Downloads

The main LTX-2.3 release includes the full development checkpoint, distilled checkpoints, distilled LoRAs, and dedicated upscalers. The official repository lists ltx-2.3-22b-dev as the full trainable model, while the distilled variants are intended for faster inference. The same repository also provides x1.5 and x2 spatial upscalers and an x2 temporal upscaler.

Development: ltx-2.3-22b-dev.safetensors — the full LTX-2.3 model for flexible inference and training. Distilled: ltx-2.3-22b-distilled.safetensors — 8-step distilled model. Distilled 1.1: ltx-2.3-22b-distilled-1.1.safetensors — updated distilled version with a different aesthetic and improved audio. Distilled LoRA: ltx-2.3-22b-distilled-lora-384.safetensors and the 1.1 variant — LoRA versions intended to bring the distilled behavior to the full model.

For users with limited VRAM, Lightricks also publishes quantized LTX-2.3 repositories. The official FP8 release contains ltx-2.3-22b-dev-fp8 and ltx-2.3-22b-distilled-fp8. Lightricks also publishes an NVFP4 variant for supported hardware and workflows.

The official LTX-2.3 repository includes separate spatial and temporal upscalers for multi-stage workflows. The available models include an x1.5 spatial upscaler, an x2 spatial upscaler, and an x2 temporal upscaler. These are supporting models rather than replacements for the main generation checkpoint.

The official LTX-2.5 repository is the primary download location for local use. Unlike the single-file LTX-2.3 checkpoints, LTX-2.5 weights are split into components following the ComfyUI-aligned structure. This means a complete local installation requires downloading the appropriate set of components rather than one large checkpoint file.

The LTX-2.5 repository includes the distilled transformer, Gemma 4 text encoder with LTX projection, video and audio VAEs, and the duration head used by supported pipelines. Additional files cover latent upscaling, LoRAs, and model patches. Always download a compatible set of components from the same LTX-2.5 release rather than mixing files from different model generations.

The official repository provides a documented download example for the LTX-2.5 distilled split pack, including ltx-2.5-22b-distilled-transformer-bf16.safetensors, gemma4-12b-with-proj-ltx-2.5-bf16.safetensors, ltx-2.5-video-vae-bf16.safetensors, ltx-2.5-audio-vae-bf16.safetensors, and ltx-2.5-duration-head-bf16.safetensors.

The LTX-2.5 ecosystem also contains quantized model variants intended to reduce memory requirements on compatible hardware. The official Hugging Face collection groups base models together with quantized models, LoRAs, and IC-LoRAs. Check the individual model card before downloading a quantized variant because supported hardware and pipeline requirements can differ.

LTX generation also requires a compatible text encoder. LTX-2.3 uses the Gemma 3 12B text encoder documented by Lightricks, while LTX-2.5 moves to a Gemma 4 12B text encoder with an LTX-specific projection. Use the text encoder specified by the corresponding LTX model documentation rather than assuming that encoders are interchangeable between versions.

Lightricks publishes additional LTX-2.3 and LTX-2.5 LoRAs and IC-LoRAs through Hugging Face. These include creative adapters, video-to-video controls, upscalers, audio-related adapters, and other specialized transformations. Community-created models and LoRAs are also available through the Hugging Face model trees, but they should be evaluated separately from the official Lightricks releases.

The safest source for LTX model weights is the official Lightricks organization on Hugging Face and the official LTX-2 GitHub repository. Avoid downloading large model files from unofficial mirrors when the same checkpoint is available from the original publisher.