Lip-syncing a 3:20 AI music video on one RTX 5090: LTX + Wan 2.2 workflow

keto_brain · reddit · 2026-11-05

A creator shares the full ComfyUI pipeline behind "Low Battery" — a 3:20, two-character, 42-shot AI music video rendered entirely on one RTX 5090 via ComfyUI's HTTP API, with Claude Code writing the graph-building and cutting scripts.

Core trick: make LTX lip-sync to your vocal. Encode an isolated vocal (split with Demucs) via LTXVAudioVAEEncode, freeze it with SetLatentNoiseMask(mask=0), concat with the video latent, and sample with CFG 1.0 / euler / ManualSigmas — all core nodes.

Settings that mattered:

Wan 2.2 (14B I2V, Lightning LoRA 0.3, 5 steps 2+3, RIFE x2) handles acting shots: 121 frames is the sweet spot; trim frame 0; prompt action sequences not states; let Wan do the "entering" motion; pre-crop input to 1024x704; hand off a Wan frame as the next LTX shot's still. Shots are Sing (LTX), Act (Wan), or Graphic (HTML — both models garble text).

Pitfalls: "slow push-in" runs away into extreme close-ups; calm seeds win; never seed from mp4 frames (h264 blocks animate as crawling squares) — use lossless PNGs. 42 shots cost 61 LTX takes + 33 Wan takes. Total cost: 8,800 ElevenLabs credits, a Claude subscription, $1 of electricity; all models are free and open.

Related event: Creator Renders 3:20 AI Music Video Entirely on a Single RTX 5090(2 posts)→

Original post →

More from coding & agent

coding & agent channel →