Hassaku aims to be a anime model with a bright and distinct anime style.
You can run Hassaku XL and use its API on SinkIn: https://sinkin.ai/m/JWknjgr
V1 use illustrious-xl as base, versions from 0.2M up are trainined with quality prompts like masterpiece or worst quality. Recommended prompts in negative are "worst quality, bad quality" and masterpiece in positiv. There is also my pony xl variant.
My Discord for everything related to anime models and art. You can support me on my patreon.
My models: hassaku, sudachi(flat 2d), koji(2d), yuzu(light semirealistic), grapefruit
_____________________________________________________
Supporters:
Thanks to my supporters Riyu, SETI and Kodokuna on my patreon!
_____________________________________________________
Using the model:
Model is trained with images with minimal disruptive features like floating texts, logos, speak bubbles and signatures as possible. If such features appearing in a image, please use prompt "signature" as negative.
Excluded metadata and franchise tags, don't use them. meta data -> something like "highres", franchise tags like "re:zero kara hajimeru isekai seikatsu" are pruned on characters like "rem_(re:zero)".
Here are some good resolution options for SDXL:
1024 x 1024
1152 x 896
896 x 1152
1216 x 832
832 x 1216 (most recommended)
1344 x 768
768 x 1344
1536 x 640
640 x 1536
License info:
Because it merges ANIMAGINE XL 3.0 below version V1 and Illustrious-XL & WAI-NSFW-illustrious-SDXL beginning from V1:
Fair AI Public License 1.0-SD
V2 is trained on its own and don't include any merge, just Illustrious-XL 0.1, but trained to include newer or missing characters, which also use the Fair AI Public License 1.0-SD license.
Model was trained on more actual data, but only on a small dataset of 50k images (mostly mihoyo games). With a month of experimenting and a hard reset inbetween, because of better training settings. New characters are relative weak trained in and the style merge mostly erased details of them, in which mostly the clothing suffers. It is more or less a prototype to not make the same mistakes of V2.1, that resulted in sampler problems.
better (all in compare to 2.1fix)
- Less orange looking images
- Style is overall more even torwards anime
- Model is a bit less restrictive and more expressive
worse for some
- more anime looking, which also makes backgrounds less detailed and less cohesive
- less mature looking faces
- Can be darker on some concepts (also more shadows on faces)
- Pupils are more filled out (less circle pupils)
worse
- hands/fingers
- text (model goes more towards japanese writing, when it wants to generate text, which is disjointed)
Can't say
- Loras should work better, because of a better inner style merge, with much less artifacts. But i can see that a strong style can backup loras to work better. It mostly depends of the loras itself.
2.1fix have a more 2.5d feel to it, like more details on bodys (like showing rips more often). Which also translate to backgrounds. If you like that, you can still use 2.1fix further.
If you question why i can't stay on the same qualitys of 2.1fix, its because i can't be more "reckless" in my mergin, without actual hurting the new knowledge in the model. Mergin another model in, erase every training I done on it, because all other models knowing less then my model. While every longer training on it, resets it to be like every other base model like noob or illustriuos, without style merge.
Files
hassakuXLIllustrious_v22.safetensors