(つ´ω`)つ

有罪！ε٩(๑> ₃ <)۶з

Hello, thanks for your attention and support. The training data for this model are screenshots from the game's gallery. As long as the dataset is large enough, the model's performance should be fairly reliable :) However, I recently compressed and archived this dataset, so I can't share it at the moment (crying)~( ´•̥̥̥ω•̥̥̥` )

@aminuoshit Hello, based on my personal superficial understanding, LoRA is actually used to increase the probability that the base model will produce the desired group of pixels. Compared to the base model, the dataset used to train LoRA is too small, so unless you want a "blurry" or "pixelated" effect during inference (usage), you don't need to label these tags to describe the quality of the dataset.

From my personal experience, when using the open-source SDXL series models, to train a LoRA with some degree of generalization, you should have at least 30 images. My personal view is: as long as the dataset is large enough, it can solve all problems, haha ：）

PS: For example, according to the metadata of this model, it used 582 images.

@aminuoshit Large datasets are actually a lazy approach. In fact, as long as each keyword has 3-5 corresponding collections of pixels, you can achieve very good results. Therefore, using just a dozen or so images is completely sufficient to create a high-quality LoRA, though this requires well-designed prompts and a relatively high number of training iterations. The large dataset approach is similar to a Chinese saying: '一力降十会' which is the method I, a foolish person, use :)

Thank you for your blessings, and I also wish you a Happy New Year. I look forward to seeing you create the LoRA you desire, and you are always welcome to come and play :)