---
tags:
- ltx-video
- text-to-video
- image-to-video
pinned: true
language:
- en
license: other
---

This is a special fork of LTX-Video designed to facilitate deployment to the Inference Endpoints, using diffusers + varnish.

If you want to know how to use it once you've deployed it, please refer to this Python demo:

https://huggingface.co/jbilcke-hf/LTX-Video-for-InferenceEndpoints/blob/main/example.py

## Why?

Using the Hugging Face Inference Endpoints gives you with reliable and controllable availability.

The goal of this wrapper is to offer flexibility and low-level settings, to add things such as upscaling, interpolation, film grain, and compression controls.

## Setup on Hugging Face Inference Endpoints

### Pick a large machine

It is recommended to use at least a NVIDIA L40S with 48 Gb of VRAM.

### Downloads all assets

Make sure to select "Download everything" when selecting the model.

Otherwise some files ending in `.pth` won't be downloaded.

### Select Text-to-Video or Image-to-Video

By default the handler will do Text-to-Video.

To do Image-Text-To-Video, you need to set the environment variable `SUPPORT_INPUT_IMAGE_PROMPT` to a trueish value (eg `1`, `True`).

It is possible to support both pipelines at the same time if you modify the `handler.py`.

But if you keep both pipelines active in parallel, this will consume more memory.

### Using private LTX-Video LoRAs

If you plan on using private LoRA models you will have to set the `HF_API_TOKEN` environment variable.

## Credits

For more information about this model, please see the [original HF repository here](https://huggingface.co/Lightricks/LTX-Video/).