Train Lightricks Videos Locally with New LTX 2.3 Support

By vramkickedin | | 2 min read
An orange translucent video camera floating above a digital rippling sea

AI Toolkit now includes LTX 2.3 support for fine-tuning the video generation model from Lightricks. A recent merge adds training capabilities for the latest version of this model, allowing users to adapt it with their own custom datasets.

The update, contributed by developer jaretburkett, introduces key changes for handling the new model's layer structure. This gives video creators and researchers another local option for AI training without relying on cloud-based services.

What's new in this update

  • Auto Frame Count feature loads each video at its true length automatically.
  • Support for mixing different length videos within the same dataset.
  • Updated LoRA key renaming for new LTX 2.3 layers.
  • Version bump to accommodate the new model architecture.
  • Both GUI and CLI interfaces available for training workflows.

Video creators working with varied source material can benefit from the new Auto Frame Count option. Instead of trimming or padding all clips to a uniform length, users can drop in videos of any duration and let the data loader handle the differences. This saves significant preprocessing time for anyone building custom training sets from existing footage libraries.

Notes from the development process

The contributor noted in commit messages that initial support is in place, but

'still needs a lot of testing to make sure'

everything works correctly. Users should expect potential rough edges with this early implementation. The code changes span 8 files with over 700 additions, suggesting substantial adjustments to accommodate the new model architecture.

This integration builds on AI Toolkit's existing support for various diffusion models, including FLUX.1 and SDXL. The toolkit emphasizes compatibility with consumer-grade hardware, making it accessible to hobbyists and small studios without enterprise equipment. Anyone trying the new LTX 2.3 support should test thoroughly and report issues back to the project.

See the new LTX 2.3 Support on GitHub.