This Version of Hassaku is the hentai variant of my other hassaku xl model. But again, like my 1.5 model with a priority for hentai art. It is trained on top of pony. Goal is to make it easier to archive good hentai anime art, without the need to use complicated source and long quality prompts.
My Discord for everything related to anime models and art.
If you'd like to support my work, you can do so at: SubscribeStar, every bit helps and is truly appreciated!
____________________________________________________________
Supporters:
Thanks to my supporters tamashiicolle, pttcot, SETI and Kodokuna
_____________________________________________________
Using the model:
Use this extension to use it easier with booru tags:
https://github.com/DominikDoom/a1111-sd-webui-tagcomplete
Use score_9 in positive and score_4 in negative
masterpiece -> score_9
best quality -> score_8
high quality -> score_7
medium quality -> score_6
normal quality -> score_5
low quality -> score_4
worst quality -> score_4
You can also use score_8_up, score_7_up and such, but effects are reduced.
Model is merged with a style lora, to make its style stable and to higher up the quality. Because of that, score_4 and score_5 have less effect, Lora /Checkpoint was trained with score_9, but can also be leaved out, for a bit more basic look.width and height should be a multiple of 64, to prevent gray borders (not a model issue, its a general thing for sdxl models)
Tag order: 1girl/1boy, character name, everything else in any order
Some concepts appear with strong 2.5d faces. Negatives like nose, lips and sources helps (in 1.1 reduced and in 1.2 rare, in 1.3 not a problem anymore)
It is a continuation of pony, source_pony, source_furry, source_cartoon, 3d, monochrome, realistic in negative works. Even with less effects on it
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)". Pony was trained on them, it seems.
Here are some good resolution options for SDXL/Pony:
1024 x 1024
1152 x 896
896 x 1152
1216 x 832
832 x 1216
1344 x 768
768 x 1344
1536 x 640
640 x 1536
License info:
Because this model is a continuation of pony, from PurpleSmartAI, it have this license:
This model is licensed under a modified Fair AI Public License 1.0-SD (https://freedevproject.org/faipl-1.0-sd/) license.
The following modifications have been added to Fair AI Public License:
You are not permitted to run inference of this model on websites or applications allowing any form of monetization (paid inference, faster tiers, etc.). This applies to any derivative models or model merges.
Description
Its a bit unstable in overall art style and have some problems with hands/limbs. But is easier to handle in compare to pony, if you want anime art out of it. It knows now artist tags and some anime characters better in compare to pony. What i want to fix/making better on it now, until SD3 comes out:
A bit more anime like (some concepts do strong 2.5 faces)
Make hands better
Less detached limbs
A bit better sfw with better prompt understanding for it
FAQ
Comments (29)
How many images did you train on? Looks nice
initial image count was in the first epochs around 800k explicit and questional images. In later epochs it got down to 200k and end eastetic runs was 20k selected images
@Ikena Oh sweet, i'm testing now and see that artist tags work like Asanagi/Ishikei that aren't recognized by pony. Do you happen to have a .csv file or something that breaks down the frequency of tags? Will be useful to see what artists/etc it knows compared to pony
@Ligmanese I have a script for that, but because of the nature of my training, reducing image count to prioritize my training for lower training time (i don't have the ressources like pony or animagine have), some tags are trained much more in because of that. Even if i count it, it would not say, how good that artist tag would work. My later 200k training had 578 artists in it. But my later eastetic run with 20k only 40. Most known artists should work or should have some style influence with it
Any chance of seeing a Sudachi in the future? I'll use this for my next releases, thanks a lot!
sdxl is hard to train with only one 4090. Sadly no time left until SD3 comes out, maybe a SD3 sudachi is thinkable :)
@Ikena Awwww, I understand 100%. I'm just making Pony characters using the trainer here 😅. Is my favorite model, you probably see me there every week lol. Looking forward to it!
"score_9" doesn't seem to be enough. "score_9, score_8_up, score_7_up" works better.
Thanks for the info! its pony behind, so score_8_up, score_7_up also works, but not as strong like on pony itself. I trained without score_8_up, score_7_up in a attempt to fix ponys mistake (that he also sayed was a misstake)
@Ikena There are a lot of problems with limbs when I do a scene with two people (for model v1). Also sometimes an extra head appears. I hope the model will be better in a future version. Will wait :)
@vickov817 I know of that problem, its in the version description, working on that :)
About this version: Its a bit unstable in overall art style and have some problems with hands/limbs. But is easier to handle in compare to pony, if you want anime art out of it. It knows now artist tags and some anime characters better in compare to pony. What i want to fix/making better on it now, until SD3 comes out:
A bit more anime like (some concepts do strong 2.5 faces)
Make hands better
Less detached limbs
A bit better sfw with better prompt understanding for it
Have you tried training without score tags? This model has a pretty significant quality downgrade compared to pony, similar to when I tried training a lora with score tags. It responds a bit better to some characters and tags, but no matter how I try to set quality tags or style lora, it's much worse than with pony/autismix.
@wewewew I trained with score tags, if you mean score_9 and such. Can you please describe the quality downgrade? is it the style or composition or stability? Because i had some discussions about it and it turned out to be mostly a style problem, because they preffer less flat looking styles.
