Lightricks's open-weight video line is now led by LTX-2 (19B parameters, native synchronized audio-video, up to 4K), released under Lightricks's own community license, with the earlier LTX-Video 0.9.8 line still available as a smaller, lighter-VRAM option. Pick it to self-host video generation with real hardware behind it: LTX-2's full model is a roughly 66 GB download that needs serious VRAM, though quantized and CPU-offload options exist.
Other good fits
Lighter local hardware: the earlier 0.9.8 line (2B to 13B, with distilled and FP8 variants) still works and needs less VRAM.
+ Pros
Read pros
Lightricks's flagship open-weight video model is now LTX-2 (19 billion parameters), which natively generates synchronized audio and video in one pass, meaning dialogue, ambience and lip sync come out of the same generation rather than a separate audio step. It supports resolutions up to native 4K and up to 50 FPS, and is distributed with full inference and LoRA-training code. The earlier LTX-Video line (0.9.8, video-only) is still maintained separately as a smaller, lighter option: 2B and 13B variants, with distilled and FP8-quantized versions for less VRAM at some quality cost.
Pick it when…
You want synchronized audio and video from one open-weight model, not a separate lip-sync or audio step.
You have the GPU headroom for a large model (LTX-2's full weights are roughly 66 GB), or you're willing to use quantization and CPU/disk offloading to fit smaller hardware.
You'd rather use the lighter, older 0.9.8 line (down to a 2B distilled variant) if your hardware can't handle LTX-2's full model.
− Cons
Read cons
LTX-2's full model asks for real hardware; without quantization and offloading, a typical consumer GPU won't hold it. The project has moved fast (from 0.9.8 to LTX-2, then to LTX-2.5 within months), so a tutorial or workflow built for one release may not carry over cleanly to another; check the current repository rather than an older guide. Running low on VRAM shows up as out-of-memory errors during decoding unless VAE tiling is used, which is an extra setting to remember rather than an automatic default in every workflow.
Pick something else when…
You'd rather use a managed cloud workflow than self-host.Kling offers a comparable cloud video generation path with different licensing.
You need lip-syncing on footage you already have, not new scene generation.LatentSync is built specifically for that.
You want a node-based creative tool that also handles broader image and video workflows.ComfyUI is the platform LTX-2's recommended workflows run inside.
Avoid if
You need a fully permissive open license; check Lightricks's community license terms for your specific commercial use before assuming it's unrestricted.
Context
LTX-2.5 is the currently recommended release within the LTX-2 line; LTX-2.3 is kept as a legacy option that the same pipelines still run on. The older, separate LTX-Video 0.9.8 line (2B/13B, OpenRail-M-era licensing) remains available in its own repository for lighter local hardware. Check which line a workflow, checkpoint or tutorial actually targets before assuming compatibility.
Requirements
LTX-2 was tested with Python 3.12+, CUDA 12.7+ and PyTorch ~2.7. Its default resolution is 1024x1536 at 24 FPS, with 4K defined as 3840x2176; width, height and frame count must follow the model's divisibility rules (divisible by 32 for resolution, by 8 plus 1 for frame count). For limited VRAM, fp8-cast quantization with CPU or disk offloading is documented, and VAE tiling is usually needed to avoid running out of memory during decoding. ComfyUI integration runs through the separate ComfyUI-LTXVideo project; a DistilledPipeline gives faster, lower-step generation for quick iteration.
Licence & cost
Read licence & cost
The weights and code are free to download. Real cost is hardware: LTX-2's full model is a roughly 66 GB download and needs serious VRAM even with quantization and offloading tricks; the older 0.9.8 line's 2B distilled variant is far lighter if your GPU can't handle LTX-2. If you'd rather not manage hardware, hosted routes exist through Fal, Replicate, ComfyUI-hosted services and LTX Studio, billed per generation. Licensing is Lightricks's own community license (ltx-2-community-license-agreement) for LTX-2, distinct from the OpenRail-M terms on the older 0.9.8 line; check the specific terms for your intended use, including any commercial conditions, before relying on either.
First task
Documentation-based suggestion — not run here.
Read first task
Generate one short clip and check it against your prompt.
Check your GPU's VRAM and pick LTX-2 with quantization/offloading, or the older 0.9.8 line, based on what your hardware can handle.
Set up ComfyUI with the LTX video workflow, or use the documented inference script directly.
Write a short, specific prompt describing a scene with no real person's likeness.
Generate a short clip and watch it fully, checking it against what you described, including whether the audio (if using LTX-2) matches the visual.
Done when the clip matches the basic content of your prompt, plays at the resolution and frame count you asked for, and, if you used LTX-2, the audio is reasonably in sync with the visuals.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on the LTX-2 and legacy LTX-Video GitHub repositories and the LTX-2 Hugging Face model card, checked 23 September 2026. No model was downloaded, installed or run for this review. Exact commercial-use conditions of Lightricks's community license were not independently verified beyond what the repository states; check the license file for your specific use.
LTX-2 repository · Sep 23, 2026 — 19B parameters, ~66 GB full download, fp8-cast quantization and offloading, resolution/frame-count rules, ComfyUI and LoRA-training routes.
Generate video from text or images through ComfyUI workflows.
Explore model-specific audio-video generation paths.
Pin releases before planning hardware or production workflows.