# LongCat Avatar 1.6 custom nodes: talking avatar from one photo and a voice file, install from ComfyUI Manager (free, MIT) — retained record

> Retained reddit record from reddit:comfyui.

Canonical: https://highsignal.app/data/records/c0c78085917c457f

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:comfyui)  published 2026-11-19 15:03 ingested 2026-11-20 02:39

[original source ↗](https://www.reddit.com/r/comfyui/comments/1y0bk5q/longcat_avatar_16_custom_nodes_talking_avatar/)

## retained text



```
one photo + one voice clip in, lip-synced talking video out. these 3 are one run each (photo + ElevenLabs line, 480p, seed 42, no lora, no editing). sound on.

install: Manager > Install via Git URL > paste the repo link, restart. the workflow is in Browse Templates. you only add the umt5 text encoder by hand, the rest (about 25 GB) downloads on the first run.

the nodes: Model Loader, Text Encode (runs on the CPU), Whisper Loader + Audio Encode, Vocal Extract (optional, takes a voice out of a song or a noisy clip), Sampler. videos can be any length, it continues window by window (93 frames, about 3.7 s each).

settings that matter: attention\_mode auto picks SageAttention, then FlashAttention 2, then PyTorch. mix\_steps 4 is the official 1.6 LoRA switch, 0 is about 15% faster but needs more VRAM. block\_num 1 streams the model through the GPU, 0 loads it all and needs far more VRAM. 480p is the default, 720p is about 3x slower per step.

speed on a 5090 32GB + 64GB RAM, Windows, 480p, for a 3.8 s clip: 7.7 min before SageAttention, 6.6 min with it, 5.6 min with SageAttention + mix\_steps 0. each extra 3.2 s of audio adds one window (1.5 to 2.5 min). I haven't tested cards under 32GB yet, if you try one tell me how it went.

built on rookiestar28's 1.5 pack (MIT), model by the Meituan LongCat team (MIT).
```



document fetched 2026-11-20 00:42 · canonical https://reddit.com/r/comfyui/comments/1y0bk5q/longcat_avatar_16_custom_nodes_talking_avatar
