README.md

---
title: LTX Director Image to Video
emoji: 🎬
colorFrom: indigo
colorTo: purple
sdk: docker
app_port: 7860
pinned: false
license: apache-2.0
---

# LTX Director — Image to Video (ComfyUI)

A Hugging Face **Docker Space** running [ComfyUI](https://github.com/comfyanonymous/ComfyUI)
with the [WhatDreamsCost LTX Director](https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI)
custom nodes and the [ComfyUI-LTXVideo](https://github.com/Lightricks/ComfyUI-LTXVideo)
extension, for **image-to-video** generation with **LTX-2.3 (distilled fp8)**.

## Usage

1. Wait for the Space to start (first cold boot downloads ~37GB of weights — this
   takes several minutes).
2. Open the **`ltx_director_i2v`** workflow from the ComfyUI workflow menu.
3. Upload a source image, enter a prompt, and run.
4. The generated video appears in the UI and can be downloaded.

## Hardware

- **GPU:** L40S or A100 (48GB). LTX-2.3 requires **32GB+ VRAM** and **100GB+ disk**.
  L40S (Ada) is preferred for native fp8 support.

## Notes / caveats

- **Cold starts are slow.** This Space uses no persistent storage, so the LTX-2.3
  checkpoint (~29GB) and the Gemma text encoder (~8GB) are downloaded on every
  cold boot.
- **Public, no auth.** Anyone can run jobs on the GPU. Configure the **idle-sleep
  timeout** in the Space settings to limit cost.
- The default workflow targets the **distilled fp8** checkpoint. The full model and
  IC-LoRA adapters are available in the node packs but are not preloaded.

## Models

- Checkpoint: [`Lightricks/LTX-2.3-fp8`](https://huggingface.co/Lightricks/LTX-2.3-fp8)
  → `ltx-2.3-22b-distilled-fp8.safetensors`
- Text encoder: [`google/gemma-3-12b-it-qat-q4_0-unquantized`](https://huggingface.co/google/gemma-3-12b-it-qat-q4_0-unquantized)

See [`docs/superpowers/specs/2026-06-25-ltx-i2v-hf-space-design.md`](docs/superpowers/specs/2026-06-25-ltx-i2v-hf-space-design.md)
for the full design.