LTX 2.3 Video Generator — run Lightricks' 22B model online

LTX-2.3 is Lightricks' open-weight video model, released in March 2026. It's a 22-billion-parameter diffusion transformer that generates video and synchronised audio together, from text, an image, audio or a video, and it's an upgrade of LTX-2 with better picture, sound and prompt adherence. On Kenerate you can run it online from text or an image, 5–20 seconds at up to 1080p.

It's the full name of the model: 22B is the parameter count, and every official checkpoint carries it — ltx-2.3-22b-dev, ltx-2.3-22b-distilled and ltx-2.3-22b-distilled-fp8. There is one LTX-2.3 model; the variants differ in precision and in how many sampling steps they need, not in architecture.

The dev checkpoint is the full bf16 model: slower, tunable and the one to train LoRAs on. The distilled checkpoint runs in 8 steps with CFG 1, so it's much faster and is what most local workflows use; Lightricks later added distilled 1.1 with a different look and better audio. Hosted generators like Kenerate pick the inference settings for you.

People search "LTX 2.3 DMD" for the few-step distilled version — DMD (distribution matching distillation) is a common way to turn a model into a fast few-step one. Lightricks' official few-step checkpoint is ltx-2.3-22b-distilled (8 steps, CFG 1); its model card doesn't say which distillation method was used, so treat community files labelled "DMD" as third-party.

Yes. Lightricks publishes ltx-2.3-22b-distilled-fp8 in a separate LTX-2.3-fp8 repository on Hugging Face, and community quantisations (FP8 and INT8) exist too. FP8 roughly halves the weights' memory compared with bf16, which is how LTX-2.3 fits on 16–24 GB consumer GPUs; fast FP8 maths needs an RTX 40-series or newer card.

The official weights are on Hugging Face at Lightricks/LTX-2.3 (dev, distilled, distilled LoRA and the spatial and temporal upscalers) and Lightricks/LTX-2.3-fp8. Lightricks recommends running them with the LTXVideo nodes in ComfyUI, or with its ltx-pipelines package. If you'd rather not download anything, run LTX 2.3 in Kenerate's video studio.

The full bf16 dev model wants a 32 GB-class GPU or more; FP8 and distilled builds bring it to 16–24 GB cards with offloading, at the cost of speed. Lightricks tested the code with CUDA 12.7+ and PyTorch 2.7. In Kenerate's studio you need no GPU at all.

The open model can: its Hugging Face card lists text, image, audio and video inputs, and hosts such as Sogni expose audio-to-video and video-to-video (ControlNet, inpaint, outpaint) variants. Kenerate runs LTX 2.3 for text to video, image to video and extending a clip by up to 20 seconds — not audio-to-video, video-to-video or LoRAs.

On Kenerate, 5 to 20 seconds per generation in 16:9 or 9:16, at 480p, 720p or 1080p, and you can extend a clip by 1–20 more seconds. Locally, length depends on your GPU memory; frame counts must be a multiple of 8 plus 1.

Lightricks' guide recommends one flowing paragraph of 4–8 present-tense sentences for a single shot, with camera moves described relative to the subject and spoken lines in quotes. Prompt-generator tools simply expand a short idea into that format; you can do the same by hand — the tips on this page follow the guide.

The weights are open under Lightricks' LTX-2 Community License, not Apache 2.0. The licence allows free use with conditions, and larger companies need a commercial licence from Lightricks — read the licence in the LTX-2 repository before shipping commercial work. Videos you make on Kenerate are yours to use under Kenerate's terms.