Video models struggle to stay still, this one learned the opposite

LTX releases the Cinemagraph LoRA for its LTX-2.3-22B video model on Hugging Face, enabling users to animate single elements while keeping backgrounds still.

LTX is releasing a LoRA called Cinemagraph for its LTX-2.3 video model. Its function: to transform a still image into a looping video where only one element moves while everything else remains frozen. A reflection rippling in sunglasses, clouds speeding across a static landscape, a neon sign flickering in an otherwise perfectly stable scene.

The highlighted angle is a reversal: where most video models struggle to keep a perfectly stable background, this LoRA has been trained to do so by default. The process comes from its training dataset, composed of scenes without moving characters, filmed with a fixed camera on a tripod to eliminate any slippage or parallax. The model thus learns to isolate the animated element and leave the rest still.

Technically, it is a LoRA fine-tuning of LTX-2.3-22B, to be loaded into ComfyUI with the trigger word CINEMAGRAPHMOTION and an explicit description of what should move and remain fixed. The tool targets single-motion scenes (water, light, reflections, neon signs) and is outside its domain when camera movement, an animated character, or multiple moving subjects are required. The LoRA is distributed on Hugging Face under the LTX-2 community license, backed by the Lightricks base model.