Hassaku aims to be a model with a bright, clear anime style. Model focus are nsfw images, but also with a high emphasis for good looking sfw images as well. My Discord for everything related to anime models and art.
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 mostly danbooru tags. No extra vae needed. For better promting on it, use this LINK or LINK. But instead of {}, use (), stable-diffusion-webui use (). Use "masterpiece" and "best quality" in positive, "worst quality" and "low quality" in negative.
My negative ones are: (low quality, worst quality:1.4) with extra monochrome, signature, text or logo when needed.
Use a clip skip 1 or 2. Clip 2 is better for private parts, img2img and prompt following. Clip 1 is visually better, because i assume, the model has more time and freedom there. I use clip2.
Don't use face restore and underscores _, type red eyes and not red_eyes.
Don't go to really high resolutions. Every model, like hassaku, get lost in the vastness of big images and has a much higher chance to greate, as example, a second anus.
_____________________________________________________
Every LoRA that is build to function on anyV3 or orangeMixes, works on hassaku too. Some can be found here, here or on civit by lottalewds, Trauter, Your_computer, ekune or lykon.
_____________________________________________________
Black result fix (vae bug in web ui): Use --no-half-vae in your command line arguments
I use a Eta noise seed delta of 31337 or 0, with a clip skip of 2 for the example images. Model quality mostly proved with sampler DDIM and DPM++ SDE Karras. I love DDIM the most (because it is the fastest).
Description
The determination of what is better/worse in compare to my other model grapefruit is more difficult, because a model can be used in varios ways and training have more impact overall, i can't test/predict everything. Some feedback would be nice.
What is worse:
- On some occasion, hands again
- Images have sometimes a more dark and muddy look to it (see what is better)
- Faces have sometimes not the details of the body and looking a bit detached
What is better/worse:
- Less motion lines
- It tents in its style on some images to be more semi-realistic
What is better:
- Better quality of images before highres (not on all, but for the majority)
- Better contrasts -> stronger lights and shadows (but it can create dark and muddy images).
- Better skin textures
- Some backgrounds like mountains looking better
- On some images, backgrounds are better structured
Details:
Training with a small dataset (800 images).
Vae included in the model.
You can support me on my patreon, where you can get other models of me and early access to hassaku versions.
FAQ
Comments (279)
Can't recreate images from examples. The hash of the model from the examples does not match the one uploaded to the site. [9ebf46e9] in the examples, but [f604a4f9] in the model file.
I use the safetensor version of my model. I'm at work, so i cant see, if it has another hash.
Images are upscaled in img2img in steps, with the selection of the best pictures in batches, with a denoising strength of 0,4 to 0,7.
Prompts used in my examples:
machine one
(masterpiece:1.1), (best quality:1.1), highress, perfect anatomy, detailed, nude, 1girl, solo, [dark skin], gold eyes, long hair, black hair, animal ears, spread legs, medium breasts, ((restrained on machinery)), arms up, huge dildo vaginal, clitoris, nipples, [[stomach bulge]], pussy fluid, [[realistic]],
fisting:
(masterpiece:1.1), (best quality:1.1), highress, perfect anatomy, nude, long curly hair, (medium breasts), black hair, gold eyes, animal ears, spread legs, pussy, clitoris, nipples, bedroom, sitting, maid, detailed, [[[[stomach bulge]]]], [dark-skin], ((fisting)), arm in pussy, 2girls, [[realistic]]
Those images are further changed in img2img, originals has prompt "loli" in it and was to young looking for the site staff. Original: https://mega.nz/file/Gh0R1JQJ#ERqf1QdxmDSqw5bdZ40sGSjk8uXzEGHKvB61MIKAYdg
Thanks for the answer. I will be grateful if you upload safetensors version too. :)
Please provide some prompt examples. I list of effective terms would help too. Thanks!
Model use NAI models, so use the typical nai quality promts, that everyone use.
I use in positive: (masterpiece:1.1), (best quality:1.1), highress, perfect anatomy, detailed
Negative: bad anatomy, bad hands, text, worst quality, signature, watermark, username, artist name, blurry, bad proportions, multiple legs, multiple arms
then add in positive danbooru tags that you want.
Got it somewhat work
Is there a pruned version of the model available?
