Trained on PonyXL with ~500 images by Seraziel I had saved on my phone, mostly NSFW.I really like Seraziel's works, but so far noone's made an SDXL lora of his style, so I thought I'd give it a go. I tried training this on my local machine, but it turned out a laptop with an rtx2070 super (8gb vram) doesn't cut it. I tried renting out a server with an rtx4090 (24gb vram), and that technically worked, but I wasn't satisfied with the results. (Probably has more to do with me than the machine, obviously, but also it was getting quite costly by my meager standards). So I turned back to civitai's built-in trainer, sinking ~2500 buzz into 30 epochs of training. It's not too bad, but it still seems I have much to learn. If anyone has any insights they can share, please leave a comment.The trigger word is <code>seraziel</code>(PonyXL version) / <code>serazielart</code> (IllustriousXL version), and I'd recommend putting something like <code>blurry</code> in the negatives if you encounter any weird effects. This should know the concepts of <code>futanari</code>, <code>human penis</code>, <code>animal penis</code>, <code>dog penis</code>, <code>knotted penis</code>, <code>horse penis</code>, etc. quite well, if you're into that.If you like the style, please support the <a target="_blank" rel="ugc" href="https://twitter.com/Seraziel">original author</a>24.04.2025 - Updated to IllustriousXL. Same dataset, but the trigger word is now <code>serazielart</code> to avoid any preconceptions the base model might have.Also, this time CivitAI finally recongnized me as worthy of using the Early Access program. But, with the site in decline, I reckon I'd rather spend my remaining buzz making something nice, than scam the community, as the site seems to encourage nowadays... I hope I'll get to build a new PC before CivitAI becomes completely unusable, finally letting me train IllustriousXL LoRAs locally. But I don't know of any good alternatives to CivitAI in terms of community model sharing. If you know any, please tell me. I know about tensorart and stuff, but they seem a lot worse and possibly even more profit-over-people-minded. If we're about to devolve back to sharing MEGA folders around, feel free to send me yours!

Seraziel style (ILLUST & PDXL)

Same dataset, different trigger word to avoid base model's preconceptions

Files with SHA256: 5ea141a5cfc0282d2aac62d7f503cde006590f4fcb10a38f0cf0c75060e7d1ca