Lightricks/LTX-2.3

Overview

LTX-2.3 is the Lightricks follow-up to LTX-2: an open-weights diffusion transformer that produces video and synchronized audio in a single model. Lightricks names four changes. The VAE was redesigned and trained on higher-quality data, keeping hair, fine texture, text and edges sharper. A new gated-attention text connector, four times larger than before, follows descriptions of timing, motion and expression more closely. Portrait video is native up to 1080x1920, trained on vertical footage. Audio is cleaner: silence gaps and noise artifacts were filtered from the training data, and there is a new vocoder.

For image-to-video, Lightricks reports less freezing, fewer clips that only pan and zoom across a still, and output that stays closer to the input frame. This page runs that mode: a first frame, an optional last frame and a prompt go in, and a clip with sound comes back at 1080P, 1440P or 2160P.

Choosing a variant: LTX-2 Distilled runs the earlier LTX-2 release from a single start image, with a fixed five-second length and output from 480P to 1080P. Use this page when you need an end frame, a longer take, or 1440P or 2160P.

What you can build

• Characters with spoken lines. The chef sample was prompted with a quoted monologue and comes back talking to himself over a tray of cookies, his expression changing as he goes.
• Transitions between two stills. An end image anchors the clip's last frame, for before-and-after reveals, product transformations or scene changes.
• Vertical 9:16 video for phone feeds, which LTX-2.3 generates natively.
• 1440P or 2160P shots with room to crop or reframe in the edit.
• Single takes of up to 20 seconds at 1080P for longer dialogue or a slow camera move. At 1440P and 2160P, keep clips to 10 seconds or less.

Inputs

• Start image (required): one image, uploaded or linked, in JPG, JPEG, PNG, GIF, WEBP or HEIC.
• End image (optional): one image in the same formats; the clip is conditioned to finish on it.
• Prompt: action, camera, look and sound, with dialogue in quotes.
• Duration: 6, 8, 10 or 20 seconds; Default gives 6.
• Resolution: 1080P (default), 1440P or 2160P.
• Aspect ratio: Default keeps the start image's proportions; 16:9, 9:16 and 1:1 fix the frame.
• Seed: 0 to 9999999, default 42.
• Image Strength: 0.0 to 1.0, default 1.0.

Outputs

An MP4 at 24 fps with H.264 video and stereo AAC audio. Frame size follows the resolution and aspect ratio; the samples here are a 6-second clip at 1920x1024 and an 8-second clip at 2560x1408.

Recommended settings

• Be specific. Lightricks says prompts with several subjects, spatial relationships and style instructions now resolve accurately, which is what the larger text connector is for.
• For a vertical clip, pick 9:16, or leave the ratio on Default and use a portrait start image.
• Hold the seed constant while comparing end images or prompt edits.
• The Image Strength slider does not change the result here, so its default of 1.0 can stay.

Limitations

Every run starts from an image; text-to-video is not available here. The card names English in its model details but tags eight more languages in its metadata (German, Spanish, French, Japanese, Korean, Chinese, Italian and Portuguese), so check non-English dialogue before you depend on it. It also cautions that output may drift from the prompt, might amplify societal biases or turn out inappropriate, and should not be read as factual. Scenes with no speech may get weaker audio.

Safety & compliance

The LTX-2.3 weights are covered by the LTX-2 Community License Agreement. Under that license, any company whose annual revenue meets Lightricks' threshold must obtain a paid commercial license before using the model, and content made with it has to be labeled as machine-generated when it is published. Start or end images of a real person, and any words you script for them, need that person's consent.