Nope, is there are easy way to prune a model? I don’t want to have a doctor title in ai to do it…
https://github.com/saftle/stable-diffusion-prune easy to prune with this (ema+fp16)
Whats the .vae to use with the model i dont see it in the dropdown but maybe im just blind xD
Click on the down arrow next to the download button.
thanks!
also rename it, if both are not named the same. This site renames the download files
How do I convert the SafeTensor file into ckpt?
i create ckpt and safetensor models simply with the merge tab in webui
Just drop this file into the folder with the models, it will work without problems
nmkd local sd has a converter
Can this do cunnilingus? To my knowledge there's 0 models out right now that can do that consistently through a prompt
It is a model merged with mostly nai based models. If NAI can't do it, my can't do it also
What negative prompts can I use to prevent the chin from becoming round, kinda hate how it looks
there isn't one that i know, generate images until you get one you like, with prompts that usually don't generate younger looking girls.
i dont get how to use the vae.pt either. automatic1111 just doesnt recognize it since it doesnt show up in the checkpoint merger. ive named them grapefruitV2.safetensors and grapefruitV2.vae.pt
thats all to do, if you use grapefruidV2, it use the vae with it. Vaes dont show up on the merge tab. If you select grape, you can see that in the bash window, that it loads with it.
hellow,have you solved this problem?
@ikena how you train this models they are to succesfull ?
you input the vea in the settings :)
Has anyone figured out how to define the size of a penetration?
You can add huge, but a definition of size is unknown to me
Can anyone explain a little bit about the installation procress?
If you have installed the auto111 web ui, download the files and rename them the same. Copy the files to models/stable-diffusion. Done, web ui should have the model in the model selection
Who knows how to insert all penises.🥲
hellow，how can i use file .pt? I can not find this model on chrome...................
.pt is the vae file, put it alongside the savetensor file, under your models. both must be named the same
If only someone would build a model for real life gapes, fistings, and large insertions :P
I download and all works, but can you please explain how to use VAE file. I rename it and put it in embbedings, but nothing work
put VAE in same folder as model, go to Settings -> Stable Diffusion -> SD VAE choose grapefruit.vae.pt. Apply settings
thx)
could anyone maybe provide link for vae? thx :)
Can anyone tell me how to best specify an artist style? danbooru tags don't have a huge amount of artists in them.
You can't really for the reason you just said. The model is trained with Danbooru's tag, so it never learned artists.
nifty model, but how did you get your example images to look so sharp?
Don’t go below a Denoise strength of 5 on highress fix, with a upscaler like latent. You can click on the pictures (full view) und download them, you can see all parameters with the png tab.
for the upscale are you just using img2img? or are you doing it through the "extras" tab? Also is the name of the upscaler you're using called latenet? or is that some setting?
I make a direct txt2image upscale with the highres fix (optional flag in txt2image). To show the capabilities of my model. There you can find the upscale options. Those options are added like 2-3 weeks ago in autos web-ui.
now don't i feel silly, didn't even think to update my web-ui. Thanks mate working now.
Thank you Ikena - is there a chance you could upload the full un-pruned ~7 gig model for fine tuning? Thank you!
Thanks, amazing model, getting fantastic and consistent results , merged it with Protogen 2.2 for more realistic less anime style characters, works beautifully , just amazing. The danbooru tags work really well
the artsyle from 2.2 looks VERY different than the one in 2.3,
other than that, the anatomy is MUCH better than 2.2, and overall looks great!
(edit:) after more testing i found that the less you add to the negative prompt, the more it looks like 2.2
v23 model, through the dynamic prompts plug-in ran almost a thousand times XY script to generate the results of random prompts.
In clip1, the probability of anatomy error is relatively high, it is the most likely to lose a leg or a hand, but the composition is the richest, with a lot of "non-girl content", but often with "pixel blocks" for which I am not sure why ( The background bokeh part looks like pixel art)
In clip2, compared to clip1, the girl's fineness is greatly improved and there are very few "pixel block" problems.
In clip3, I think the quality is the worst, very little background detail, character limbs are easily mistaken for the background and "melted", and there are often "pixel blocks"
In clip4, the performance is exceptional, the background is close to clip1 and sometimes even better, the character is complete and detailed, and the overall texture is clean, but too clean and lost in the mod I personally liked "soft glow skin". Hardly ever see any "pixel blocks"
In clip5, it is common to see the wrong objects interlaced
In clip6, limbs are often lost, but the composition is often very different from 1-5, and occasionally very nice compositions appear
Overall, I personally recommend using 1,2,4, and I basically use 4 now
After testing, when generating often use xy script to generate the results of clip1-4, after dozens of generation I found that 70% of the content worth preserving are clip4, 20% are clip2, a few are from clip1, no clip3
Wow nice! I personally only tested clip 1 and 2 and came to the conclusion to use 2. I can’t run the script xy many times, because my pc is good. But not „i run it thousand times“ good. I look up what 4 offers, thanks!
how do i use clip skip? thanks in advance
thank you for this feedback !
