For two years the deal with AI video has been simple and a little insulting: you get whatever the model already knows. Want your character, your look, your motion? Tough. On June 17, 2026, Lightricks quietly changed that — they rebuilt the LTX Trainer so that thirteen different ways to fine-tune a video model now live behind a single config file. And then they did something nobody else has: they shipped a skill so Claude Code can run the whole thing for you.
The Story
LTX-2 is Lightricks’ open-weight, audio-and-video diffusion model — the one we covered back when LTX-2.3 dropped 4K with synced sound in a single pass. Open weights are great, but the real unlock for creators was never inference. It was training. A base model gives everyone the same vocabulary; a LoRA gives you a private dialect.
The problem, until now, was that training one was a coding project. Every objective — text-to-video, image-to-video, audio-to-video, video extension, inpainting, outpainting — was its own script, its own glue, its own way to break at 3am. The June update collapses all of that. The unified LTX Trainer exposes 13 training modes — T2V, I2V, A2V, V2V, video and audio extension, inpainting, outpainting, plus IC-LoRA for video, audio and joint audio-video references — and you select between them by editing one YAML file. No custom code per task. Want a character LoRA instead of a colorization LoRA? Change a few lines.
The headline trick, though, is the agent. Lightricks dropped a train-model repository skill — a SKILL.md file — into the repo. Point an agent like Claude Code at your footage and it probes your data and your hardware, picks the matching training mode, writes the config, kicks off the run, and babysits the job. The thing standing between “I have 30 clips of my character” and “I have a working LoRA” stops being Python knowledge and becomes a sentence.
Why You Should Care
This is the part that matters for anyone building a real pipeline rather than posting one-off clips. A LoRA adapter trains only a sliver of the model’s parameters while the base stays frozen, so the file it spits out is small — typically a few hundred megabytes — portable, and stackable. Your character adapter plus your studio’s look-dev adapter plus a motion adapter, layered at inference. That’s the difference between “cool demo” and “house style.”
Lightricks’ own example LoRAs make the point concrete: locking a single character’s identity across shots, a day-to-night converter, and a grayscale-to-color restorer. Those aren’t toys — they’re exactly the consistency and continuity problems that have kept AI video out of serious production. And because LTX-2 is an audio-video model, the trainer reaches into sound too: audio-to-video, video-to-audio, audio inpainting. You can fine-tune what your world sounds like, not just how it looks.
Try It / Follow Them
Let’s be honest about the gate: comfortable training wants an NVIDIA GPU with 80GB+ of VRAM, Linux, and CUDA 13+. That’s a cloud rental for most of us. The good news — there’s a low-VRAM config using INT8 quantization that runs on a 32GB RTX 5090, which puts character LoRAs inside reach of a high-end home rig for the first time.
- The repo: github.com/Lightricks/LTX-2 — the monorepo holds
ltx-core,ltx-pipelinesandltx-trainer. - The trainer docs: ltx-trainer README — start here for the YAML modes and the dataset workflow.
- The weights: Lightricks/LTX-2 on Hugging Face.
- The recipe: Lightricks’ step-by-step LoRA guide — organize scenes, auto-caption, cache latents, run.
One gotcha worth tattooing on your monitor: resolution buckets must use spatial dimensions that are multiples of 32, and frame counts that satisfy (F-1) % 8 == 0. Get that wrong and the run dies before it starts.
IK3D Lab Take
We keep saying the interesting frontier in AI creative tooling isn’t bigger models — it’s control. LTX Trainer is a near-perfect specimen. The 13-modes-in-one-config move is solid engineering, but the agent skill is the genuinely new idea: the open-weight model, the trainer, and a coding agent finally form a closed loop where the artist’s job is to curate the dataset and describe the goal, and the machine handles the plumbing. That’s the right division of labor.
The 80GB requirement keeps this out of reach of the casual crowd for now, and we’d love to see the 32GB path mature into something a creator can run overnight without a CUDA degree. But the direction is exactly right. When training your own video model stops being a research task and becomes a YAML file an agent fills in, the question stops being “what can the model do” and becomes “what do you want it to be.” That’s the version of AI video worth getting excited about.
