The framework is free and open source, and the IC-LoRAs that ship with it are free to download from HuggingFace. The cost of training your own LoRA on top of LTX-Video is the GPU time, which is hundreds to thousands of dollars per adapter on rented hardware depending on rank and dataset size.

IC-LoRA, in LTX terminology, is in-context LoRA. The LoRA is conditioned on a reference input at inference time, so a single trained adapter can produce different outputs depending on the reference image, video, or audio you feed it. The Ingredients effect in the LTX Trainer launch thread is the clearest demonstration: one reference sheet becomes a fully realized video world.

Do I need to train a LoRA to use LTX Trainer?

No. The ten demo IC-LoRAs (water simulation, inpainting, day-to-night, colorization, and so on) are free to download and use. Train your own only if you have a use case the pre-trained adapters do not cover.

Reviews

Jun 18, 2026

5 min read

LTX Trainer: 10 Fine-Tunable Video Effects, and the First Agent-Driven Training Loop

LTX just shipped a training framework, not a model. Ten demo LoRAs, open source, and the first time a major video lab has shipped a tool an agent can drive end-to-end, including the training step.

By Igor Gridel - Founder, Scopeful

Pricing verified Jun 18, 2026·2 tools live in calculator

Short version

What to remember

01
LTX Trainer is a training framework, not a model: LoRAs and IC-LoRAs for video, audio, cross-modal, and reference-conditioned workflows
02
10 demo LoRAs ship with the release (water simulation, ingredients, inpainting, day-to-night, colorization, instant shave, cross-eyed, deblurring, decompression), all trained on LTX-Video with the new trainer
03
The first major video lab to ship a first-party agentic skill for training, not just prompting
04
Open source: framework, IC-LoRAs, and the agentic skill are all on GitHub and HuggingFace
05
Cost reality: pre-trained IC-LoRAs are free, but training a LoRA on a 19B base is hundreds to thousands of dollars in GPU time
06
Two-layer choice: free IC-LoRAs + rented GPU for custom training, or skip training and pay per generation through a hosted route like fal.ai

LTX just shipped a training framework, not a model. That distinction is the entire story, and most of the coverage is going to miss it.

On June 17, 2026, Lightricks introduced LTX Trainer, a single framework for training LoRAs and IC-LoRAs across video, audio, cross-modal, and reference-conditioned workflows. Fully open source. GitHub, documentation, and HuggingFace all linked from the announcement. Ten demo effects ship with the release, all of them LoRAs the LTX team trained on top of LTX-Video using their own trainer.

The tweet thread is here if you want the source: https://x.com/ltx_io/status/2067258965369528397

The ten effects, straight from LTX's launch thread. Each links to the demo video LTX posted:

Water Simulation. Add rivers, surf, rain, floods, or splashes to any shot while preserving composition and motion. https://x.com/ltx_io/status/2067284780081905685
Ingredients. Combine characters, props, locations, and styles into fully realized video worlds generated from a reference sheet. https://x.com/ltx_io/status/2067284859668877723
Inpainting and Outpainting. Extend and modify scenes beyond the original frame. https://x.com/ltx_io/status/2067284934847528965
Day to Night. Transform daytime footage into nighttime scenes. https://x.com/ltx_io/status/2067285012060537021
Colorization. Bring grayscale footage back to life. https://x.com/ltx_io/status/2067285088761708813
Instant Shave. Remove facial hair while preserving identity and expressions. https://x.com/ltx_io/status/2067285165546873140
Cross-Eyed. Make anyone go cross-eyed while keeping expressions, motion, and framing intact. https://x.com/ltx_io/status/2067285243690930534
Deblurring. Turn blurry footage into sharp, clean video. https://x.com/ltx_io/status/2067285318454378784
Decompression. Remove compression artifacts and restore visual quality. https://x.com/ltx_io/status/2067285396728521147

The tenth tweet is the closer, not a new effect: "All trained with LTX Trainer. What will you build?" https://x.com/ltx_io/status/2067285408468337046

(All nine demo videos are LTX's own assets, linked back to the original tweets. If a video breaks, the tweet is the source of truth.)

What actually changed

Read the announcement the way you would read a release note, not a press release, and three things stand out.

First, the trainer is a framework, not a single skill. It supports video, audio, cross-modal, and reference-conditioned workflows in one place. That means a LoRA you train can be conditioned on a reference image, a reference video, an audio track, or a combination. Most of the existing video model ecosystem is still one-input, one-output. Multi-modal conditioning at the training stage is a real shift.