@asdpro123
Setting--User Interface--quick setting list--sd_model_checkpoint
add ",sd_vae,CLIP_stop_at_last_layers" after sd_model_checkpoint
save and restart
You can find sd_vae and CLIP_stop_at_last_layers on the top
Is there a CKPT version of this model? I am using DiffusionBee on Mac. Thanks.
I know if you rename the vae to the model and place it next to it, it will work. However, does anyone know how to make it work when I put it in the vae folder?
Automatic1111: Settings -> User Interface -> Quicksettings list add "sd_vae"
mine looks like this: sd_model_checkpoint, inpainting_mask_weight, sd_vae
after restarting the ui - on top of the window is now a option to choose the vae
i put them here:
\stable-diffusion-webui\models\VAE
@swat You are a G! Thanks!
@Swat Thank you!!! ♥♥♥
hehehe this is good man
Is the AOM-2nutmegmix = 33% grapefruit, 33% AOM2-Hard and 33% 7th_anime_v2_A?
I couldn't find any information about it anywhere and was really interested in trying them all out for my uh..... highly involved research.
Nutmeg is a mix of many models:
https://rentry.org/hdgrecipes#nutmegmix-aa3e502b
It existed before grapeV1. The version in my compare images is a AbyssOrange2 + 25% nutmeg.
There is a torrent out there, with all versions. See on 4chan -> hentai -> stable diffusion threads.
@ikena Just wanted to say thank you! I ended up having to make a different account and now this probably looks confusing, but thank you so much! I appreciate it and I love the model!
Why does it produce weird outputs, not anime ones? i have the safetensor and vae in the same folder with the same name.
the outputs look like comics and very weird almost also a bit abstract , nowhere near the anime or samples im seeing in the comments section.
Just saw that you put restrictions on your model. Like people have to credit you when they use the model, people can't sell the images they create etc. ;
While I get the hosting services and to not sell your model.
I was wondering why you want to claim practically every output that anyone does with your model?
Crediting me for using is dumb, i changed that. I thought „use“ means merging without crediting me. I don’t claim every output, that would be ridiculous.
@ikena I appreciate that you answered and really appreciate the work that you put in your model, but sadly I don't think I'm going to use it. While the crediting restriction is gone, not having the full right to do with my output what I want including selling it, is just too much of a restriction. I normally put many hours into refining and changing the output and after that not having the full copyright to the pictures is just not very useful. ;
Still appreciate your work though.
@VyneNave it's unenforceable anyway, just ignore it
@vynenave926 cringe. sd is made on copyright cementry. the ones generated from it are the same
I tried to generate "futa" images several times. Unfortunately, I failed. The new version v3 supports "futa" pictures? Do I need to use specific words?
Doesn't quite answer your question, but you might wanna give https://civitai.com/models/4109/futagen a shot for this kind of stuff
I think, the best is to use a hypernetwork or a lora for it. One is here on cevit: https://civitai.com/models/4076/futa-frontal-hypernetwork if you want to use my model.
Lora or this tips in your prompt (it's a "replace by") :
[(vaginal):penis:0.1]+
Hi! Love your work, it's simply amazing.
So I've been recently experimenting with mixes and merges, and I tried my hand at adapting the already existing stuff built on Anything v3 (basically everything at this point) to Anything v4, a different model which I suspect is wildly different to v3 when it comes to weights, because every mix I try to make using v4 comes out deformed and nowhere near the polish of their v3 counterparts - botched anatomy, far too much realism, the works. Even the composition differs a lot. My only successful model was when I MBW'd v4 and Basil after about five or so fiddlings with weights, and since then I was unable to add any other model to it, since any merge basically wrecks it.
Do you have any advice on how to make v4 viable? I'm not sure if converting v3 models to v4 by adding difference (v4 - v3) with M=1 to existing v3 models is a viable strategy, but I could try that and see if it sticks. Thanks in advance!
AnythingV4 is a scam of some guy, it is AnythingV3 with one of the AbyssOrange mixed in it. Every mix with AnythingV4, with a model that don't have already basil or similar in it, results in the addition of realism in it. Like you said, results are then far into realism or get butchered. Because the mix simply mixed more realism in it.
