One RTX 5090 renders a 3:20 AI music video: LTX 2.5 lip-sync + Wan 2.2 in ComfyUI

keto_brain · reddit · 2026-11-05

A 3:20 AI music video on a single RTX 5090

The author made the second music video for an openly AI music act — Low Battery, 3:20, two characters, 42 shots — rendered entirely on one RTX 5090 via ComfyUI's HTTP API. No video editor was opened; Claude Code writes the scripts that build the graphs and cut the result.

The core trick: make LTX lip-sync to YOUR vocal

LTX is a joint audio+video model that generates its own audio if asked to sing. The fix: split the vocal with Demucs, encode it via LTXVAudioVAEEncode, and apply SetLatentNoiseMask with mask=0 so the sampler only generates video. All core nodes, no custom ones. Key settings:

Wan 2.2 for non-singing shots

LTX barely moves the body; Wan 2.2 (14B I2V + Lightning LoRA 0.3, 5 steps split 2+3, RIFE x2 to 24fps) handles walking, sitting, picking things up. Lessons: 121 frames is the sweet spot (longer drifts back to frame 0); prompt action sequences from the still's pose, not states; let Wan handle entering the frame; pre-crop inputs; a Wan frame can seed the next LTX sing shot.

Pitfalls and cost

"Slow push-in" in LTX prompts runs away into extreme close-ups; calm prompts win; never seed from darkened mp4 frames (h264 blocks get animated into crawling squares) — use lossless PNGs. 42 shots took 61 LTX takes and 33 Wan takes. Total cost: ElevenLabs Starter (8,800 credits), a Claude subscription, and $1 of electricity — every model is free and open source.

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 →