One of the styles used by artSAM.

Trigger word is @artsam. The source material are AI-generated images. Use appropriate negatives for any undesirable biases you encounter. Three versions of the model are uploaded:

• Original: Uses 36 images from the author's pixiv. The images are heavily (even excessively) censored, so putting censored, mosaic censoring in the negatives and uncensored in the positives is highly recommended, even for SFW generations. If the censorship persists, you might need to reduce the LoRA weight or use inpainting.

• Augmented: Uses the same 36 images plus another 12 generated with the original LoRA and inpainted to remove the censorship. Censorship-related negative and positive tags are still necessary, and the style is somewhat diluted due to recursive training, but it is at least feasible to get a decent uncensored image on the first try.

• Uncensored: Uses the 36 images from the author's pixiv, but manually uncensored with the help of Krita and the augmented LoRA. This is the version I recommend to use. The style adherence is on par with the original version, and the censorship bias is completely eliminated.

This has actually been a much more challenging undertaking than I had originally intended. Initially I stumbled upon the artist's generations at random. The style appeared vaguely familiar, and I quite liked it, so I messaged them with a question about it, but they never replied. I tried looking through boorus to find the style it reminded me of, but the search was inconclusive. So I decided to make a LoRA. The small available dataset and the extensive censorship have made it rather tricky, however. Decensoring the images accurately turned out difficult even with the frontier local models (Qwen-Image-2.1 failed to respect the style, and Anima Edit struggled with the balance between removing the mosaic and preserving the overall details). A LoRA trained on the censored dataset, as expected, had a strong bias for censorship, despite the tagging efforts. At some point I considered giving up and purchasing access to full collections of the author's generated images for $4, but the store didn't accept bank cards and the paypal account I created for the occasion seems to be stuck in some initial validation limbo. Eventually I came up with and implemented the three-step approach described above. Even so, decensoring the images took many iterations of inpainting, reference edit transformations, and manual editing in Krita. It took me between 4 and 10 hours to make this LoRA in dataset preparation alone. If you're looking to undertake a similar project, one big piece of advice I can give is to first blur the mosaic-censored area to the point that the pixelation is no longer noticeable. It is quite hard for a model to ignore the mosaic patterns, while a blur is far easier to lose in the denoising process.

I'd like to take this moment to express my disapproval of the style author's approach. I am fundamentally against selling AI-generated content. (Do note that all of my LoRAs are released on CivitAI with an explicit prohibition of selling the generated images, the models themselves or their derivatives). I am in fact generally against selling anything digital, because a copy of a digital resource costs nothing to produce once at least one copy exists. (Do note that all of my LoRAs are released for free, because once I've trained it for myself, it costs me nothing to make it available to others). But I can tolerate it when a person is genuinely striving to provide something to others, and asking for money out of necessity. On the other hand, when a person does something with the sole intent of profiting from it and takes efforts to prevent others from benefitting from the fruits of their labor, I find it rather despicable.

I will still link to the author's page, but I am not encouraging anyone to support them