AbyssOrange2 has this in it:
AnythingV3.0 - BasilMix
1,0.9,0.7,0.5,0.3,0.1,1,1,1,1,1,1,0,0,0,0,0,0,0,0.1,0.3,0.5,0.7,0.9,1
Fake AnythingV4 has then a similar but weaker distribution with basil in it. If i asume, its AbyssOrange2 mixed in, it can also be AbyssOrange.
Personaly i say, mix only with models without UnetMerges.
Because all layers are clean and have the same style.
I mixed lemon with a weaker distribution to basil:
anyGape - BasilMix
1,0.9,0.7,0.5,0.3,0.1,0.5,0.5,0.5,0.5,0.5,0.5,0,0,0,0,0,0,0,0.05,0.15,0.25,0.35,0.45,0.5
Mixed then with grapefruit, simply to get grapefruit less realistic on those layers.
Ahaha, it all makes sense now. Interesting. So the addition of Orange screwed the layers, and since Orange is built using Basil, by adding more I was just piling on the realism and breaking the model. You learn something new every day.
Well, back to the drawing board then. Thanks for the response, I really appreciate it!
Has anyone figured a prompt that removes facial shadowing? Seems common for all of these animé models and it makes the image hard to work with outside of just a proof of concept. Tried things like "direct lighting" etc but nothing seems to work
same for my case
I used anything v3 before, and the result always blow me up, but when I use grapefruit, with same prompt, same setting, the picture never seem to be colorful and the light effect doesn't seem very good, althought this model has some of anything v3 in it
Negative: Shadow on the face
@limitless177 this has never worked for me. Can you make a review here that shows a before and after?
I made a random prompt xy plot of 200 times clip1-6 for the V3 version
Setup:
DIMM
25 steps
CFG 11
512*512
no hires
Clip1: good quality in most cases, bloated limbs in a few cases
Clip2: some problems with the eyes, distorted if away from the camera, but occasionally the best quality image with the same seed, overall good composition and lighting performance, but some problems with details (eyes or fingers)
Clip3: Still the option that is neither too good nor too bad
Clip4: still seems to be the best quality in general, but the prompt is not as accurate as clip1
Clip5: The image texture is closer to Anything than the glossy texture of this model
Clip6: The glossiness is further reduced, sometimes the composition is very different from 1-5, and it is the worst with prompt
I would recommend clip1 or clip4
This is only my personal opinion, I uploaded it to googledrive, please download it and compare yourself
Next I will try other samplers for quality comparison and turn on hires 2x latent to see if the details of clip2 will improve
BTW, does anyone know how to share and view a 20M or even 200M size image online?
20 times clip 1-4 with hires latent 1.5x
Hires doesn't seem to help with clip2's eye error
Uploaded a sampler test in googledrive and its clip2(I want see different about eyes and it does), hard to say which one is good or bad, seem like other sampler draw eyes better than DIMM(See the seed 30083575 and 41539608 is most notable)
But DIMM has the richest composition and the least amount of cropping and the best performance in its class with Eular A/Eular , so I guess it might be a little hard to choose
Heun seems to be the best quality, but it is known to be extremely slow
DPM++ 2M seems to be a moderate choice, with moderate speed and rich composition
DPM++ 2S a Karras I personally like some of its compositions, but that might just be the lucky shot
https://drive.google.com/drive/folders/1oqQoJzo59hz8_D_CyCK70qqb1KxTfnFH?usp=sharing
Wow, again, many thanks!!
I wish i had 4 3090, i have one measly peasant 12gb 3060 at my disposal. I looked through your grid and yes 1 and 4 seems the best. Pains me that many pictures are cropped through. Before your test, I already shifted more to 1 then 2, 3 of my new examples are out of clip 1 because of it. Hope that V3 is a little better then V2. I admit, my goal was simply, to get V2 to a more anime look, lemon was the result of it and I mixed whopping 50% with it. Now my plan is to refine V3 like I did with V2. But it may take more time, because I have less time for merging now. Really appreciate your effort to test it out.
I think, i make my quality checks now on clip 1 then on 2.
@ikena
I always forget to say this thing: what configuration of machine do you need to help you with your work please?
