Thanks... I trained it with Ostris Ai-toolkit. I used the recommended config from the Ostris video to train LTX 2 loras with 24 GB VRAM (a RTX 3090 in my case) but with some changes: Using Prodigy as the optimizer instead AdamW8bit, LR 1 and Weight decay 0.01. I used no captions, only the trigger world/phrase (in my case, 'ani-mat, an anime scene of'). This works really good for styles. I trained it 8000 steps, with a dataset of 421 images, all AI generated, the same dataset used with my Mature anime screencap style lora of Z-image Turbo. The image dataset was generated by combining an Illustrious checkpoint with mature body types (one of the versions of Mature Ritual, I don't remember which one) with a lora with the style of T-rex Studio (I like the colorful screencap style of that studio). I found the resulting style very beautiful and eye-catching, so I decided to train a lora specifically in that style.

For other loras, for example multi-concept loras like the Hentai version of Mature Anime Screencap style for Z-image Turbo, i used Joycaption Beta One to caption the dataset images, with the Taggui app to facilitate editing each caption and manually reviewing each one of them (approximately 950 images in the case of the lora mentioned above).

@jaimd In theory, it shouldn't be a problem, although 24 GB of VRAM is already a bit insufficient for that. Just a few days ago, Ostris published a video tutorial on his channel specifically for training with LTX 2.3, although it was for training a character. However, it's a good starting point.