LongCat Video Avatar 1.6 Now Runs in ComfyUI

ComfyUI Wikinews

Meituan's audio-driven avatar model runs in ComfyUI now: Comfy-Org repackaged weights, a community native node pack, and core support in review.

LongCat Video Avatar 1.6 has a ComfyUI path. Comfy-Org repackaged the weights on November 18, a community node pack runs the model today, and Kijai's core support is open for review. The model itself shipped on November 13, when no 1.6 node path existed yet.
Two frames from each of three segments of one talking avatar clip

Two frames from each of the three 93-frame segments. Every segment continues from the last 13 frames of the one before.

Three ways in

PathStateWhat it is
Comfy-Org weightsReady to downloadOfficial int8 DiT, both LoRAs, the Whisper encoder and Wan 2.1's shared VAE and text encoder
ComfyUI-LongCat-Avatar (community)Working todayNative node pack by Pedro (LatentSpacer), 8-step distilled and 50-step base video nodes
Core support (CORE-720, Kijai)PR 17831, openAdds comfy/ldm/wan/model_longcat.py and the audio encoder wiring; not merged yet

The node pack installs from ComfyUI-Manager as LongCat Avatar Native, or with comfy node install longcat-avatar-native. It was tested on ComfyUI 0.44 and needs no packages beyond ComfyUI's own.

What you download

From Comfy-Org/LongCat-Video-Avatar, folders mirror ComfyUI/models:

FileFolderSize
longcat-avatar-1.6_int8_convrot.safetensorsdiffusion_models14.9 GiB
longcat-avatar-1.6_dmd_lora_rank128_bf16.safetensorsloras1.2 GiB
longcat-avatar-1.6_opsd_lora_rank128_bf16.safetensorsloras1.2 GiB
umt5_xxl_int8_convrot or umt5_xxl_fp16text_encoders5.3 or 10.6 GiB
whisper_large_v3_encoder_fp16.safetensorsaudio_encoders1.2 GiB
wan_2.1_vae.safetensorsvae242 MiB

The text encoder and VAE are the Wan 2.1 files, so an existing ComfyUI install probably has them already.

Speed on one 4090

The port's author measured 5.1 s per continued step at 768×512 with int8 ConvRot and ComfyUI's ck attention: a 253-frame clip (10.1 s of speech) in three segments took about 162 s. He also shipped W6A8 and W4A8 copies, 12.3 and 8.8 GiB against 14.9 GiB for int8 and 29.5 GiB for bf16, and both LoRAs apply to all of them.

A man in a red jacket sings into a microphone on a smoky stage

One clip from the node pack: the mouth follows the speech across all three segments.

The base 50-step node runs with text and audio guidance and no LoRAs, and since each step runs the model three times it is many times slower than the distilled path.

Availability

Comments

Sign in with GitHub to join the discussion.

Loading comments…
LongCat Video Avatar 1.6 Now Runs in ComfyUI | ComfyUI Wiki