Video - https://youtu.be/IOJ0jA-9Ccc

Overview

Some new LTX 2.3 models are here and they boast both better audio quality (much needed!) and different visuals - https://huggingface.co/Lightricks/LTX-2.3/tree/main

Because my LTX workflow isn't large enough yet, I figured I'd also add another couple of groups - this time for extending videos :)

Workflow Information

For V1.1 you will need to download some new models, of which there are many options. The notes in the workflow have the various updated links.

It does all the things the previous workflow versions can do - only now with the addition of video extending!

This means it still has a great many features, including:

VRAM Saver

Text to video

Image to Video

Custom Audio

Union Control Net

LoRAs

First Image / Last Image

Best Face Swap

Note: Fixed an issue with the -ve prompt for sampler 1 in version 44a! Was using +ve prompt twice. Doh! Not much of an issue when using distilled models, of course... but should now be better for non-distilled!

Custom Nodes

You will need to install a few custom nodes in order to use all of the features, as detailed in the workflow note:

Ollama can be installed from https://ollama.com

Switches - Your easy way to control group operations!

Simply flip any of the available switches to quickly and easily enable / disable the specified option. For example, Start Image Only mode is off by default as the workflow is configured for basic text-to-video operations:

If you want to use a starting image with text-to-video, then you can enable Start Image Only mode by flipping the switch:

If you instead want to use the first image with ControlNet, again you would simply flip the switch to disable the image bypass:

Groups also come with notes containing additional information:

There are many switches, typically in black:

You may wish to check previous posts and videos for even more information!

Dragons!

You don't need to look at the "engine" outside the map, but exploring it can also be fun if you're looking to see how things work, or what order options are chosen in:

There are many different options, and the switches will try their best to override mistakes in options, using the latents and conditioning in the order shown:

Note that the workflow isn't magic though, thus you will be able to select options which make no sense... such as trying to use Best Face Swap without loading a Best Face Swap video! Be sure to check the notes in each section for more information.

Bookmarks

Bookmarks are also available to quickly jump to the most often used workflow groups:

Just press the relevant number key to jump straight to that section of the workflow.

Model Download Options

There are a great many different options to choose from, but I'd suggest starting with Kijai's FP8 conversions.

The official LTX v1.1 options are currently only for the full model...

... with their FP8 options still only being for the previous version:

Kijai has a couple of distilled FP8 options:

So far as I can tell, there doesn't seem to be much difference in quality, and for me at least I think the slightly smaller "mxfp8" version works just fine. Give them a go and see which ones you prefer.

Also note, that whilst the Distilled LoRA has a new file name (ltx-2.3-22b-distilled-lora-384-1.1.safetensors), the new Union LoRA has the same filename and so will overwrite your existing one (unless you decide to change the name).

Choice of Model Loader

There are 3 different options for loading, and chances are you'll only want one of them... unless you're like me and want to try everything out.

The Loader Group defaults are for Kijai's new v1.1 files:

If you want to use the checkpoint files, then just connect that group to the Init block instead:

Likewise with the GGUF loader. You can, of course, simply delete any model loader group(s) you don't want to use.

Extending Videos

The two new groups are for extending your initially generated video, or any existing one:

Initial video end:

Extended video end:

There is a second prompt group for extending a video:

You can pick where to extend the video from using the LTXVAudioVideoMask node:

You don't have to start exactly at the end of the initial video, and can start at any point.

You can also extend an existing video (not just the initial one you generated). Simply flip the switch to enable the load Video to Extend group:

The Video to Extend group is just underneath:

If you are simply extending a video, you can turn off the usual generation using the switches in that group:

With output, stage 1 & stage 2 and Final Decode groups off, only the video extend groups will run.

Extending an existing video works pretty well, and often the voice sounds very similar to the original - though prompting is king here!

If extending a video, it is just a 1-stage sampler pass, and the Load Video node itself is used to control video resolution and initial frame count, unless you're extending a video you just generated.

Being a 1-stage sampler, this does mean it will take a bit longer than the 2-stage (downscale - upscale) initial generation.

I haven't run out of memory yet, but if you have issues with the VAE Encode stage, you could try swapping in a tiled version of the node:

I believe the standard VAE Decode auto-switches to the Comfy tiled code if necessary but I'm not sure if they do that with their VAE Encode. I've not had a memory issue yet though!

Enjoy! :)