Overview
The LTX Trainer exposes the same toolkit we use to train production LTX models. Use it to fine-tune LTX on your own data,
from a quick style LoRA to full multimodal fine-tuning. The train-model agent makes training approachable even without ML
expertise, while experts keep full manual control over every parameter. It’s built around a single conditioning
framework that covers every training scenario.
Start with the agent. Describe what you want and /train-model trains it for you — run it in Claude Code, or
follow the Quick Start to train by hand.
What You Can Train
One strategy, every mode. A single flexible conditioning framework expresses 13+ training modes — switch between them by editing a few lines of training config (set which modality is generated and compose conditions), rather than picking a separate strategy or writing code. Text- and image-to-video, video and audio extension, inpainting and outpainting, audio-to-video, video-to-audio (Foley), text-to-audio, and in-context (IC-LoRA) transformations are all driven by the same config block.
Joint audio + video. Train both modalities together through the model’s cross-modal attention, or freeze one to condition the other (e.g. generate Foley from a fixed video, or video from fixed audio).
In-context control (IC-LoRA). Learn transformations from paired videos or audio — depth and pose control, style transfer, deblurring, colorization, and more.
LoRA or full fine-tuning. Train lightweight, portable LoRA adapters, or update all model parameters with distributed FSDP for larger adaptations.
Multiple model versions, one API. The trainer supports LTX-2, LTX-2.3, and LTX 2.5 through the same configuration
schema — the architecture is detected automatically from the checkpoint metadata, so there’s no model-version flag to
set. Just point model_path and the matching text_encoder_path at the version you’re training.
Runs on accessible hardware. 80GB VRAM is recommended, but a low-VRAM path (INT8 quantization, 8-bit optimizer, reduced rank) brings LoRA training to 32GB consumer GPUs such as the RTX 5090.
Across model versions, the large majority of LoRAs and IC-LoRAs trained on LTX-2.3 run on LTX-2.5 without changes; a small number of exceptions exist, so validate an adapter previously trained before production use.