LTX-2.5 is a 22-billion-parameter open-weight world mannequin that generates synchronised video and audio from textual content, photographs or current clips, with help for native 4K HDR output.
LTX-2.5 is an open-weight world mannequin developed by LTX, an organization spun out of Lightricks. The mannequin is designed for video technology, real-time functions and bodily AI use instances resembling robotics simulation. It could generate text-to-video, image-to-video and video-to-video content material with synchronised audio, supporting resolutions from 720p to native 4K and clip lengths of six to twenty seconds.
The structure relies on a 22-billion-parameter uneven dual-stream diffusion transformer with unbiased streams for video and audio linked via bidirectional cross-attention. This permits each types of media to be generated collectively as an alternative of manufacturing the video and audio individually. It additionally makes use of a 12-billion-parameter Gemma 4 textual content encoder and immediate enhancer to interpret directions involving characters, digicam actions, lighting and actions.
One of many goals of LTX-2.5 is to scale back the computing necessities for native video technology. The distilled checkpoint makes use of an eight-step technology course of, whereas FP8 quantisation can scale back VRAM necessities by as much as 40%. Native deployment is supported via ComfyUI and Diffusers, with {hardware} necessities beginning at 16 GB of VRAM relying on the checkpoint and quantisation used.
The mannequin options native multishot technology, the place a number of consecutive photographs are generated whereas preserving character id, lighting, surroundings, sound and visible fashion inside a single technology. It additionally makes use of Diffusion Constancy Rendering, which allocates rendering effort in line with scene complexity. A brand new video decoder is designed to enhance nice particulars and stability in high-motion scenes.
LTX-2.5 is accessible as open weights on Hugging Face and is built-in into ComfyUI. Builders can even use the LTX API or an area Python-based pipeline. The mannequin could be fine-tuned utilizing the LTX-2 Coach, permitting builders to adapt it to their very own datasets and functions.
