Despite recent advances in creative tech, AI video generators still struggle with consistency. You can run the same prompt 100 times and you’ll likely get 100 slightly different clips. If you’re a filmmaker or studio trying to get AI content to feature recognizable characters or mimic an iconic design style, you can’t have AI making changes. That’s why AI company LTX built a framework, named LTX Trainer, to help you train models in your own style.
Instead of having an off-the-shelf AI model that replicates all kinds of art styles, you can use LTX Trainer to create your own AI model that is only trained on your stuff and creates content in your unique style. These custom, fine-tuned models can be built without extensive coding, thanks to new updates rolling out today on GitHub, LTX co-founder Yaron Inger exclusively told CNET this week.
“It was very important for us to release this, so the community can fine-tune their own data,” Inger said in an interview. Creative companies that want to experiment with AI but need to protect their intellectual property can download the company’s publicly available tools and run models locally, he said, avoiding the issues that come with uploading their IP to outside networks or cloud servers.
While AI video hasn’t really taken off for the general public beyond oodles of AI slop — RIP, Sora — it’s a booming business for professional creators and entertainment studios. Studio heads have lauded AI for its ability to speed up workflows, while some creators enjoy experimenting with it. For professionals, being able to customize an AI model opens new creative avenues beyond just using mass-market slop-generators that produce content that often all looks the same.
But creative AI is controversial, with many artists fearing that AI uses their work without permission and can unethically replicate their hard-built styles and brands. That’s why companies such as LTX and Adobe, which introduced a way for businesses to create custom AI models earlier this year, say they don’t train on their customers’ content. LTX’s open models and trainer, when run locally, also keep your data in your hands.
Trainer can now do audio-only and joint audio-video training, expanding from its previous video-only abilities. You can prompt between modalities, such as audio-to-video or image-to-video.
Creators can also customize their AI content to match their unique style using adapters called LoRAs and IC-LoRAs. Essentially, these are your custom, fine-tuned models that rely on LTX’s underlying world AI model to consistently replicate your style across generations.
This process lets people customize AI models without spending months and thousands of dollars, as companies such as Google and Meta have the resources to do. A new agentic assistant makes it easier to build, so you can describe in natural language what you want to create.
You can also apply specific editing conditions to your custom models. For example, you can create a model that denoises every audio file it generates or upscales video clips to a higher resolution.
At the beginning of the year, LTX dropped its second-gen model, which it built in partnership with Nvidia. LTX-2 is different from other AI video models. For one, it’s built to run on-device; AI video generation is a very compute-intensive process and uses way more energy than chatbots, which is why it’s usually run in the cloud.
LTX-2 is also an open-weight model, which gives us some insight into how it was constructed and functions. (Truly open-source AI would also share the content the model was trained on, experts say, which no major AI company has wanted to disclose.)
The company’s end goal with this update, Inger said, is to make a tool where you feel your intellectual property is secure and you can experiment with AI.
“These kinds of tools basically allow you as a creator to customize things exactly the way that you like, while having all the upsides of AI models that can generate things very, very quickly,” Inger said. “But you can then keep the IP; it’s yours.”
Read the full article here