This one uses a very small dataset, because my intention here was to capture not so much the general style, but the portrayal of some particular concepts (check out his arts of <code>entoma_vasilissa_zeta</code> to see what I mean). I wasn't sure if I wanted to publish it at all, but I guess there's no harm in it.In terms of the training process, this time I tried to account for the advice I received from the comments on my last LoRA, and settled for under a thousand steps overall. (The published version is an early epoch, though). I also used gelbooru tags directly instead of auto-captioning.If you like the style, please support the <a target="_blank" rel="ugc" href="https://www.pixiv.net/users/7454">original author</a>25.04.2025 - Updated to IllustriousXL. See update notes for details and the trigger word. This version seems better, but there is another one, closer to the original. Tell me if you want that published as well.

4shi style (ILLUST & PDXL)

Due to the small dataset, I believe later epochs are somewhat overfitted. That was desirable for me, but for general use I feel an earlier epoch works better. If you want me to publish a more overfitted version too, feel free to ask.

Files with SHA256: f52bf199ede28b2c7048aa5e849b5160f2185dc8a209ec1c2e6d75d943ad28a5