LongCat-Video-Avatar

LongCat-Video-Avatar 1.6 is an audio-driven digital human video model from Meituan LongCat, supporting both single-person and multi-person scenarios. Officially, it claims improvements in lip-syncing, expression and body coordination, and consistency in long videos exceeding 10 minutes. Comfy-Org has repackaged it into files for ComfyUI: an int8 main model, two rank-128 LoRAs (dmd and opsd), the umt5_xxl text encoder (available in fp16 and int8 versions), and the Wan 2.1 VAE, along with a list of the directories where each file should be placed.

LongCat-Video-Avatar 1.6 is an audio-driven digital human video model from Meituan LongCat, supporting both single-person and multi-person scenarios. Officially, it claims improvements in lip-syncing, expression and body coordination, and consistency in long videos exceeding 10 minutes. Comfy-Org has repackaged it into files for ComfyUI: an int8 main model, two rank-128 LoRAs (dmd and opsd), the umt5_xxl text encoder (available in fp16 and int8 versions), and the Wan 2.1 VAE, along with a list of the directories where each file should be placed.

Why it matters The original version requires manually configuring conda, FlashAttention, and two sets of dependencies; the official package provides an int8 main model, two LoRAs, a text encoder, and a VAE, eliminating the need for manual conversion.

Hug Face Likes 7 · Hug Face Upvotes 7 · Hug Face Downloads 0

License: Commercial use allowed · ComfyUI: Official package · VRAM: Not specified

Reasons

  • Published by the official Comfy-Org account, MIT license, original model allows commercial use
  • The original model card includes a technical report and human preference evaluation

Usage

  • Place the files into the diffusion_models, loras, text_encoders, and vae directories under ComfyUI/models respectively; the package page does not include a workflow, so refer to official ComfyUI templates or update notes when setting up

Risks

  • The package page only lists files and directories, without specifying VRAM requirements or including a workflow
  • Digital human videos involve portraits and voices; authorization is required to use others' likenesses

Open

抱脸 image-to-video 新模型

Explore the latest AI Radar updates