LTX-2 ComfyUI Power-Workflow! (Patreon)

Video - https://youtu.be/tufeXrzYgrs

Overview

LTX-2 is a "source available" (yes, I'm not calling it open source due to the various limitations ;) release for generating AI videos which also include audio.

For more information on the limitations, see their licence file - https://huggingface.co/Lightricks/LTX-2/blob/main/LICENSE

There are various models available to suit almost any GPU, though the quality may vary as the model file size decreases.

Generation time will vary depending on the resolution, number of frames and your own hardware, but with a distilled model needing just 8 steps for the 1st stage generation, videos can take under 30 seconds to generate :)

This is quite a chunky workflow and can do lots of things all in one place, but I haven't included the various canny/depth control things. What you can do, however, is:

Use the distilled model

Use the non-distilled models

Enable Stage1 and/or Stage 2 Distill as you like

Try 2x FPS "upscaling"

Use a Starting Image

Use First/Last Frame (both or just last)

Use Custom Audio

Use Negative Prompts - even with CFG1 (via NAG)

Have all your common settings close together and easily accessible

Workflow Information

There are quite a few features packed into this workflow, with all the common settings you'd want to change easily accessible via the various settings panels

You will need to install a few custom nodes - mostly for firstframe/lastframe (TTP) and some maths (ComfyMath):

Note: Workflow version 8a EXCLUDES the TTP Toolset nodes for those who can't install in on their systems for whatever reason.

Model links are included in the note in the workflow:

The default setup is for the distilled FP8 model with the Gemma3-12B text encoded, but other models are available as well.

If using a GGUF, you will need to replace the Load Checkpoint node. City96's GGUF UNet loader was recently updated, so be sure to have the most up-to-date version.

Also note that the fp8-distilled model was updated by LTX, so download the latest version for a VAE fix!

Gemma3 also has as fp4-mixed option, though I've not tested it personally.

The model loader has some settings to take note of:

As mentioned, the default is for the fp8-distill, but if you use the other models (such as fp4-dev or fp8-dev) then you should ENABLE "Use Distill Stage 2" even if you're not using the LoRA for Stage 1. This is because the second stage is only 4 steps and so really needs the distill LoRA :)

Optional LoRAs can be set in the PowerLoRA loader:

This includes things such as the camera and detail LoRAs.

Video Settings is where you set the main video resolution, fps and frame length:

If you want to use images, there are 2 options available by default...

... though you can easily extend to use mid-frames. I've put an example node up in the "engine" area:

You'll need to duplicate as many of the "load image" groups as you want to add middle frames for, set their image constant names, and of course set a position for each.

I personally prefer only using 1 image at most, but hey!

Prompting has a "NAG" option, allowing use of negative prompts even with distilled models as well as different negative prompts for audio and video:

Sampler settings has the common things for samplers, such as steps and CFG:

Comfy has manual sigmas for their IMG-2-VID workflow, so that option is available as well via the "Enable Manual Distill Sigmas" option. If enabled, the steps will be IGNORED so you should use CFG1, and either the fp8-distilled model or enable the distill LoRA for Stage 1.

Note that Stage 1 samples at 1/2 resolution, then Stage 2 sampling does the 2x upscale. This makes things faster :) Stage 3 "upscales" the FPS but I found the image quality could be slightly impacted. I tried running it through a custom sampler as well, but then the audio went out of sync, so that isn't included here :/

You can use your own custom audio as well:

I tried using Comfy's "record" node as an option, but it complained the audio wasn't in stereo.

The "Mask" setting acts as a sort of "strength", with values about 0.75 resulting in significant changes from the loaded audio.

Most of the things in the "dragons" section you won't need to change or even look at...

... but of course feel free to examine and play to your heart's content!

Have fun! :D