v1.2.69 - LTX2.5: Generation & Training

LTX-2.5, running on your own GPU, Train Clip & IC LoRa

Lightricks' 22B open-weights video model is now a local model in Inline Studio.
It generates video and its soundtrack in the same pass, on your own hardware,
with no account and no per-second billing.

• Three nodes. Text to video, image to video, and first-and-last frame. Each
  has a fast mode (distilled, 12 steps) and a quality mode (the dev build, guided),
  and each takes a LoRA. Clips run from 1 to 20 seconds at 24fps.

• Bring the weights. Drop the checkpoint pack into models/ the way you would
  any other model. The download is about 71GB for fast mode, and the app tells you
  what is missing and fetches it for you if you would rather not do it by hand.

• What it needs. An Ampere card or newer. A 2 second clip at 960x576 renders in
  about 229 seconds on a 48GB card once the model is warm. On a card too small to
  hold the transformer it streams the weights from system RAM instead, which works
  but takes roughly four times as long.

Train a LoRA on it(Clip & IC Lora)

Two kinds, both trained on the dev transformer, both teaching the picture rather
than the sound.

Clip LoRA learns a look and how it moves, from single clips. Feed it a folder
of footage and the adapter carries that style into anything you generate
afterwards.

Control LoRA learns a transform rather than a look. Every item is two clips:
a reference and the result you want from it. Lightricks call this an IC-LoRA.
Training is available now and produces a standard adapter file. Driving a render
with a reference clip is not wired up yet, so this release covers training a
Control LoRA rather than generating with one.

Training peaks at 42GB, so a 48GB card is the floor rather than a comfort. There
is no 4-bit path for this architecture, and the Trainer says so plainly instead
of accepting the setting and ignoring it. A step takes about two thirds of a
second at 512px, so the startup costs more than the training on a short run.

A new way to add training data

The dataset controls are now one dialog, Add/Manage Training Data, which both
builds a dataset and edits one you already have.

• Pull from Hugging Face. Paste a repo id and see the item count, how many are
  paired, and the download size before anything is fetched.
• Import a folder from the machine running Inline Studio, without pushing
  gigabytes through the browser.
• Pairs are found for you. A dataset.json or metadata.jsonl is read where
  the dataset ships one, and filenames are matched otherwise, so bear.mp4 and
  bear_reference.mp4 become a single training item rather than two.
• Captions come from the dataset where it has them, and the local captioner
  fills only the rows that have none rather than overwriting good text.
• Nothing reaches your project until you press Import. Checking a repo or
  loading a folder cannot change a dataset by accident, and Import stays greyed
  out until something has actually changed.

Smaller things

• LTX-2.5 in the starter cards, matched against the GPU you actually have.
• Large datasets no longer stall the browser. Clip previews load as you scroll
  rather than all at once, play on hover, and open full screen on a double click.
  A paired item opens both halves side by side, which is the only way to see
  whether a reference and its target line up.
• A guide at
  inlinestudio.art/lora-training/ltx-2-5-lora-training-locally,
  with measured VRAM, an example dataset, and the licence terms.

Before you start

The weights are gated. Open the model page on Hugging Face, accept the LTX-2
Community License, and make sure the account you accepted with is the one your
token belongs to. Without that, every download comes back as a permission error
rather than a file.

Upgrading

From the repo (the usual way):

cd Inline-Studio
git pull
cd core
./webui.sh --install --extra all
./webui.sh

What's Changed

• LTX-2.5: local video generation, Clip and Control LoRA training, and the dataset editor for it by @imprsnst in #47

Full Changelog: v1.2.68...v1.2.69