---
title: LTX-2.3 LoRA Trainer
emoji: 🎬
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 6.18.0
app_file: app.py
pinned: true
hardware: cpu-basic
hf_oauth: true
hf_oauth_scopes:
- read-repos
- write-repos
- manage-repos
- jobs
---

# LTX-2.3 LoRA Trainer (HF Jobs)

Train a **LoRA / IC-LoRA for LTX-2.3** from your own videos, entirely on Hugging Face
infrastructure:

- **Sign in with Hugging Face** — the dataset, the job, and the pushed LoRA all run under
  **your** account and billing (no pasted tokens);
- the **Space** (this app, `cpu-basic`) collects your videos + hyperparameters and submits a job;
- training runs on **HF Jobs**, reproducing the trainer environment from the monorepo
  lockfile (`uv sync --frozen`);
- your dataset is staged on **HF buckets** (`hf://buckets/<you>/ltx2-train-<run>`);
- the trained LoRA is pushed to the Hub model repo you choose.

You only pay for the Job's actual GPU runtime.

## Setup

Just **sign in with Hugging Face** in the app — the OAuth token (scopes: repos + `jobs`) is
used to create buckets, submit the job, download the gated Gemma encoder, and push the LoRA,
all as the signed-in user. No Space secret is required.

The job pulls the trainer source from the bucket **`LTX_SRC_BUCKET`**
(default `linoyts/ltx2-trainer-src`). To use your own, upload the monorepo
(`pyproject.toml`, `uv.lock`, `packages/`) to a bucket and set the `LTX_SRC_BUCKET`
Space variable.

## Usage

1. Sign in with Hugging Face.
2. Pick a mode and upload your videos:
   - **IC-LoRA (in-context control):** upload each target `clip.mp4` **and** its control video
     `clip_reference.mp4` (depth, pose, edges, an inpainting-masked version, …) — matched by
     filename. References are user-supplied; nothing is auto-derived.
   - **T2V / I2V:** upload only the target clips.
3. Add one caption per line (in filename order of the target clips), set hyperparameters and a
   Hub model id, then **Submit training job**.
4. Copy the job id into the monitor box and **Refresh** to stream logs.

The job: sync buckets → `uv sync` → download base checkpoint + Gemma → `process_dataset.py`
→ `train.py` → push LoRA to the Hub.

## Notes

- Resolution: W,H divisible by 32; frames `F % 8 == 1` (25, 49, 81, …).
- `a100-large` (80 GB) fits the 22B model in bf16; INT8 quantization is optional.
- First job is slow to start (env build + ~67 GB of model downloads) before training begins.