LongCat Avatar Whisper

LongCat_Video_SM_WhisperModel (display name "LongCat Avatar Whisper") loads the audio encoder that makes the whole avatar pipeline tick. LongCat's audio conditioning isn't a spectrogram hack - it's Whisper-large-v3 embeddings, which is exactly why the model understands speech rather than just noise. This node loads that encoder from ComfyUI/models/audio_encoders/ and hands it forward as an AUDIO_ENCODER, which both LongCat Avatar Audio Encode and the vocal-extract nodes consume. It's small, it's boring, and it's mandatory - no Whisper, no talking head.

What it actually checks

Beyond loading the file, the node enforces a contract: Avatar 1.5 is Whisper-only, so Wav2Vec2 audio encoders are rejected. That's not a bug - it's a guardrail. The Avatar 1.0 era used Wav2Vec2, and the README is explicit that this pack targets 1.5 only; loading the wrong encoder "can load permissively and break lip sync," so the node fails fast with a clear error instead of letting you produce a video with a mouth that doesn't track the words. If you see that error, you're feeding it the wrong file.

The input and output

One input, one output:

• audio_encoder - a dropdown of everything in ComfyUI/models/audio_encoders/. Pick whisper-large-v3.safetensors.
• Output is an AUDIO_ENCODER that plugs into LongCat Avatar Audio Encode.

The encoder also carries the Whisper feature extractor along for the ride (the node wires both up), so downstream nodes get everything they need in one object.

Getting the model file

The README is specific about this one: download the official whisper-large-v3/model.safetensors from the Avatar 1.5 repo, rename it, and drop it in the right folder:

# after downloading model.safetensors from
# huggingface.co/meituan-longcat/LongCat-Video-Avatar-1.5
mkdir -p ComfyUI/models/audio_encoders
mv model.safetensors ComfyUI/models/audio_encoders/whisper-large-v3.safetensors

Note the auto-download on the model loader does not fetch this file - VAE, Whisper, and the distill LoRA are always manual, only the sharded DiT and text encoder are in the bounded download manifest.

Install

Shared with the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/rookiestar28/ComfyUI-LongCat-Avatar
cd ComfyUI-LongCat-Avatar
pip install -r requirements.txt

or ComfyUI Manager → search ComfyUI-LongCat-Avatar → restart. NVIDIA + CUDA required; the pack has no CPU/MPS path.

Where people get burned

Almost entirely the file problem: a user who grabbed an Avatar 1.0-era or other Whisper variant, named it something else, and gets the rejection error - rename it whisper-large-v3.safetensors and it just works. The other recurring trap is expecting the model loader's auto-download to cover this node. It doesn't. If the dropdown is empty, the file isn't in audio_encoders/. Fix that, and this node becomes a "select and move on" step you'll stop thinking about.