Hassaku aims to be a anime model with a bright and distinct anime style.
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:
The model was trained primarily on images with minimal disruptive elements, such as floating text, logos, speech bubbles, or signatures. If any of these elements appear in an image, please include "signature" in your negative prompt.
Danbooru Metadata and franchise tags are not used in training, so please avoid using them in prompts. Metadata such as "highres" and franchise tags like "re:zero kara hajimeru isekai seikatsu" are not used in training.
A CFG scale between 3 and 7 is recommended. For best results, use CFG 6 with the Euler a sampler.
Refer to the example images, to understand, how the model should be used.
The model is designed to be simple and straightforward to prompt. Use the following tag order:
Number of characters
Character Tags
All remaining tags
Example: 1girl, rem \(re:zero\), standing, masterpiece, upper body
Here are some resolution options for SDXL:
1536 x 640
1344 x 768
1216 x 832
1152 x 896
1024 x 1024
896 x 1152
832 x 1216 (most recommended)
768 x 1344
640 x 1536
Lora use:
NoobAIs Loras working for the most part better in compare to Illustrious loras
Versions 2 and 3 is less compatible to most loras trained on Illustrious or Noob, because it got further trained away of both. Style A is still the most compatible for it.
______________________________________________________
Version and License info:
Below V1 merges ANIMAGINE XL 3.0
Version V1 use Illustrious-XL & WAI-NSFW-illustrious-SDXL with additional training
V2 is trained on its own and don't include any extra merge. Base was V1, it is trained to include newer or missing characters. It was also used for further training tests
V3 is a merge of Illustrious-XL and V2.2, to fix some issues and was also trained to include newer or missing characters
All Models using the Fair AI Public License 1.0-SD license.
Description
It behaves and looks mostly like Style A, less flat looking but also more drawn in a way. Model style is now build up from the ground up, without any other merge, like mergin with wai, simply to actualy update my model on new concepts. Because of that, its basicly a new model and a bit less stable in compare to style A.
Don't expect, because its named V2, thats magicaly much better. When you train, merge or make any changes, it tends to impact other concepts significantly. V2 means a new base to work on.
Compared to 1.3 A, a honest review over my own model:
Improvements:
- Better at prompt following
- Clearer images with less of a "fog" over the image (for concepts that had that)
- Slightly better at handling furry-related prompts
- For most concepts, more filled and coherent backgrounds
- Better at artist tags
Improvements for some:
- Model have a bit more freedom and variety, that comes also with a wider array of style on a singe concept (some images are more detailed and others are more flat on the same prompt)
- Stronger contrast, but some concepts have a yellow tint
- Bodys have more details, with less 2d flatness, featuring a slight shine and sometimes stronger outlines
- Faces appear slightly more mature
- It likes to draw fingernails more often
Worse:
- Nights are less dark
- Some images can look overexposed or "burnt" on color
- Bodys, because they are more detailed, can look washed out, at some prompt combinations
- Artifacts are more frequent; slightly worse than style A, but much better in compare to style B
- Eyes forming less well-formed pupils (dependent on concept and character distance), in combi with more artifacts, makes it harder to have clear eyes. A point to use Style A over V2. But it also have a bit to do, that faces are more mature looking and having smaller eyes with it.
- Looks sometimes "dirty"
- A bit more unstable
The model was further trained on an dataset of around 60k (more recent characters) to 600k (total images).
Will work first on a update, to make eyes, lines and overall look more clear and harmonic
FAQ
Comments (71)
Is it a problem with the Civitai server? Various models and LoRAs have randomly shown the error message:
"404 The page you are looking for doesn't exist."
I am afraid. I fear that the high-quality models and LoRAs created by talented contributors might disappear.
Some models got unpublished by civit accidentally, models and loras should be back in some minutes
@Ikena I was afraid because LoRAs uploaded by various other uploaders also showed a 404 error. I always appreciate your contributions.
I'm really relieved. I want to be able to see the works of talented creators who make AI art-related checkpoints and LoRAs for a long time. Even though it was a short moment, I was truly afraid.
