Open Model Radar · Dossier

LTX-2.3

A major version update to LTX-2, with improvements in image quality, audio quality, and prompt adherence; the distilled version runs 8 steps at CFG=1 and is one of the most realistic 22B video options on 16GB GPUs.

Base

Highlights

  • Distilled version ltx-2.3-22b-distilled: 8 steps, CFG=1, fast output
  • Official fp8 and nvfp4 weights provided, specifically for small VRAM
  • Supports image-to-video, text-to-video, first/last frame, and image+audio-to-video (IA2V)
  • Official recommendation is to use the built-in LTXVideo nodes in ComfyUI Manager

Scope

  • Good for: small VRAM users wanting 22B-class video
  • Good for: image-to-video + first/last-frame transitions
  • Not for: projects requiring long-duration coherent narrative

Tips

  • CFG fixed at 1; increasing it has no effect
  • Use the official upscaler for 2× resolution; it saves VRAM compared with directly raising generation resolution
  • Write audio prompts as a separate paragraph

Cautions

  • Community license; multi-stage workflows double VRAM requirements

Parameters

Distilled version
8 steps, CFG 1
Full version
Multi-step, trainable
Resolution
Mainly 720p class, with spatial upscaler for multi-stage
VAE/upscaling
Official x1.5 / x2 spatial upscalers

Deployment

16GB conclusion
Runs with fp8 or nvfp4 distilled weights; 8 steps is the key to making it viable on small cards.
VRAM note
If weights are still in the 20GB class, offload is still required; keeping duration to 3–5 seconds is most stable.
Suggested starting point
fp8 distilled + 8 steps, first output 720p/3 seconds.

Comparison

Predecessor to 2.5, with 850k downloads and still growing. Its value is that the combination of “distilled 8 steps + fp8/nvfp4” has been thoroughly run by the community, with lower trial-and-error cost than 2.5; new projects can use 2.5 directly.

Facts

Type
Base
Org
Lightricks
Released
Mar 4, 2026
Downloads
854,604
Daily downloads
3,665
License
LTX-2 Community License