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
I repaired the eyes, but it still as example generates green or red eyes when prompted for brown eyes. However, it no longer produces damaged eyes in my tests. Additionally, it now creates fewer blurry, white, or messy genitals. Unfortunately, I had to remove a bit of the SDXL merge to make these repairs. This part was responsible for making the backgrounds nice and providing knowledge outside of my normal anime dataset and tags. As a result, the backgrounds are a bit worse and less detailed, and the model seems a little less stable due to my merging back to a version without eye problems and my additional training on it.
Using training for better backgrounds resulted in much simpler and worse backgrounds, so I dropped that attempt and created this compromise model, which still generates good backgrounds. You can still use version 1.3, but if you don't care much about backgrounds and don't mind slightly less colorful images, you can use this model instead.
What is better:
Eye artifacts should be now very rare
Fewer incorrect eye colors.
Better formed out pupils (including special pupils like hearts-shaped pupils)
Reduced blurry, white, or messy genitals.
What is worse:
Colors are more muted.
Worse and less detailed backgrounds and landscapes.
Minimal more issues with hands and composition.
Some sfw concepts are a little bit less precise
I am now working on my natural language model to generate NSFW captions for images and on SD3.
FAQ
Comments (148)
I really like this model. I don't find it hard to use, just make sure you know the booru tags for what you want, natural language won't always cut it (though can be sprinkled in). I have been extremely impressed with it's respect for colours as well. The 1.5 model, if you specified colours for any part they would often merge and also apply a colour scheme to the whole image. Which ngl did create some cool images, but in this version the colours stay firmly where they're supposed to even doing things like heterochromia. Score tags can create a variety of relatively consistent styles, I don't think always going for the highest scores is even the way to go.
All around thumbs up.
"Euler a" there are many cases where the eyes are broken or the color of the left and right eyes are different.
"DPM++ 2M Karras" has normal eyes.
"DPM++ 2M Karras" is recommended for model usage.
thanks for the helpful info, reason why there is a better eyes version, but a sampler without that problem is better
also ran into that problem, with an particular lora (i pretty much only generate with loras, usually 1-2 at a time).
raising the amount of steps to 30-35 dealt with it largeley.
When i played with dpm++2m karras i had more problems with hands and geometry/perspective.
would need more testing for true statistical relevance though i guess.
I would like to explain something ...
only TOP Creators have the possibility to use the civitai-generation service ... all others do not ... BUT that does not mean that these models are the best "by far not"
Really good checkpoint, truly excellent.
Hands, by the way, are rendered perfectly well most of the times. Well done!
Since SD3 was an overwhelming disappointment and proved underwhelming, will there be further updates to this model? This has been the best iteration of Hassaku, at least imo.
I am still waiting for proper SD3 training tools to decide if I will drop SD3 completely. Also, the license makes me hesitate to use it further. It's too early to say that I will drop SD3. In the meantime, I tested Pixart Sigma and I am pleasantly surprised by how well sigma is trainable. For now, I am working on a VLM (vision language model) to tag NSFW images with natural language. This will provide prompt adherence instead of the current randomness where stuff appears with pure booru tags. However, I don't have enough data to do this correctly so far.
@Ikena I can only wish you the best on your current endeavors, XD
My knowledge does not extend very far.
hi, first of all you did a wonderful job on this checkpoint, I love it :) Second thing, can you tell me how to create a checkpoint like yours? I would really like to try, I know how to create models but I don't know how to create Styles and checkpoints :(
I use onetrainer to finetune or for creating loras out of pony. There are many tutorials for it :)
@Ikena hi, I found the onetrainer program, but I can't find the tutorials on how to create checkpoints :(
@ApexThunder With tutorials, i mean finetuning tutorials in general. Because i can't tell you here, how to do it, tutorials can do that better.
Here is the link to the onetrainer wiki: https://github.com/Nerogar/OneTrainer/wiki
If you plan to fully finetune a SDXL model, you need the newest hardware for it. I suggest you try first to make LoRas.
As for me, i use 1.3mio images to finetune my models. Each update of my model here is one epoch complete, that takes 2weeks with a additional week of merging, lora & stabilisation training. Some other trainer using less images for training and thats also good. Some pony models are also only lora merges, to force the style to something else. All viable solutions to create a model.
