Voxta docs

LTX-2.5

Local video with a synchronized stereo soundtrack, from Lightricks' LTX-2.5 (22B, open weights), from a prompt or a first frame.

LTX-2.5 is Lightricks' 22B open-weights video model. It generates a clip and its stereo soundtrack together, from a prompt or a first frame. An end frame also works, and using the same picture for both makes a loop. A Gemma 4 12B text encoder reads the prompt. The two models never share the card.

Experimental. Only an admin can add it. Made for a 24 GB card. The distilled int8 build (~21 GB) streams the part that does not fit from RAM. The weights are gated on Hugging Face.

Setup

Accept the license

On Hugging Face, accept the LTX-2.x Community License on Lightricks' LTX-2.5 page, with the account your Hugging Face token belongs to. The token goes in Settings › Account › Connected Accounts (from the Hugging Face Integration module).

Add the service

Add Modules → Services → LTX-2.5 → Install, then follow the install steps. The Configuration step sets a folder per asset type (models, text encoders, LoRAs, VAE, upscalers), so you can point them at an existing ComfyUI collection. The Models step offers the weights to download now.

Models

SettingDefault
ModelLTX-2.5 22B distilled int8 (~21 GB, 8 steps)
Text EncoderGemma 4 12B for LTX-2.5 int8 (~15 GB)
Video VAE, Audio VAELTX-2.5's own (~1.4 GB, ~0.4 GB)
Spatial Upscaler, Temporal Upscalerx2 each (~1 GB, ~0.3 GB), used by the two-stage render

Settings

  • Two-Stage Render (on): renders at half size, upscales, then refines at full size in 3 more steps. This is Lightricks' own pipeline. Keeps fast motion whole at the same speed.
  • Half-Rate First Pass (on): with the two-stage render, drafts the motion at 12 fps and lets the temporal upscaler double it. Faster, and smoother in testing.
  • Resolution: 480p to 768p (1344x768 at 16:9). Sides are multiples of 64.
  • Duration (seconds): 2 to 10, at 24 fps. Default 5.
  • Generate Audio: keep the soundtrack. Off by default, but requests can still ask for it.
  • LoRAs: LoRAs for the LTX-2.5 transformer, merged as it loads.

On an RTX 4090, a 5-second 544p clip takes about 22 to 30 seconds with the two-stage render, once the model is loaded. Prompt conditioning is cached, so re-rolling the same prompt skips the text encoder.

On this page