LTX-2.3 22B
Model, video model, by Lightricks
Assessment
The heaviest thing on a two-card box and the only one that makes a face speak convincingly. Worth it for finished pieces, not for experiments. Budget half an hour a clip and give the LLM a fallback for the whole window.
2026-10-09
Strengths
- Lip sync and voice-over: a face that speaks given audio, at quality that survives a close look
- Long clips in one pass when split across two cards
- The same model set serves plain video generation
Limitations
- Twenty to thirty minutes per 40-second clip on two 12 GB cards
- Needs both cards, so every other GPU user waits, and a queue of three keeps the LLM away for over an hour
LTX-Video is Lightricks' open video model family; the 2.3 release at 22B parameters is the one that made lip sync and voice-over practical on consumer hardware, provided two cards can be split with a DisTorch-style loader in ComfyUI.[1] It generates video with its own sound, and conditions well on a reference face and an audio track.
What it is for
Lip sync: a person in a still or a short clip speaks the words in an audio file. Voice-over: the same without a source recording, driven by a text-to-speech track. Plain video generation on the same weights, where it competes with Wan at higher cost and often higher polish.
What it costs
Twenty to thirty minutes for a 40-second clip on two RTX 3060s, and the model stays loaded between queued jobs, so three clips keep the cards for well over an hour.[2] Any language model that lived on those cards is gone for the duration; the fallback tier is what answers.
Things to know
- Keep the reference face frontal and well lit; the model forgives little there.
- Render once at the target length rather than stitching short pieces; the join is where it shows.
- An hour-long Studio window after the last job is a design choice, not a bug: it avoids reloading 22B of weights for the next clip.
History
- 2026-10-09: article written and published.
- 2026-10-09: entry created.
Comments
Public comments on each entry are coming. Nothing is collected here yet.