LTX opens LTX Trainer to train its LoRAs on the LTX-2 model.

Lightricks launches LTX Trainer, an open-source framework to train LoRAs and fine-tune its LTX-2 video-audio model using an agentic assistant.

Lightricks, the publisher of LTX, completes its LTX-2 video-audio generation model with LTX Trainer, an open-source framework for training and refining the model rather than relying solely on prompts. From a single framework, it manages the training of LoRAs and IC-LoRAs across video, audio, cross-modal, or reference-conditioned workflows, as well as full fine-tuning.

Numerous modes are covered: text-to-video or audio, image-to-video, video and audio extension and inpainting, video outpainting, or generation of one from the other. Additionally, it offers flexible conditioning and freely available IC-LoRAs.

The most significant new feature is an agentic skill: an assistant that examines data and hardware, selects the appropriate training mode, prepares the dataset, starts the job, and monitors it, using documentation as a reference. The entire system is fully open source, available on GitHub and Hugging Face. Regarding hardware, training targets an 80 GB Nvidia GPU, with a lightweight INT8 configuration allowing for use with 32 GB cards.