Did v2 is trained on ILXL 1.0 or 0.1?
on 0.1, or better say on the model i train on since V1
@Ikena understood. Because i made lora mostly using your model, so i need some confirmations xD
Why is no recommended settings in description. Like Sampler, scheduler etc?
it works great no matter what
Use what you like, i don't want to force someone to use specific settings, because i say so. I test and generate my images over basic EULA A with balanced 20 steps,. Simply to ensure that people using simpler settings having great images as result. But nobody said that you can't use another sampler or more steps like 30.
Guys how do you know the specific description tags for a certain character from booru? Is there a excel file or something like that?
@lqc thank you, btw your Ie Style is very good, i just found it 3 days ago and loved it since
@chuberongonline03307 :)
im looking forward for v2. ive been using your refine since i first learn how to do ai using my own gpu. thank you so much.
Same here. Started with the Pony model, and just got Early Access to v2. Thank a lot!
🦊👍
Style B is just the best. I have tried a dozen different illustrious checkpoints, and this is just the most beautiful one
I'm in total agreement! Style B produces very beautiful images even without any lora.
Very good!
So description said "don't use any metadata" except "Masterpiece"? Am I correct?
Anyway this model is very unique style and beautiful.
Hassaku series and style B are the most amazing model.
Have to say v1.3 style B is the most marvelous model I have seen. It is unique! Cannot find any other model having similar effect. I have been totally besotted with it! Please keeping updating it, Great Ikena!
What is new in v2?
Can someone please guide me how to avoid multiple others/person in the background. Are there any tags that i should put in negative prompt to ensure that it does not happen. I was using style B model.
Sorry i don't remember exact tags i used. But what i tried to do was:
"person in medieval clothing style standing infront of a mirror, in candle lit room, looking at his own reflection".
So 1/5 times there was someone else in the room. The prompt even had "1boy" only not anyother tags like "multiple others" or "others" or "1girl" etc
Use tag solo in the positive prompt
@MV_QT Thanks for replying, will test it out
"(other person, other boy, other girl:1.4)" in negatives - as I recall, this worked to some degree for original SDXL, to a much lesser extent for Pony and Illustrious. Better to combine with "(solo:1.3)" in positives.
V2:
- Better at following the prompt, creating characters with specific poses without using LORAs
- Worse at creating eyes and hands
Overall a nice attempt to improve style A. I'm looking forward to new update of V2 since i'm a style A fan as well.
V2 will be allowed to use on site?
Can only have two models on the generator and because Style A is a bit more stable, i think of let it on style A, until i make v2 better with at least the eyes
@Ikena eyes on Civitai generator are terrible indeed, they're add so many things that nobody asked for and not a sight about adetailer... 😅
Well, hope you manage to add it, we need good illustrious models on site ❤️
Thanks for replying
I actually pointed that out in "About that Version":
- Eyes forming less well-formed pupils (dependent on concept and character distance), in combi with more artifacts, makes it harder to have clear eyes. A point to use Style A over V2.
I work on that, to make it more clear and clean.
It's a perfect model
i love it
I feel happy to be able to access early access and there is no regret spending buzz for it.
Throughout my time trying V2:
- Some characters like Yanagi zzz are not so well recognized in V1.2 but in V2 they can be recognized from her clothes and hair. of course with the help of lora it is much more precise in her clothes
- My Lora style works well in V2. I really like this considering my Lora works well in V1.2 but there is a difference in V1.3, (this is the reason I use V1.2 rather than V1.3), I am happy that in V2 the results are not so different from V.1.2
Minus:
- The pupils are a bit messy compared to V1.2, but can be reduced a little with the help of a detailer. I have read the version details on V2 and I think it's not too bad considering it can still be slightly improved with adetailer and hiresfix
- Fingers will be a bit messy in some scenes, for this I changed the sampling from Eular A to DPM++2M SDE, as long as I see the image becomes sharper from the lineart side and the fingers are better.
I have also supported through Ko-Fi and hope that it can help the author to create fantastic models in the future.
Overall I really like it, hassaku has become my favorite checkpoint when I was learning SD1.5 until now. It's nice that the author is still active until now ^^
Thanks review! i'm aware that eyes are curently messy, will work on that the most. Style in this version is more like 1.2, at least for bodys and details around that. Maybe a reason why the loras you use resulting in a more 1.2 like look👍
@Ikena Yep ! i really like v1.2, but for 1.3 I don't really like it because it seems to have its own style so that the lora style effect is not so visible. especially in V1.3 there are versions A and B. well anyway some people have their own tastes especially it seems that many like v1.3B and we have to respect that.
