LTX 2.3 ComfyUI Tutorial: Text to Video & Image to Video

What is LTX 2.3?

LTX 2.3 is a free, open-source video generation model built by Lightricks. At 22B parameters, it handles text-to-video, image-to-video, and native 9:16 portrait output without any cropping — and it runs locally through ComfyUI with no subscription required.

It's roughly 18× faster throughput than WAN 2.2 14B, and audio-to-video is built directly into the generation process — synced during generation, not tacked on afterwards. ComfyUI is the best way to run it locally: full control over every setting, access to the official workflow, and zero dependency on third-party services.

See the results before you download anything

All three model versions produce solid output. The differences mostly show up at the limits of your hardware. Here's the same prompt run through each version — watch before you commit to a 23–39GB download.

Not sure which to pick? If you're on an RTX 4070 or better, download the FP8 version and stop reading this column. Only go BF16 if you need absolute maximum quality and have 24GB+ VRAM to spare.

What you need before you start

Two things break most installs before they begin: the wrong ComfyUI version, and not enough VRAM for the chosen model. Check both before downloading anything.

ComfyUI version — check you're on v0.22 or later

LTX 2.3 nodes are built into ComfyUI natively from v0.22 onwards — no custom nodes needed for BF16 and FP8. The version number is shown in the bottom-left corner of the interface.

Which model should you download?

The v1.1 distilled models are newer and better than the original LTX 2.3 release. Use one of these, not the original dev weights, unless you specifically need dev-quality for a final render.

Running on less than 8GB VRAM?

If your GPU is under 8GB, RunPod lets you rent an A100 for under €1 per hour and run LTX 2.3 at full quality without touching your local machine.

Download LTX 2.3 — the right files and folders

Most failed installs come down to a missing file or a file in the wrong folder. Start by downloading the workflow file below, then follow the model download instructions — this way you have the workflow ready and the downloads listed in order.

Step 2 — Download the model files

LTX 2.3 needs more than just the main model — here's every file and exactly where it goes. Use the download buttons to get each file directly from HuggingFace.

Step 3 — Put each file in the right folder

Two things catch people out: (1) GGUF models go in checkpoints/, not diffusion_models/ — wrong folder means ComfyUI won't see it. (2) Both text encoder files are required — missing either one causes black frames.

How to load LTX 2.3 in ComfyUI

You've already downloaded the workflow JSON above. Here's how to open it in ComfyUI.

Load the workflow — step by step

1. Open ComfyUI in your browser.
2. Drag and drop the downloaded JSON file directly onto the ComfyUI canvas — it will load automatically.
3. Alternatively, click the Load button in the ComfyUI menu bar and select the JSON file.
4. The workflow opens with all nodes connected. If you downloaded the main workflow (BF16 + FP8), you'll see a model selector node — point it to your downloaded model file.
5. To switch between BF16 and FP8, change the model file in the checkpoint loader node inside the workflow.
6. To switch between text-to-video and image-to-video, use the T2V / I2V toggle button inside the workflow — no need to load a different file.

Red nodes or errors on first load?

Two causes cover most cases. First, check your ComfyUI version — you need v0.22+. Second, confirm both text encoder files are in models/text_encoders/. A single missing file produces either red nodes or silent generation failures.

Speed up generation with SageAttention

The workflow uses SageAttention — an optimized attention implementation that can speed up generation by 30–50%. If you haven't installed it, generation will still work but will be slower. Install SageAttention once and all your ComfyUI workflows benefit.

Text to Video (T2V) with LTX 2.3

Switching between BF16, FP8, and T2V/I2V in the workflow

The main workflow covers everything — you don't need separate files for each mode. Inside the workflow, look for the model checkpoint loader to switch between BF16 and FP8. The T2V / I2V toggle is a button near the top of the workflow canvas.

Add your workflow screenshot here — shows the checkpoint loader (to switch BF16/FP8) and the T2V/I2V toggle button at the top of the canvas.

Add a close-up screenshot of the T2V / I2V toggle button — so users know exactly where to click to switch modes.

Settings that work — what to set and why

These are reliable starting values. Don't change anything until you've confirmed a basic generation works.

How to write prompts for LTX 2.3

LTX 2.3 responds to cinematic, scene-specific language. Vague prompts produce vague motion. The more specific you are about scene, lighting, and camera behaviour, the more consistent the result.

Want the full breakdown of prompt structure, camera language, and mode-specific tips? Read the complete LTX 2.3 prompting guide →

Download the T2V / I2V workflow

The main workflow covers both T2V and I2V — switch modes with the toggle button inside. Download the version that matches your hardware.

Image to Video (I2V) with LTX 2.3

I2V works differently from T2V. You provide a still image and the model generates motion from it. Your prompt describes how things move — not what the scene looks like, because the scene is already your image. Use the I2V toggle button inside the workflow to switch to image-to-video mode.

📸 I2V workflow — where to add your image and switch modes

screenshot here showing the I2V toggle switched on and the image loader node — so users can see exactly where to load their input image.

Example of the input image you used — helps users understand what makes a good I2V source image.

What makes a good input image

LTX 2.3's I2V is notably better than LTX 2 — sharper edge preservation, more consistent subjects across frames. For best results:

• Clear subject with no motion blur — the model needs a clean frame to start from.
• Good, even lighting — extreme contrast or heavy shadows create flickering across frames.
• Avoid overly busy backgrounds if you want the subject to stay consistent.

Settings and prompting for I2V

Settings are similar to T2V. The key difference is denoise strength — start at 0.85 for subtle motion. Lower it for more movement from the source image.

The prompt describes motion, not the scene. Here's an example using a portrait photo as the input:

Download the I2V workflow

Use the main workflow (BF16 + FP8) and toggle to I2V mode, or use the dedicated GGUF I2V workflow if you're on a lower-VRAM card.

Frequently asked questions