@Ikena and instead for the creation of Styles? how should it be done?
@ApexThunder create a lora and that lora can be merged into a model
@Ikena Ok,sensei :)
Muito Poggers, pretendo usar
excelent i love
Es uno de mis favoritos. Me encanta.
really nice, really bright colourful checkpoint compared to autismmix.
I find it understands stuff that would otherwise require using a variety of other LORAs
The best 2D Anime PONY model I have used so far, and the characters trained with it can be restored very well.
I will probably remake all my loras with this base model.
But the generated background does lack variety and quality, such as my commonly used prompt words: Concert Stage, City. It cannot produce good results at all. This is done very well in Anything XL.
But I am more inclined to think that the problem is the foundation of PONYXL v6, without much background training.
It would be better if more lighting conditions and background images could be added.
Nice model！Easy to HCG！
it is the easiest to use pony model I have found and it generates excellent images.
style and details are good,but can draw too little things,can't draw many common concepts.
Does this model support different artist styles? If so, is there a list?
don't know why, but mine is looking...overcooked, would this be due to me using the newer Sdxl Vae that has the fp16 fix or another issue?
looks like the common bug of auto1111 or forge. If you use direct numbers to strengten weights like (PROMT:1.5) or if you use :3 at the end of all prompts. If you remove some or all of that, it should resolve it.
@Ikena oh, ok, thanks for letting me know.
@Ikena figured it out, was my lora's causing the overdark and overcooked issues, specifically extreme bukkake lora.
@pokezilla39966 I've found these kinds of excessive effects can be greatly reduced by setting the LORA weight around 0.50~0.70
One of the top 3 anime models I have used.
Any chance for an AuraFlow version in the future?
AuraFlow is to big to train on consumer hardware, it is bigger then sdxl. It is maybe possible with FSDP and DeepSpeed, but still extremly slow. I don't have the financial means to go bigger then a single 4090 (mostly because of the running energy costs).
how to use the "better eyes" version on civitai image generator? when i chose the model, it always use the basic v1.3 version
generic results, cute finish but tried different prompts with the same character and barely no change, same position, same proportions
Hi, this is great but I feel like this is a very different style from Hassaku (hentai model) for SD1...
Are you planning to make something closer to that one for Pony please ? I'm asking because that version was Especially Goat in expressing beautiful detailed leg muscles and anatomy details without getting two realistic or too thicc
don't planning to do that and its hardly possible, the style of hassaku 1.5 was archived with unet block merges. Its not possible with sdxl to do that. The style of this model was archived with training and merging my own loras to it.
Will there be more updates? Hassaku is always one of the most stable ones since SD1.5. I hope the Pony version keeps improving ^-^
I plan to train on another model, like flux. I currently work on the other Hassaku XL (nsfw & sfw) model, because it is in a poor state
Hey, what comes next for Hassaku now that Flux is here and Pony v7 is (soon to be) on the table?
Though not sure if Pony will be a candidate given the fact that it is being built on Auraflow.
I'm not sure so far, i only have a single 4090, so i need to consider a model that can be trained in a reasonable time. Auraflow is relative big, so is also flux, but flux have more support torwards training possibilities. Then there is SD3.1 and the new model innitiative of civit. It save to say that i test some of the available models to train and look to what the general audience going to for now. Don't want to put time into a model, that don't see much use, because the userbase uses another model
Stumbled across this while trying a variation of someone else’s image. Amazing
Emphasis mode:No norm
i found this to be one of the best check point I have used. Very good generation compared to others
atm this is the best model at least for me that uses correctly my loras made here in civitai.
No viable vae could be found
A vae is integrated into the model
@Ikena I get Noise when none vae
@2328091835224 then you use a GUI that don't support vaes that are already implemented in the safetensor format
@Ikena I don't know if that means I need to change my hardware or change anything
Make sure yall try the better eyes version (you can swap at the top). i didnt know that this existed until just recently and the results are sooo much better imo. Eye artifacts suck.
great one, fast and accurate
ControlNet not working???