Just to clarify, Illustrious 1.0 is not in the training data for v2 correct?
correct, no merge with Illustrious 1.0 (the last image should clarify that, put it there, because i think some would asume that)
@Ikena Thank you
V2 really doesn't play nice with AlignYourSteps schedulers. Not a big deal, just making a note.
I don't understand this decision to prune franchise names. Sometimes it helps to get the style of a certain anime without calling out characters from those franchises.
Because some franchise names are extremly long and takes a lot of the tocken lenght in training, that and you need to put the franchise name to a character name, to have the best depiction that way (less training with one uniqe name tag). Yes, its helps to style results to a franchise. But i never seen that used much + there is the option to use artist tags. So i decided to prune it in training
Does that model have problems with fingers? A lot of generations end with molten or deformed fingers. I use just the model without LORAs, and Euler A as scheduler. I also tried to put "bad hands" and "bad fingers" into the negative, but it did not work properly.
try "bad hands, extra digit, extra fingers" together and use more then 20 steps
OK, sorry to say but the prompt handle is in someway to simple. Meaning if you use more complex prompts to specify what you want, it gets more and more bad. If you do simple prompts, it's doing a bit better Pics but not what you want.
not sure why but the 1.3v version seems to work better for me, maybe I need to re-tweak the prompts specifically for the newest one...
V2 is more of a new start and 1.3 had some months to get better from V1. V1.3 is more stable and a reason why i let the generator on 1.3👍
Has a problem where pussy always glows or is distorted.
Sadly a common problem on every model (except base model like noob, animagine, illustrius). It comes with mergin styles into a model.
Does this model incorporate Illustrious XL 1.0 as the base?
no merge with Illustrious 1.0 (the last image should clarify that, put it there, because i think some would asume that)
v2 will generate noisy images if I use karras scheduler, is that an intended behavior?
No, will look into that
@Ikena
here is an example of an image with normal scheduler
https://files.catbox.moe/w7fq50.png
and this with karras scheduler
https://files.catbox.moe/tr8bwe.png
using character lora make the noise even worse
@TOF_enjoyer Thanks, will fix this in the next version
oh i thought it was just me.
First of all, very good models and i love Style B ! Maybe not the best model concerning LoRa but surely incredible for the rest ! Good job !
As For V2, will we get V2 Style B in the future ? or maybe planning to make style B a LoRa ? Eager to see how this absolutely stunning model will evolve in the future.
Keep up your amazing work !
For 2.1 maybe, want to make V2 more stable first
I really liked v2 more than other models, it's amazing!!
It's now actually and visibly worse than 1.3B. Worse details, more boring compositions.
v2 is a Style A variant general purporse model, while B have a much stronger style merge with another goal in mind. Its artstyle is forced for more details, but have its own flaws. But i disagree on more boring compositions. Style B have a more narrow variety of compositions, filled with much more random wrong placed contents -> like random armored guys with tag ruins in the background, or bunny ears creates bunnys in the background, while A does not.
@Ikena huh, that's weird. Maybe I did something wrong. I'll keep testing!
Thank you for your hard work. (I'm a big fan of 1.3B and use it for chkpoint merges.)
Maybe im doing something wrong but i just cant seem to make v2 not produce blurry/artifacty images
Is V2 gonna be available for online generation?
next week are all models on the generator
I've been using V2 for a couple of days, and it's a very good, versatile model. I created a wonderful space scene. I sincerely don't think the pictures posted by other users represent this model correctly (some of them, but 95% not).
cool
Details
Files
hassakuXLIllustrious_v2.safetensors
Mirrors
hassakuXLIllustrious_v2.safetensors
hassakuXLIllustrious_v2.safetensors
hassakuXLIllustrious_v2.safetensors
hassakuXLIllustrious_v2.safetensors
hassakuXLIllustrious_v2.safetensors
hassakuXLIllustrious_v2.safetensors
hassakuXLIllustrious_v2.safetensors
hassakuXLIllustrious_v2.safetensors
hassakuXLIllustrious_v2.safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.