Load Audio Encoder

Most of the model loaders in ComfyUI hand you something that turns text into images. Load Audio Encoder is different - it hands you a model that turns audio into a representation the diffusion pipeline can condition on. This is the node behind audio-to-video workflows: you feed it a trained audio encoder, and further down the graph the soundtrack's content can steer generation. If your workflow has a "reference audio" input, this loader is almost certainly in its ancestry.

It shipped with core in late August 2025, added specifically to support wav2vec2 as an audio encoder. ComfyUI's audio encoder stack currently knows two formats - wav2vec2 and whisper - and this loader is how you get either one into a graph.

How it works

The node scans models/audio_encoders/, loads the selected file with safe_load=True, and builds the encoder object. If the file isn't a recognized audio encoder, it raises a RuntimeError telling you so - no guessing, no silent fallback. The output is a single AUDIO_ENCODER object.

That object doesn't do anything on its own. It becomes useful when you wire it into AudioEncoderEncode (category model/conditioning), which takes the encoder plus an AUDIO input (waveform + sample rate, from Load Audio or similar) and produces the encoder output embeddings. From there those embeddings feed into the conditioning side of audio-aware video models - in core today, the Wan video nodes are the main consumer of this chain.

The one input that matters

• audio_encoder_name - a dropdown of everything in models/audio_encoders/. That's the entire interface. There are no optional inputs, no strength knobs, no dtype switches.

One output:

• AUDIO_ENCODER - the loaded encoder, ready for AudioEncoderEncode.

Getting the encoder

Drop a wav2vec2 (or whisper) safetensors file into ComfyUI/models/audio_encoders/ and refresh the node. Note that ComfyUI doesn't fetch these for you - audio encoders aren't in the standard "download me" set, so you'll be pulling the weights yourself from wherever your chosen video model expects them.

Common issues

The dominant gotcha is format. This loader is not a generic "load any audio model" - it expects one of the formats ComfyUI's audio encoder code understands. An ONNX export or a raw Whisper checkpoint in the wrong layout will error at load time. If you're following a workflow that uses this node, grab the exact model file it names rather than improvising.

Second, don't expect this node to be useful in a text-to-image graph. It exists for audio-conditioned video generation, and wiring it up without the corresponding encode and conditioning nodes gets you nothing but a dead wire. It's a niche loader - but if you've ever wondered how those "make this clip say anything" and audio-to-video workflows actually hear the soundtrack, this is the first stop.