Lightricks/LTX-2

Overview

LTX-2 is Lightricks' open-weights model for generating video and audio together. The model card lists 19B parameters, and the LTX-2 technical report splits them into a 14B-parameter video stream and a 5B-parameter audio stream joined by bidirectional cross-attention, so speech, foley and ambience are generated in step with the picture.

This page runs Lightricks' FP8 build of the distilled checkpoint as image-to-video. Lightricks' card says the distilled version samples in 8 steps with classifier-free guidance fixed at 1, and there is no guidance or negative-prompt setting here. You supply a still frame and a prompt and get the frame back in motion, with a soundtrack that includes any dialogue you wrote.

Choosing a variant: LTX-2.3 is the newer release and adds an optional end image, clips of up to 20 seconds at 1080P, and 1440P or 2160P output. This page fits when one start image and a five-second clip at 480P to 1080P are enough.

What you can build

• Talking-head clips from one portrait. The default sample starts from a photo of a man in glasses; in the clip he faces the camera from start to finish and says the line from the prompt.
• Toy and character animations. Another sample animates a photo of a small astronaut figure standing in a cracked eggshell on a lunar surface: the figure pushes out of the shell and drifts into space.
• Camera moves described in words. Example prompts on this page ask for a 360-degree orbit around a classic car, a slow move along a forest path and a glide toward a moonlit pool.
• Animatics: turn storyboard frames into five-second shots to check framing and how a line of dialogue plays.

Inputs

• Start image (required). Upload a file or paste a link in the Input Image URL field.
• Prompt. Action, camera, lighting and sound, with spoken lines in quotes.
• Resolution: 480P, 580P, 720P (default) or 1080P.
• Aspect ratio: Default follows the start image's shape, while 16:9, 9:16 or 1:1 give a fixed frame.
• Seed: a number from 0 to 9999999 (default 42).

There is no duration setting.

Outputs

One MP4 with H.264 video and an AAC audio track. Every run returns 121 frames at 24 fps, about five seconds; both samples here run 5.04 seconds.

Recommended settings

• Lightricks' prompting advice for LTX-2: one flowing paragraph under 200 words that opens with the main action, then covers movement, appearance, setting, camera and lighting in the order things happen.
• Script the sound in the same paragraph: who speaks, the exact words, the tone of voice, the background noise.
• Keep the seed fixed while you rework the wording, so differences between runs come from the prompt.

Limitations

• A start image is always required; there is no text-only mode.
• The card's model details list English, while its metadata also tags German, Spanish, French, Japanese, Korean, Chinese, Italian and Portuguese, so test dialogue in those languages before relying on it.
• Lightricks warns that audio without speech can be lower in quality. The astronaut sample has no dialogue, and its soundtrack is close to silent.
• The card also says the model can miss parts of a prompt, amplify societal biases or produce offensive content, and is not meant to provide factual information.

Safety & compliance

LTX-2 is released under the LTX-2 Community License Agreement. Companies with annual revenue at or above the license's threshold need a paid commercial license from Lightricks to use the model. The license's use restrictions also require generated content to be clearly disclosed as machine-generated wherever it is published. If the start image shows a real person, or the prompt gives that person a voice, get their consent first.