LTX-LoRAs

LTX-LoRAs

Adds LoRA files for the LTX 2.3 video model that fill masked areas of a video, with or without a visual reference, plus an experimental video-editing LoRA. Made for video creators who work with LTX 2.3.

Made by Alissonerdx on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository holds LoRA files trained for LTX 2.3, a video generation model from Lightricks. Most of them do inpainting, which means replacing a masked area of a video. One version also accepts a visual reference, and an experimental Edit Anything LoRA handles add, remove, replace and style changes.

Who it suits: It suits video creators and tinkerers who already run LTX 2.3 and want to edit part of a clip, or who want to test video editing and build synthetic datasets.

Industry | Who uses it | How they use it | Example
Video production | Video editors | Replace an object in part of a clip | Swap a masked item for a new one
Visual effects | Compositors | Guide a replacement with a reference image | Use a reference to steer the new object
AI research | Dataset builders | Make synthetic edit pairs, especially style-focused ones | Run the Edit Anything LoRA on test clips
Hobby creators | Experimenters | Test prompts and edit behavior | Compare the 2500-step and 10000-step files

If you do not run LTX 2.3, or you need polished production output, you can skip it.

Where to use it: These files load into an LTX 2.3 workflow that supports LoRAs and the guide nodes the card names.

• A workflow that uses the LTXVAddGuideMulti node
• The Blockify Mask node from KJNodes
• Windows, macOS or Linux, for downloading the files

Use case | Who | What they get back
Video inpainting | Editors | The masked region of a video filled according to your prompt
Reference-guided inpainting | Effects artists | A replacement shaped by a visual reference
Experimental editing | Researchers | Add, remove, replace and style edits for testing

How to set it up: You need the LTX 2.3 base model and a workflow that can load LoRAs. Prepare the files and a guide video with the mask built in.

1. Install the Hugging Face hub tools.
2. Download the repository files.
3. Pick a checkpoint from the card's file list.
4. Load it in your LTX 2.3 workflow, with the mask embedded in the guide video.

pip install -U huggingface_hub
hf download Alissonerdx/LTX-LoRAs

How to use it: Keep the mask out of its own channel. Combine the mask video and the guide video into one video, then feed it in through the LTXVAddGuideMulti node. The colors matter. Paint the mask magenta, and color the reference's green zone with chroma key green.

Tool | What it does
LTXVAddGuideMulti | Passes the guide video, with the mask inside it, to the model
Blockify Mask | Makes the mask blocky like the training data, with size 8 as the default
hf download | Fetches the LoRA files to your computer

Example: You have a clip where one object should change. You paint that area magenta in the guide video, write a prompt that describes the replacement, and load the reference-based inpainting file with a reference image. The model fills the masked region as the prompt and reference direct.

For newcomers and experts: A LoRA is a small add-on file that gives a big model a new skill. A first try is the 2500-step inpainting file with a simple mask and a clear prompt.

The inpainting files learned from one fixed guide and mask layout, using 8 by 8 blocks, so inference should copy that layout. The 2500-step checkpoint follows prompts better, while the 10000-step one uses the masked area better but may overfit. You can expand the mask up to 512, which makes it a full rectangle.

Download the LoRA files
Prepare a guide video with a magenta mask
Pass it in with the guide node
Write a clear prompt
Compare checkpoints on your clip

Good to know: Results depend a lot on the checkpoint, the prompt and the mask, so the card suggests trying more than one.

• Mask colors: magenta for the mask and chroma key green for the reference area.
• Prompting: for the reference-based file, prompt quality and mask size matter a lot.
• Edit Anything: experimental, trained on 8,000 video pairs, and may change as training goes on.
• Quality: the Edit Anything file is for testing, not polished professional output.
• License: Apache 2.0.
• Cost to run: none named. It needs LTX 2.3 and your own hardware.
• Data it reaches: the videos, masks and reference images you provide.
• Not for: professional-grade output from the Edit Anything file.

On Hugging Face, 13 November 2026

• Library: ltx
• Downloads in the last 30 days: 8,461
• Likes: 231
• License: apache-2.0
• Based on: Lightricks/LTX-2.3
• Tags: lora, inpaint

Links

• Model: https://huggingface.co/Alissonerdx/LTX-LoRAs
• Files: https://huggingface.co/Alissonerdx/LTX-LoRAs/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Alissonerdx/LTX-LoRAs

Then follow the usage in its model card. Nothing in this kit runs by itself.