Sign In

Hassaku XL (sfw&nsfw) beta

Checkpoint Trained
Feb 2, 2024
Base Model
SDXL 1.0

Model is a WIP! expect 3-5 weekly updates, later beta versions are on my patreon for a short time (only have a single 4090, can't train faster). My Discord for everything related to anime models. You can also support me on my patreon.

All versions up 0.2M are trainined with quality prompts like masterpiece, best quality. Recommended prompts in negative are "worst quality, low quality, normal quality, signature, blurry", with additional nsfw, if you want sfw art. Alpha to 0.2 don't use quality prompts.

Undertrained concepts appears sometimes to oversaturated, reverts to realism or appear basic and flat. Most training data was simply to put nsfw into the model. Later versions are more stable and don't want to make everything see-through with nipples, skintight and nsfw.

My models: hassaku, sudachi(flat 2d), koji(2d), yuzu(light semirealistic), grapefruit(old hentai model)


Sponsors: with its amazing creators program supports all kinds of creators like me! Preinstalled with 80+ high quality models, hundreds of loras and a new animation feature, join their discord community here! hosts the best Stable Diffusion models on fast GPUs. You can run Hassaku at:


Thanks to my supporters Riyu, SETI, Jelly, Kodokuna and Gpr0mpt on my patreon!

You can support me on my patreon, where you can get other models of me and early access to hassaku versions.


Using the model:

  • Use mostly danbooru tags. For better promting on it, use this LINK or LINK. But instead of {}, use (), stable-diffusion-webui use ().

  • Exclude quality prompts like "masterpiece", "best quality", "worst quality", "low quality". My model was not trained with them, I train only with masterpiece and high aesthetic images. I have not the resources to train lower quality images with low quality tags for more diversity. Use 1girl, 2girls, 3girls.... instead, it is more or less the "masterpiece" prompt of this model. Recommended is also hair and eye colors.

  • Model is trained with images with minimal disruptive features like floating texts, logos and signatures as possible. If such features appearing in a image, use prompt "signature" to use it as negative. I pruned everything down to that single prompt "signature".

  • Excluded meta data and franchise tags, don't use them (meta data -> something like "highres", franchise tags like "re:zero kara hajimeru isekai seikatsu" are pruned on charaters like "rem_(re:zero)")

  • For now, artist tags does not work, but will work in the future

License info:

Because it merges ANIMAGINE XL 3.0 with 0.2M
Fair AI Public License 1.0-SD