Genshin TCG Style [LTX-2.3]

https://civitai.com/models/1768496/genshin-tcg-style-wan-14b

Trigger Word: Genshin_TCG
Model: LTX-2.3 22B dev
For inference used gguf_Q8
Fork of musubi-tuner was used for training.

Training Details

The same dataset was used as for Genshin TCG Style [Wan 14B] without sound. The training launch was the same as that of [OIIA LTX-2.3]. I used the same training parameters. But I took 3500 steps for the best lora. I noticed a slightly strange behavior of LTX in relation to prompts, it differs from Wan. In my training dataset, there is a detailed description of the golden frame that is in the original TCG dataset. The Wan model, after training on this dataset, did not generate a golden frame during the output, in order to get this frame from it, it was necessary to explicitly specify this in the prompt. But LTX has a different situation, it generates a gold frame whenever it wants. Which is generally strange, since lora doesn't need to memorize the objects explicitly described in the prompts to the training data. Moreover, the longer lora studied LTX, the more often she generated a golden frame. And after the 6000 step, it was constantly generating golden frame... Perhaps it's a side effect of rank 32. But I didn't do any tests at the 16th rank. Well, basically, this lora isn't very popular, so I didn't have the motivation to do a lot of experiments with her.