ComfyUI/models/foley/.

Then the surprise: SigLIP2 and CLAP don't come from a file you download. The loader pulls them from HuggingFace - google/siglip2-base-patch16-512 and laion/larger_clap_general - on first run, then caches them locally so later sessions are offline. That means your first queue is the one that needs internet, and if it stalls, the logs will show it downloading a model you didn't realize was part of the deal.

Installing

Same story as the rest of the pack - ComfyUI Manager (search ComfyUI-HunyuanVideo-Foley) or:

cd ComfyUI/custom_nodes
git clone https://github.com/phazei/ComfyUI-HunyuanVideo-Foley

Then restart. The pack's requirements.txt is where the real dependencies live: diffusers, timm, accelerate, transformers, sentencepiece, einops, omegaconf, loguru, plus torchvision transforms used for preprocessing. None of these are exotic, but transformers and timm are the kinds of packages that other custom nodes fight over - if you've got a node that pins an old transformers, that's your likely conflict source, and that's a known tax on the audio layer of ComfyUI generally, not this pack specifically.

The one real troubleshooting tip: if the dropdowns come up empty, the files aren't in ComfyUI/models/foley/ - the folder is registered on import, but it won't auto-move your downloads there. Put the files in the right folder, hit refresh, and the dropdowns populate.