You can download the lora from my Civit - https://civitai.com/models/1553692?modelVersionId=1758090

I've used the official trainer - https://github.com/Lightricks/LTX-Video-Trainer

Trained for 2,000 steps.

The new model is pretty insane. I used both previous versions of LTX, and usually got floaty movements or many smearing artifacts. It worked okay for closeups or landscapes, but it was really hard to get good natural human movement.

The new distilled model quality feels like it’s giving a decent fight to some of the bigger models while inference time is unbelievably fast. I just got few days ago my new 5090 (!!!), when I tried using wan, it took around 4 minutes per generation which is super difficult to create longer pieces of content. With the new distilled model I generate videos at around 5 seconds per video which is amazing.

I used this flow someone posted yesterday:

https://civitai.com/articles/13699/ltxvideo-096-distilled-workflow-with-llm-prompt