Anymatix Audio Encoder Loader

When models learned to generate audio, they borrowed the whole playbook from image generation - including latent space. The trick that makes Stable Audio and its relatives work is an audio encoder: it turns a waveform into a compressed audio latent, the diffusion model denoises that latent, and a decoder turns it back into sound. AnymatixAudioEncoderLoader is the node that loads that encoder, and it's the audio-side cousin of the VAE loaders you already know from image work.

In ComfyUI terms it's a niche but real category. The loader takes a single string, resolves it against models/audio_encoders (after stripping to a basename), reads the weights, and hands you an AUDIO_ENCODER object. There's no stock-download magic here - the file has to exist, which is where AnymatixFetcher comes in: point it at the encoder's URL with type audio_encoder, and this loader gets its file on first run.

What it feeds

The output plugs into the audio side of your workflow the way a VAE plugs into the image side - an encode node turns your audio into latents for generation, or the encoder is part of a larger audio-model chain. If you're doing image-to-audio or text-to-audio work (Stable Audio style, or the newer audio-capable video models), this is the node that brings the encoder weights in. If you're only ever generating images, you'll never touch it - and that's fine.

Input and output

• audio_encoder_name (STRING) - filename (resolved against models/audio_encoders) or a fetched path to the encoder weights.
• Output: AUDIO_ENCODER → the encoder input of your audio generation chain.

Install

Standard for the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Anymatix/anymatix-comfy-nodes

then restart, or install via ComfyUI Manager (search "anymatix-comfy-nodes"). The audio encoder weights are typically a few GB and arrive by download or fetcher.

Issues to expect

Most of the pain here is outside the node. Audio generation in ComfyUI is younger and less forgiving than image work - mismatched encoder and decoder families fail in confusing ways, and model packs sometimes expect specific encoder weights that aren't in your folder yet. The loader's error message when a file is missing is honest, so if you see "audio encoder file is invalid" or a file-not-found, the first thing to check is whether the right weights actually landed in models/audio_encoders. Pair the loader with a matching fetcher and it mostly stops being a problem.