LTX 2.3 - LTX Image Generator (LORA + Workflow)
Make sure to download the workflow from the 'optional downloads' as well!
A conceptual lora and workflow to create 1920x1280 images with LTX 2.3.
Remember image generation on WAN? Now you can generate images with LTX too.
The workflow captures the last frame out of a fade-in video (lora trained) where the resulting image appears.
Usage:
The workflow is absolutely necessary
(As it has extremely specific parameters)
Preferably, use the following sentence to activate the image generation
jxlstyl3, A fade in from black to a static image
Keep the initial parameters, although you can experiment with
' cfg' , 'custom lora strength' and 'distill lora strength'
Please generate on 1920x1280 resolution
Although other resolution may work, it was mainly trained for 3:2 at 1920x1280
Feel free to post your results, though keep it decent
Let me know your feedback, as this is the only image generator workflow that I know exists for LTX so far :)
Known issues:
This is a concept for generating high res images, and I think it works pretty well
However, LTX 2.3 is a video model, not an image model. Realise that the model is being tricked and mis-used into behaving like an image model.
A limitation of LTX is that reflection and text do not work really well, which this lora did not resolve
The workflow may complain about an image needed to be loaded, just feed it an image (ignore that it will have any effect though)
You may need to either install Kijai's nodes, or bypass the ChunkFeedForward node inside the subgraph, follow the install instructions here:
If you really need to use your own workflow or use the on site generator (I haven't tried), then these are the required parameters
Pass 1 - 30 sigma steps @ CFG 2, Lora strength 2, Lora distill strength 0
Pass 2- 4 sigma steps @ CFG 1, Lora strength 2, Lora distill strength ~ 0.19
And somehow take the last frame from the video
But just use the supplied workflow if you can
Why make this?
I was curious to see what LTX 2.3 could do when tricked into creating static images. By using a fade-in transition on a high-res image dataset that has been pre-processed (added film grain and some bleach bypass) it became possible to let LTX 2.3 generate a static transition video of which an output frame can be extracted as an image.
While it took a few working days and many hours (40?) of lora training experimentation, to get to a working state, I am now content with the result. There is some room in the adjustement of parameters though, so if you happen to have better results, let me know!