rookiestar28/comfyui-longcat-avatar

This repository is an extension for ComfyUI, a visual workflow tool used to run AI image and video generation pipelines. The extension adds support for LongCat Avatar 1.5, an AI system that generates realistic talking-head videos of a person from a still photo and an audio recording. You provide an image of a person and a spoken audio file, and the model produces a video where the person's lips and expressions move in sync with the audio. The system works on NVIDIA graphics cards only and requires a substantial amount of model files downloaded separately from Hugging Face. The core model is called a DiT (Diffusion Transformer) and comes in several formats: a set of official weight files split across multiple files, a compressed INT8 version that uses less memory, or a single merged file for simpler loading. Audio is processed using Whisper large v3, an audio transcription model repurposed here to drive the lip and facial animation. A separate small model called a distill LoRA is also required and makes the generation work in just 8 steps rather than many more. Once installed and configured inside ComfyUI, you connect nodes in a visual graph to load the model, encode text, encode audio, and generate the video. The extension supports generating at 480p or 720p resolution, working with a single audio track or two simultaneous audio tracks for two-person scenarios, and optionally separating vocals from background music before feeding audio into the model. Installation involves cloning this repository into the ComfyUI custom nodes folder and installing Python dependencies. FFmpeg must also be installed on the system separately. The extension intentionally does not install PyTorch or attention acceleration libraries to avoid breaking an existing ComfyUI setup, those must already be in place. The README lists which model files go in which ComfyUI folders and provides direct download links for each. Several attention backends are supported for users who want faster inference.