generated: '2026-07-19'
method: searched
source: https://ltx.io/changelog
docs: https://ltx.io/changelog
title: LTX Models Changelog
scope: model
scope_notes: >-
  Lightricks publishes a dated changelog for the LTX model generations and the ComfyUI plugin, not for
  the REST API surface itself. API-surface change is announced separately as dated deprecation notices
  in the docs (see lifecycle/lightricks-lifecycle.yml). Detailed per-release notes live on GitHub.
versioning_scheme: model-generation
current_version: LTX-2.3
tracks:
- name: LTX Model
  entries:
  - version: LTX-2.3
    date: '2026-03-05'
    title: LTX-2.3 22B Model, Native Audio, Portrait Video, Desktop Editor
    breaking: false
    additions:
    - 22-billion parameter architecture, upgraded from 19B
    - Synchronized audio and video generated in a single forward pass
    - Up to 4K resolution at 50 FPS, up to 20 seconds per clip
    - Native portrait (9:16) vertical video
    - 24 and 48 FPS output options; last-frame interpolation
    - Spatial and temporal upscalers; desktop video editor for local generation
    - 'LoRA adapters: Motion-Track-Control, Union-Control (depth/pose/canny), Detailer'
    highlights:
    - Redesigned high-fidelity VAE; reduced texture drift over longer sequences
    - Upgraded HiFi-GAN vocoder, stereo audio at 24 kHz
    - 4x larger text connector for better prompt adherence
    - Apache 2.0 license, fully open weights, commercial use allowed
  - version: LTX-2
    date: '2026-01-06'
    title: LTX-2 open-source release — full weights and codebase
    breaking: false
    additions:
    - Full open-weights release on Hugging Face and GitHub
    - Fast / Pro / Ultra model variants
    - Text-to-video, image-to-video, video-to-video, audio-to-video
    - Multi-input conditioning (text, image, video, audio, depth maps, reference video)
    - Keyframe interpolation pipeline; LoRA fine-tuning in under an hour
    highlights:
    - First production-ready audio-video generation model with truly open weights
    - Technical report published (arXiv 2601.03233)
    - HD through 4K output, up to 10 seconds at launch
  - version: LTX-2
    date: '2025-10-23'
    title: First unified audio-video foundation model
    breaking: false
    highlights:
    - First DiT-based audio-video foundation model with synchronized generation
    - Native 4K fidelity at up to 50 FPS
    - Multi-keyframe conditioning; up to 50% lower compute cost vs competing models
  - version: v0.9.8
    date: '2025-07-16'
    title: 60-second generation, IC-LoRA Detailer, distilled 2B model
    breaking: false
    additions:
    - Extended generation up to 60 seconds
    - New 2B distilled model; FP8 quantized weight options
    - Control LoRA support for depth, pose and canny
  - version: v0.9.7
    date: '2025-05-14'
    title: Distilled 13B — HD video in 10 seconds
    breaking: false
    highlights:
    - HD generation in ~10 seconds on H100; no classifier-free guidance required
    - LoRA variant runs in 1GB VRAM
  - version: v0.9.7
    date: '2025-05-05'
    title: 13B model, multi-scale rendering pipeline
    breaking: false
  - version: v0.9.6
    date: '2025-04-15'
    title: Resolution expansion, 15x faster distilled variant
    breaking: false
  - version: v0.9.5
    date: '2025-03-05'
    title: Commercial license, keyframes, video extension
    breaking: false
    highlights:
    - Commercial use permitted under OpenRail-M
    - Keyframe conditioning and video extension added; LTX Studio launched
  - version: v0.9.1
    date: '2024-12-20'
    title: STG / PAG support, macOS, research paper
    breaking: false
  - version: v0.9.0
    date: '2024-11-21'
    title: Initial release
    breaking: false
    highlights:
    - Text-to-video and image-to-video, 5-second clips at 768x512
    - ComfyUI integration, Hugging Face model repository, native Python inference
- name: ComfyUI Plugin
  entries:
  - version: LTX-2.3
    date: '2026-03-05'
    title: LTX-2.3 day-0 support
    breaking: false
    requirements: ComfyUI 0.16.1+
  - version: LTX-2
    date: '2026-01-06'
    title: Performance optimizations and NVIDIA partnership
    breaking: false
    highlights:
    - Native PyTorch-CUDA optimizations for RTX GPUs, up to 3x faster
    - Up to 60% VRAM reduction
api_change_announcements:
- date: '2026-07-15'
  kind: deprecation
  summary: LTX-2 (ltx-2-fast, ltx-2-pro) deprecated; requests transparently served by LTX-2.3 at LTX-2 prices.
  url: https://docs.ltx.io/ltx-2-deprecation
- date: '2026-08-15'
  kind: removal
  summary: LTX-2 removed; requests specifying ltx-2-fast or ltx-2-pro return an error.
  url: https://docs.ltx.io/ltx-2-deprecation