LongCat, MuseTalk and Duix: local avatar pipeline

Aillex explains a local AI pipeline for making talking-head videos from one photograph and a voice track. The four stages use three distinct engines, with MuseTalk running twice. The presenter says this setup produced the channel's animated face.

LongCat supplies posture, hand movement and weight shifts. Its exaggerated mouth movement prompted the creator to keep that mouth as still as possible and replace it with MuseTalk's audio-driven lip sync. Duix, described as the open source HeyGem avatar engine, then reduces facial exaggeration. According to the presenter, that pass softens the lip shapes, so a second MuseTalk pass restores articulation.

The episode also explains failures that shaped the workflow. Repeated body footage produced recurring gestures, sometimes in reverse. Reusing a render name caused the mouth engine's face cache to composite the wrong face without an error. Small face renders led the creator to set a minimum render resolution.

The reported setup runs in and around ComfyUI on one RTX 5090; the presenter calls 32 GB of VRAM comfortable, rather than stating a minimum. A linked guide shares the stage order but withholds exact tuning values. Commercial use requires checking each model's license. The presenter also reports that LTX 2.5 lost to this pipeline for talking faces in the channel's tests, while performing well for action.