Skip to main content
LTX-2.5 is Lightricks’ 22B open-weights video model, and it generates a soundtrack with the picture. Omnichar Studio trains LoRAs for it on video clips, in two modes:
  • A Clip LoRA learns look, style and motion from single clips.
  • A Control LoRA learns a transform from paired reference and target clips: edges to video, depth to video, one look to another. Lightricks call this an IC-LoRA.
Both train the picture, not the sound. The training forward pass runs the video branch only, so the soundtrack stays with the base model.

What you need

  • A 48GB card. Training peaks at 42GB on an L40S, and the base alone is 38GB once loaded. There is no 4-bit path, and a 24GB card can’t train it at any resolution. Generation is far more forgiving, because LTX streams its own weights.
  • Ampere or newer. A T4 isn’t supported.
  • The dev transformer: ltx-2.5-22b-dev-transformer-bf16.safetensors (42GB) in models/diffusion_models/. It’s the only base the trainer accepts. Lightricks publish it beside the distilled build specifically to be trained.
  • The licence. The weights are gated under the LTX-2 Community License. Accept it on the Hugging Face model page with the account your token belongs to, or every download returns a permission error.
The adapter you train loads on all three LTX nodes, in both Fast and Quality mode, because the distilled and dev builds share an architecture.

Clip length

The video VAE encodes clips on an 8n + 1 frame grid at 24fps, so the trainer rounds your Clip length down onto that grid. The floor is 9 frames. The panel shows the resolved number next to the field before you start. Each clip is trimmed from its start, once, at precache time.

Train a Control LoRA

1

Switch the mode

In Adjust, set Training mode to Control LoRA.
2

Pair every item

Each dataset item needs a reference clip as well as its target. Reference and target must have the same frame count. Both halves are trimmed from the same end, so a pair of different lengths trains misaligned instead of being rejected.
3

Train

Press Start. Both modes adapt the same layers, the video attention projections plus the feed-forward layers, so only the dataset and the forward pass change.
The dataset editor in Control mode: each row pairs a clip with its edge-map reference and a prompt, with 90 of 90 paired The fastest way to see one work is Lightricks’ Canny-Control-Dataset: 90 clips, each paired with an edge map of itself. Paste the repo id into the Hugging Face tab and the Trainer reads the pairing itself. It also reads the clip.mp4 plus clip_reference.mp4 naming. To build pairs from clips you already have, Lightricks’ dataset preparation guide includes compute_reference.py, which derives Canny edges and can be swapped for depth, pose or anything else.

Speed

About 0.67s a step at 512px on an L40S. Setup is the expensive part: encoding 25 short clips and loading the base takes around 13 minutes. A 300-step run spends 3 minutes training and 13 getting ready; a 3000-step run spends 34 minutes training against the same 13. The encode is cached and reused for the same dataset, resolution and clip length, so prefer long runs.