Second, the open-source IC-LoRAs ship with the framework. IC-LoRA, in LTX's vocabulary, is in-context LoRA: the LoRA is conditioned on a reference input at inference time, so a single trained adapter can produce different outputs depending on the reference you feed it. The Ingredients effect (number 2 above) is the cleanest demonstration of the pattern. The LoRA turns a single reference sheet of characters, props, and locations into fully realized video worlds.

Third, and the one most people will gloss over: a new agentic skill ships alongside the trainer. The line is buried in the announcement copy, but it is there. "Plus: new agentic skill, flexible conditioning, free IC-LoRAs, fully open source." The agentic skill is what lets an AI agent, not a human in a CLI, drive the training loop. Pick the base model, point it at a dataset, configure conditioning, kick off training, evaluate the output. That is the same shape as Scopeful's own MCP server surface, and it is the first time I have seen a major video lab ship a first-party agent surface for training rather than only for prompting.

Why this matters if you are not a researcher

The reason I am writing this on Scopeful, where the beat is pricing and tooling, is that the gap between "an open-source trainer exists" and "you can actually use it" is enormous, and LTX is one of the few teams actually closing it.

Most open-source video training setups require you to wire together the base model, the training loop, the conditioning pipeline, and the inference harness yourself. LTX Trainer ships all four as one project. The free IC-LoRAs are reference adapters you can download from HuggingFace, run locally, and use without retraining. The agentic skill exists so an agent, Claude, GPT, whatever you run, can use the trainer the same way it uses any other tool.

That is the part with real Scopeful relevance. We have been saying for a year that the agent layer is the most underexplored surface in creative AI. Most vendors are still shipping prompt UIs. LTX just shipped a tool that an agent can drive end-to-end, including the training step. If the trend holds, every video model lab is going to ship an agent surface within twelve months, and the ones that ship it first will define what "agent-friendly" means in the category.

The cost reality

The training framework itself is free and open source. The compute is not.

A LoRA fine-tune on a 13B-to-19B parameter video model is not a laptop job. You are looking at a multi-GPU box or a cloud rental, and depending on the dataset size and target rank, the bill lands somewhere between a few hundred and a few thousand dollars per adapter. The 10 LTX-shipped IC-LoRAs save you that cost, because you can download and use them directly. The 11th adapter, the one you train on your own footage, is where the cost lives.

If you do not want to train anything, the underlying LTX-Video model is also billable through hosted inference, including via fal.ai, which is what the Scopeful calculator already prices out. The LTX-2 family and the LTX 2.3 family are both in the catalog with per-second pricing. So you have a clean two-layer choice: use a pre-trained IC-LoRA from HuggingFace for free, train your own on rented GPUs, or skip training entirely and pay per generation through a hosted route.

What I would actually do

If you are curious, the move is to grab the free IC-LoRAs from HuggingFace and try the Ingredients reference-sheet workflow first. It is the most generally useful of the ten, and the demos are convincing. Pair it with an LTX-Video inference route, either the open-source weights locally or a hosted route, and you can do reference-conditioned video generation today, on a laptop, for the cost of inference only.

If you have a use case the free adapters do not cover, train your own. The trainer is the right primitive for that. Budget realistically: a single LoRA on a 19B base is a meaningful spend. Treat it like hiring, not like a subscription.

And if you are building an agent that touches video models at all, watch LTX's agentic skill carefully. It is the first time a major video lab has shipped a tool designed to be driven by an agent, not by a prompt. That is the move the rest of the category is going to make, and the projects that adopt the pattern early are the ones that will set the standard.

Questions & Answers

Is LTX Trainer free?: The framework is free and open source, and the IC-LoRAs that ship with it are free to download from HuggingFace. The cost of training your own LoRA on top of LTX-Video is the GPU time, which is hundreds to thousands of dollars per adapter on rented hardware depending on rank and dataset size.
What is IC-LoRA?: IC-LoRA, in LTX terminology, is in-context LoRA. The LoRA is conditioned on a reference input at inference time, so a single trained adapter can produce different outputs depending on the reference image, video, or audio you feed it. The Ingredients effect in the LTX Trainer launch thread is the clearest demonstration: one reference sheet becomes a fully realized video world.
Do I need to train a LoRA to use LTX Trainer?: No. The ten demo IC-LoRAs (water simulation, inpainting, day-to-night, colorization, and so on) are free to download and use. Train your own only if you have a use case the pre-trained adapters do not cover.
What is the new agentic skill?: A first-party agent surface that lets an AI agent, not a human in a CLI, drive the LTX Trainer training loop end-to-end: pick the base model, point at a dataset, configure conditioning, kick off training, evaluate the output. It is the first time a major video lab has shipped this kind of surface for training rather than only for prompting.