LTX-2

// Overview

LTX-2 is the official Python inference and LoRA-trainer package for Lightricks' LTX-2 family of audio-video generation models. The project describes LTX-2 as the first DiT-based audio-video foundation model to carry all the core capabilities of modern video generation in one model — synchronized audio and video, high fidelity, multiple performance modes, production-ready output, API access and open access. In practice this repository is how you run those models on your own hardware rather than through a hosted service.

The recommended checkpoint is LTX-2.5, published on Hugging Face as one file per component so you download only the parts a given pipeline needs: a 22B transformer (a `dev` build for the guided two-stage pipelines, or a `distilled` build that runs in far fewer steps), a Gemma 4 12B text encoder fine-tuned for LTX with the projection bundled in, separate video and audio VAEs, and latent spatial and temporal upscalers. The quick-start download is roughly 66 GiB. Stock Gemma 4 is not a substitute for the bundled encoder — loading checks the encoder version against the one the checkpoint was trained with.

Around those weights sits about a dozen pipelines. `DistilledPipeline` is the fast starting point; `DFRPipeline` (Diffusion Fidelity Rendering) is the production-quality text/image-to-video path, adding generated keyframes and a spatial detailing IC-LoRA at the cost of runtime and VRAM. Beyond those there are guided two-stage and single-stage text/image-to-video pipelines, IC-LoRA video-to-video, keyframe interpolation, audio-to-video conditioned on an input audio file, a Retake pipeline that regenerates a specific time region of an existing video, an HDR IC-LoRA path that writes linear float suitable for EXR export, and Dub-It, which rephrases speech while matching speaker identity and lip movement.

The documentation is unusually direct about the memory/quality trade-offs. Defaults are 1024×1536 at 24 fps, with UHD 4K available at 3840×2176. `--quantization fp8-cast` with `--offload cpu` or `--offload disk` brings the footprint down on constrained GPUs, FP8 scaled matrix multiplication is available on Hopper and later, and FlashAttention 3 or 4 wheels can be installed for supported hardware with PyTorch SDPA as the automatic fallback. Every pipeline in the repository also still runs on the legacy LTX-2.3 checkpoints, though files are not interchangeable between the two model generations and a LoRA only works with the model it was trained on. The code and weights are released under the LTX Community License Agreements rather than a standard OSI licence.

// What it does

• Synchronized audio and video generation from a single DiT-based model
• DistilledPipeline for fast generation and DFRPipeline for production quality using the same distilled transformer plus a detailing IC-LoRA
• Guided two-stage and single-stage text/image-to-video pipelines with CFG/STG and 2x upsampling
• Keyframe interpolation, audio-to-video, video-to-video via IC-LoRA, and a Retake pipeline that regenerates one time region of an existing clip
• Dub-It: rephrasing speech while matching speaker identity and lip movement
• Native HDR and EXR support with BT.2020/HLG mastering
• Component-wise weight downloads, so you fetch only the transformer, VAEs, upscalers and LoRAs a pipeline needs
• Optional duration head that predicts clip length from the prompt, letting you omit --num-frames
• FP8 quantization, CPU/disk offload and FlashAttention 3/4 backends for memory- and speed-constrained hardware
• A LoRA trainer alongside inference, for adapting the model to your own footage

// Getting started

Clone the repository, sync dependencies with uv, download the LTX-2.5 components for the pipeline you want, then generate. The natten extra is the fastest backend for the diffusion video VAE and is Linux + CUDA only — the same command works elsewhere because it is skipped automatically.

git clone https://github.com/Lightricks/LTX-2.git
cd LTX-2

uv sync --extra natten

hf auth login
hf download Lightricks/LTX-2.5 \
    diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
    text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
    vae/ltx-2.5-video-vae-bf16.safetensors \
    vae/ltx-2.5-audio-vae-bf16.safetensors \
    latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
    --local-dir models/ltx-2.5

uv run python -m ltx_pipelines.distilled \
    --transformer-path       models/ltx-2.5/diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
    --text-encoder-path      models/ltx-2.5/text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
    --video-vae-path         models/ltx-2.5/vae/ltx-2.5-video-vae-bf16.safetensors \
    --audio-vae-path         models/ltx-2.5/vae/ltx-2.5-audio-vae-bf16.safetensors \
    --spatial-upsampler-path models/ltx-2.5/latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
    --num-frames 121 \
    --seed 42 \
    --output-path output.mp4 \
    --prompt "A medium close-up shot of a man looking directly at the camera, speaking in a satisfied tone. The camera remains static throughout, with a shallow depth of field keeping him in sharp focus against a softly blurred beige wall."

hf download Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler \
    ltx-2.5-22b-ic-lora-pixel-spatial-upscaler-x2-1.0.safetensors \
    --local-dir models/ltx-2.5/loras

uv run python -m ltx_pipelines.dfr_pipeline \
    --transformer-path models/ltx-2.5/diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
    --detailing-lora   models/ltx-2.5/loras/ltx-2.5-22b-ic-lora-pixel-spatial-upscaler-x2-1.0.safetensors \
    --num-frames 121 --seed 42 --output-path output_dfr.mp4 \
    --prompt "..."

--quantization fp8-cast --offload cpu

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Generating short clips with matching dialogue and sound effects from a single prompt, without a separate audio pass
• Turning a still image or a pair of keyframes into motion for storyboards and previsualisation
• Fixing one bad stretch of an existing clip with the Retake pipeline instead of regenerating the whole thing
• Dubbing or rephrasing a line while keeping the speaker's identity and lip movement with Dub-It
• Producing HDR footage in a linear float workflow that exports to EXR for grading and tonemapping
• Fine-tuning a LoRA on your own footage for a consistent character or visual style

// How LTX-2 compares

LTX-2 alongside other open-source video generation tools AI/TLDR tracks, ranked by GitHub stars.