Updated: Jan 13, 2026
concept
Two trained LoRAs for LTX 2 T2V (official trainer, full weights no quants, 1.5mp dataset resolution):
LTX2_SS_Motion_7K / trained on 508 video clips
LTX2_SS_Details_30K / trained on 780 images
I was thinking of using a two-stage training approach like I do with Wan 2.2: → Motion LoRA for the first stage → Details LoRA for the second stage
Unfortunately the results were not as good as I had hoped, especially when using the distilled model.
I got better results with the non-distilled model when:
using a slightly higher step count (~12 steps)
using the distilled LoRAs at lower strength (~0.4) to avoid overcooked look.
I still don’t have any clear recipe for the perfect strength mixing, so I consider these models experimental for now. I decided to share them anyway, maybe I’ll retrain better versions later.
All example videos have the full workflow embedded, so feel free to download them and simply drag & drop into ComfyUI to check the exact settings used (prompts are generated from local files so you will only get the prompt from the previous run).
Don’t mind the stupid slop prompts lol, I just ran a batch and forgot about it.
I inject some money to train models on runpod but It wouldn't be possible without your help, so I appreciate those who used my refferal link, providing both of us free credits to use their service for longer sessions. You can also support me on buymeacoffee If you want to, it's appreciated. Thanks.