For custom production use cases, the biggest gains in an AI video model come from fine-tuning it, not from chasing the perfect prompt. Training LoRAs and IC-LoRAs is a way to teach a model a specific style, subject, character, or motion, then reuse it across projects with full ownership of the outputs. That's why LTX has shipped a trainer for its open weights to its enterprise customers and open-source community since the initial release of LTX-2.
Today we're announcing a major upgrade to LTX Trainer, establishing a flexible conditioning framework for training LoRAs and IC-LoRAs that spans video, audio, cross-modal, and reference-conditioned tasks directly out of the box for training LoRAs and IC-LoRAs.
What's New: A Flexible Conditioning Framework
Today’s video models can generate content from source images, video, text, and more. However, even in those instances, there’s still a level of chance involved in the generation process. With the release of these upgrades to LTX Trainer has the ability to customize the model itself, fine-tuning it for the specific use case at hand so it matches the unique style and requirements of a project. Through a new flexible conditioning framework, available out of the box, LTX Trainer supports training across:
- Video: Text-to-video (T2V), image-to-video (I2V), video extension (forward and back), inpainting, and outpainting.
- Audio: Text-to-audio (T2A), audio extension (forward and back), and audio inpainting.
- Cross-modal: Audio-to-video (A2V) and video-to-audio (V2A).
- Reference (IC-LoRAs): Video-to-video (V2V), audio-to-audio (A2A), and audio-video-to-audio-video (AV2AV).
Previously, each of these training objectives needed its own custom code, and in practice the trainer shipped only two: T2V and V2V. The upgrade replaces all of that with a single approach driven by one config file. Instead of writing code, users describe what they want for video and audio, whether the model should generate that modality or treat it as a fixed input, plus any conditions to apply, like a first frame, a prefix, or a mask. Those few choices define the training mode, so moving from I2V to video inpainting is a quick edit, not a different program.
The conditions can also be combined and mixed. Several can be combined on one modality, each with its own probability of being applied, so one run can teach more than one behavior at once. For example, applying a first frame to half the clips trains I2V and T2V together.
The use cases are concrete. A studio can train a character LoRA on its own IP and generate that character consistently across an entire campaign. A team can train an audio-to-video LoRA so a soundtrack drives the visuals. A creator can train a LoRA on a recurring character or motion and reuse it shot after shot. Each one becomes a reusable asset, not a one-off generation that has to be recreated from scratch every time.
Simplifying Training with an Agentic Assistant
The expanded set of modes makes the trainer far more capable, but it also widens the space of choices. To make that approachable, we’ve integrated an agentic assistant into LTX Trainer that guides the entire training process from start to finish. The agentic assistant runs directly inside of an agent such as Claude Code, and is built to help users initiate a training run, even for users without prior experience training models.
The experience starts as a conversation. Users describe what they want the model to learn (i.e. a character, a style, a motion, a specific sound effect) and point it at the footage. From there, the assistant will:
- Suggest the right conditioning mode for the specific goal.
- Inspect the user’s data and GPU.
- Caption the clips.
- Generate a training configuration and tune it to fit the user’s hardware.
- Launch the run and monitor its progress.
- Render the finished model against test prompts for review.
Most importantly, none of this is a black box. The agentic assistant pauses for user approval when a decision is needed for mode selection, caption generation, or any compute-intensive actions. Users keep complete ownership of the data, the model, the decisions, and the outputs, and the barrier to training drops from machine learning (ML) expertise to a clear idea and good source material.
Owning the Model, Not Renting It
Much of the AI market is moving toward rented access: closed models, cloud dependency, and output billed by the generation. LTX remains committed to going in the opposite direction for the benefit of our community of more than 10 million users. With LTX Trainer, developers and creators train a model on their own data and own the result. It’s models built on their characters, their styles, their pipelines, that are reusable across projects and deployable on their terms.
There are community-built trainers for other models out there. What is different here is that LTX Trainer is maintained by the team that builds the model, with new, composable conditioning modes, an agentic assistant, and production-ready examples and documentation. It keeps pace with the model instead of lagging behind it.
Get Started
LTX Trainer is available now on GitHub. The documentation, example configs, or the LTX Discord are good places to start. We cannot wait to see what the community builds.