I aim to have a default style reminiscent of re:shimashima or masami_chie
Pony tends to make more realistic looking nsfw art, why, don't know why a pony model does that
autism use a aki99 lora merge to archive its style
My goal is to make it more anime like overall, so most images in training are more simple looking anime images. Only the last 20k images of my training are there to define its default style with score_9. maybe there i a problem and i need more training on that
@Ikena Style, composition and stability all seem to take a hit, compared to pony + style lora to fix pony's unstable nature. Less detail, especially on backgrounds, less coherent and creative, less cohesion between foreground and background. It's not as simple as pony being more realistic; that's only when no style lora is in use and pony randomly "selects" a style. The downgrade on this model may be partly because pony style loras don't work as well with this model, but I wonder if it's not because pony's score tag system is easily disrupted. Probably not something you can fix without redoing the base model.
On a related note, I also tried comparing models using the exact method for score tags in your model's description against pony's method with "_up", and that also lowered the quality on all models, including this one.
I've also seen from autismix that, at least for one of the versions, it's just a lora merge, but somehow it improves over pony, especially stability, much better than any style lora added to pony normally in prompt. Adding another style lora on top of autismix also works way better than adding 2 style loras to pony, which usually looks hella overbaked. Dunno if it's just low weight, carefully block weighted or what else.
Here's a comparison grid: https://files.catbox.moe/mexqn1.png
@wewewew thanks for the answer! stability/coherent is one of the points i want to make better. Less detail in backgrounds is hard to change, most images i train have simple backgrounds. I asume that pony have also non anime/pony/furry images in the dataset. I test a bit if i can get that better.
ponys style loras can work worse here, because i don't try to merge loras (autismix use 2 loras, one for composition and one for style, so far i know) but I actualy do a redo of the pony model with danbooru/gelbooru images. Can you please provide me with the prompts of the first row and the last row of the comparison grid? It would help to evaluate my training better. Especially the first one should not look like that. Thanks! :)
@Ikena The prompts are randomized with wildcards, and I had not looked at them myself before. I just saw the first prompt had an emoji in it, and removing it vastly improved the output. (https://files.catbox.moe/bakgcf.png)
Dunno if that's something you can specifically fix, but it's interesting that it became a problem on this model. (I have not tested if it's a definite issue with all emojis) Not a very important issue anyway. I think the 4th image is a more common display of the quality downgrade. But here's the prompts you requested.
first row: 1girl, palutena, very long hair, tiara, (green hair:0.9), hakama pants, fantasy clothes, nylon panties, fishnet tights, sideless outfit, fantasy outfit, smug, doyagao, wet, outside, boreal minimalist landscape, standing, 👨👨👧👧, winter, score_9, score_8_up, score_7_up<lora:conoghi_xls-000008:0.8>
last row: 1girl, huge breasts, sylphlike, very long hair, bun, open robe, animal ears, indoors, nuclear powerplant, walking away, spring, a hot air balloon flying over a city, score_9, score_8_up, score_7_up<lora:kanbaki[14]XLS:0.8>
You can find negative prompt and other settings on the grid's metadata.
Nevermind, after some testing, the improvement after removing the emoji had a lot more to do with bringing the token count under 75. I tried removing "minimalist" too but that didn't do much. Going under 75 tokens really makes a difference though.
@wewewew Thanks! Need to test a bit and consider that in my further training. But i can't promise that loras working as great like on pony or autism self, because my training goes further and further away from pony itself. I don't prune artist tags like pony do as example out, its also a bigger difference to what pony is trained on and differentiate my model even more.
What config do you use to make it run? It seems that 6GB VRAM + xformers are not enough, one single image takes 10 min to generate. Maybe it's time to upgrade my desktop... But need 💵
The SDXL models are very large in size, so I believe the issue is that your PC doesn't have enough VRAM. If your PC has 16GB or more of memory, using lllyasviel's WebUI Forge instead of Automatic1111's WebUI might resolve the VRAM shortage problem.
you could alternatively try comfyui
You should give ComfyUI a try. I have 6GB VRAM as well and without changing any settings regular SDXL image generations take around 2~4 minutes here.
Yippee!
I like Hassaku, and have for sometime. However, I can't quite get a consistent output of similar styles. I have generated some images that I really like. Largely speaking, that style doesn't repeat much if at all. Is this still early in development or am I doing something wrong?
It is a sdxl thing, consistency is also a problem on pony. My next version (on patreon) use a lora merge like autism and momoi does, to have a much stronger uniform consistent style. That problem remains, but much reduced in the next version in compare to the current version here.
Additional, my model understands artist tags to some degree, maybe you like a artist style, hassaku can reproduce it
V1.0 notes:
Not gonna leave an actual review this time but here's a little tip for consistency (both for Ikena and anyone passing by, albeit it'll probs be solved by the "style merge" in the next version.)
drop a source_pony, source_furry, source_cartoon, 3d, monochrome, realistic in negative and you'll get something which, while not exactly 1.5 Hassaku's style, is quite consistent-looking.
Hope it helps.
my images for this version (using this neg)
a bit of research on HassakuXL(pony) and negatives:
actual background (imgur didn't let me upload this one for some reason)