Should work, don't got any message that it don't work
Made a poll if the better eyes version should be up on the generator. Better eyes version won, so i switched it to be the top model. The generator should switch to the top version later.
1.3 works much better,need it back
The poll decided, i can only put one version on the generator. Most want less eye artifacts, instead of better backgrounds. Both versions are near similar in most cases
1.3,need it back
Not a fan of the new eyes version pls go back to previous 1.3 version
Is it only me having memory crashes with using both better eyes version and face detailer?
reduce size of image, may be you are out of memory
Since better eyes is now up for some time, i made another poll that runs for a week: https://discord.com/invite/zSR5FcYWWE
You can decide, which model should be up on the generator.
This will make a fine addition to my model merge.
Calling it now: there will be a Flux or AuraFlow version of this soon in 1-2 years... Cus it will be even more flexible
Kemonomimi characters always have chocolate pocky in their mouth for some reason
Better eyes version won again, so it stays up
Is there a way we can go back to the dynamic creative looks and stylization of 2023 painterly-vibes (like dark sushi creative-like) models with the updates? Everything seems too similar nowadays towards realism. Anatomy is fine, just the style seem all too similar just in different artstyles, if that makes sense for those who spend several several hours to find a certain look, but can't because of the common replication.
I'm not 100% sure but I think "better eyes version" actually produces WORSE eyes in smaller face area scenarios (i.e. full body, from below or wide shot).
I'm Reverting back to plain v1.3
Going XL?! ❤️🔥
works pretty well for me
Are you working on a noobai model?
I working and experimenting currently on my own model
Ok, this its gonna be one of my favorites now
pretty good
Is the default style of this model based on a particular artist or a combination of artists?
Any suggestion to make anime screencap like images? Close to realistic anime coloring?
Best done on a model specifically meant for that. I have no suggestions as I would not know, it's not an aesthetic I go for.
Damn, good AI. Perfect for JAI bot profiles.
When my generations have weird looking eyes, this model is a lifesaver. Thanks!
So Good, best of its kind out there.
Got best results by keeping prompts short
Dear friends, please share some pony models that are good at generating realistic Asian women. I've been looking for a long time and haven't found any. Thank you
a bit strange to ask that on my model
You can try AnterosXXXL, but perhaps it's a bit too realistic
@Ikena Thank you for the comments provided by the author
Thank you@kunjabihariji317
@dj9wqurf My commend is not helpful at all and was simply my through, because my model is far from realistic. Sry i can't help much, maybe ask on the discord of civit, maybe that helps. I myself not use or test any realistic ones
Can i the images from this model to post on my patreon?
What model should I train my LoRAs on to work well on Hassaku XL Pony? v1.3?
On pony itself or Hassaku XL v1.3
what happened to the illustrious model?
I'd still love to know how to replicate this models' style. Probably my favorite style of any model I've played around with.
Absolutely phenomenal details and excellent responsiveness to tags when paired with the GTHAN style lora.
It has some strange filter on the images that leave it looking glare on the camera. Had to try this one because the one I regularly use isn't available due to this bullshit auction system and I find it lacking.
This seriously needs to be upvoted in the auction system. I find Hassaku is the must go for hentai art.
100% AI image that I generated are always HassakuXL (Pony). It creates a nice and beautiful art of any prompts (including NSFW). Sadly, now you no longer generate this checkpoint.
Its the bidding system, not me that decide that
@Ikena I supported u in petrean .you have made a great things. I promise i will always support u
@Ikena Oh I see. I don't know what happen to civit.ai nowadays that some of checkpoints/LoRa now no longer generate-able instead, changed into bidding system.
@fiew14 Thank you for your support, it helps me to work further on my models!👍
@Ikena Could u re-upload the model to solve the problem? Or just the setting problem?
@Ikena by the way, I gave you a 5000 buzz tips . Plz keep going
@Ikena Glad now I can generate with your Checkpoint again. +1 Like to my favorite model 👍
What happened to the model? It's gone.... I cant make my characters anymore....
The creator did nothing wrong. Suddenly some of LoRa and even Checkpoints I discovered changed into Bidding system. I don't know how it really works but, you need to donate yellow buzz to the creator to make the Checkpoint/LoRa become generate-able. I reccomend use another checkpoint alternatives such as Pony Diffusion or AutismMix. If you want to create a Hassaku XL-like quality, I reccomend Hoseki LustrousMix.
@keanuyoga704 the yellow buzz don't goes to me, the bidding system is only there to decide, what models are on the generator. But since its now on the 3rd place, its sure next week available
@Ikena I hope so. I can't wait your Hassaku XL (Pony) can be generated again. Since then, I used the alternative which is Hoseki LustrousMix. But, suddenly, the civit.ai generator unstable again which is I can't generate anything due to the server problem.
When will the model be re-enabled?
@Ikena ok
Is the model going to be re-enabled soon for people to freely create using it again?
yes, in about 2d, you see that info in the auction
@Ikena Thanks. My pics are only good with this model, to me, there’s no suitable replacement.
There’s a message saying the model is not enabled for generation. Is something still wrong?
I’m trying to make images with the model again, but an error message is showing up saying it is not enabled for generation. Is something still wrong?
Wait until tomorrow to see if it leaves.
What happened to the Illustrious model?
duplicate detection system of civit unpublished it, not my fault
When will it be fixed?
?
@Ikena In other words, when would we let you use the model because when you use it you don't let us
@Lobo1992 ? i still don't understand that sentence. Its not me but the bidding system, that puts it up on the generator. The model itself is downloadable for everyone to use.
@Ikena So that you understand me, I am using a model where it cannot be generated well
This is one of the best model i've ever seen
I've been getting pretty good results on the IL version, but I've never tried this one. Is there a significant difference besides the prompt syntax?
I'm not sure, I've been using this checkpoint extensively and get good results but the IL version always comes out looking overstaturated/washed out. That could just be user error, a difference in the underlying models, or a difference in the versions. If you do local generations, I'd recommend trying the same general prompt with both versions and seeing which one you like more.
He told me this a day ago) The following resources are experiencing a high generation failure rate, possibly due to temporary generator instability. This usually resolves with time and does not require action from you. This notice will be removed once performance stabilizes.
Hassaku XL (Pony) - v1.3(better eyes version)
why it is very slow
pretty good when it's loaded
really solid results as baseline
Is it disabled again?
al parecer
It would seem so
josue200555362 en 7 días se deshabilitar osea va a estar activo
In 7 days it will be disabled, meaning it will be active.
Lobo1992 vavava gracias
It’s down again. Is it that thing of wait for seven days until it is enabled again?
Great model
Will you STOP making it unavailable for use every few times?! It’s annoying.
The model will be back online next week, or will it be down for longer?
I don't deciding that, its the bidding feature
@Ikena Got it, thanks!
why it cannot be used?
Details
Files
hassakuXLPony_v13BetterEyesVersion.safetensors
Mirrors
hassakuXLPony_v13BetterEyesVersion.safetensors
hassakuXLPony_v13BetterEyesVersion.safetensors
hassakuXLPony_v13BetterEyesVersion.safetensors
hassakuXLHentai_v13BetterEyesVersion.safetensors
hassakuXLPony_v13BetterEyesVersion.safetensors
hassakuXLPony_v13BetterEyesVersion.safetensors
hassakuXLHentai_v13BetterEyesVersion.safetensors
hassakuXLHentai_v13BetterEyesVersion.safetensors
hassakuXLPony_v13BetterEyesVersion.safetensors
hassakuXLPony_v13BetterEyesVersion.safetensors
hassakuXLHentai_v13BetterEyesVersion.safetensors
hassakuXLHentai_v13BetterEyesVersion.safetensors
hassakuXLHentai_v13BetterEyesVersion.safetensors
hassakuXLHentai_v13BetterEyesVersion.safetensors
hassakuXLPony_v13BetterEyesVersion.safetensors
hassakuXLPony_v13BetterEyesVersion.safetensors
hassakuXLHentai_v13BetterEyesVersion.safetensors
hassakuXLHentai_v13BetterEyesVersion.safetensors
hassakuXLPony_v13BetterEyesVersion.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.