I have a unused i3-2100/16GB ram/120Gssd (can add, have a few unused free ssd)/3090 machine, you can use it with sunlogin or teamviewer if you need I guess
It's sitting in storage on the other side because I was going to use it as a free stable diffusion bot on my discord channel, but not many people seem to use it, so it's basically sitting completely unused
Please let me know if it helps you with your work
@omnisaa nah, i'm good with my pc. I test clips now on more explizit images and clip 2 has less problems to make good looking private parts. Test it further out now.
This is one of the best, but is there any way to remove the gape aspect from the outputs?, I tried all kinds of negative prompts for gaping but in the end I always get a wide pussy & asshole.
that's weird, because I always get normal ass & pussy even thought I added gaping to my prompt
Avoid the word pussy in your prompts, you can use groin and nsfw as alternative, if the pussy is to undefined. I play with the idea to create a tamer model, in there you can prompt pussy and get not a gaping thing down there.
noticed that as well - my friends who I've been generating the explicit ones for, say that a lot of the time pussy is so overblown it resembles a snail lol :D
@ikena I'd like that tamer model, I love using this but the gape is too much
You have so many models. It would be cool to have a x/y plot comparison between them to better see the differences.
It's quite simply. Grape is old, use grapefruit V2.4, V3 or lemon. V2.4 is more realistic then V3 and lemon is near similar to V3 with more anime in it. Every V2 versions are similar with detail improvements on each step.
Edit: I removed grape and the older Versions, less cluttered.
Edit2: Lul i doesn't know it delete also all reviews related to them. I'm now a bit sad about it.
@ikena Oh this is unfortunate. Don't be sad you will have new reviews soon.
Sorry for all that reviewed my model and got deleted.
I removed grape and the older Versions, to get it less cluttered. Was not aware, that it also deletes all reviews related to them. I'm now a bit sad about it. But I can't do anything about it.
Absolutely loving this model! My biggest problem however, are the eyes. For some reason I can't seem to generate anything where the eyes look even slightly cohesive, despite tags like "Beautiful eyes, detailed eyes" etc. Do you have any more tips on how to make faces actually representable and anything like the ones in your examples? ._.
I’m working on it, mixed already models that have better eyes. But everything around gets worse, so that was not a option. One attempt with AniReal was promising, but backgrounds got worse.
You can also try clip1 out, clip1 is better with eyes but worse with nsfw things
@ikena I .. apologize for sounding very lame /noobish right now but where can I find these "clip1 out" or clip1 options? ._.
@Sreliata don't apologize, it is a valid question. But if you don't know of this option, then you are likely already on clip1. Tab Settings -> row Stable diffusion -> last slider named "Clip skip" on that mask.
@Ikena Ah! Thank you so much for your response, Ikena. I will try that asap then! ^.^
@Sreliata If you need further help, you can add my Discord - Sato#9999 - and I'll help you out/give you an invite to a Discord server you can ask a community more questions in realtime ^^
Hey just a quick question. Did you give mage.space permission to use your model on their hosting service? Because I'm quite sure there is paywall to access your model/You need to be a pro member to have access and you clearly stated in the restrictions that you don't want your model used for services that create money by generating images/hosting services.
So I'm just curious if they actually asked you and you gabe them permission or if they just decided to ignore your restrictions?
Nope, they simply ignored my restrictions. I contacted them.
I gave mage my permission to use my model on a paid service. It doesn’t make any sense to have a restriction to sell images out of my model anymore. Because images out of mage are free to sell. So i lifted that restriction.
I have a question, i feel like the quality of my images is not to good, i copied the prompts of other images from here and i get the sameimage but with a low quality, someone have a idea of what happend?
Pd. Am using google colab
Thanks
Many images have a added high resolution highres fix after the first generation.
I think that the newer versions of Automatic1111 have some issue. Not tried this model yet but with my own and others, every update of Automatic1111 changes the image result
Try changing your CLIP SIZE in settings from 1 to 2. It doesn't show in the examples unless you copy/paste. Overall using clip size = 2 works much better on this model for me
Also consider changing the ENSD
I tried to reproduce an image from the samples and it looks a bit wrost (not too bad in general, but wrost) and all my generations looks very opaque in comparison with the samples.
Don't forget to use a vae with it
@Ikena sorry for the ignorance, but could you explain how to do it?
Try changing your CLIP SIZE in settings from 1 to 2. It doesn't show in the examples unless you copy/paste. Overall using clip size = 2 works much better on this model for me
@Ikena oh, I forget to setting up the vae... now it looks better in the colors! =D, but with an extra hand in front xP
Should I switch to a new version every time, cuz it's always better than previous one?
Im not sure but with 3.2beta, sometimes i got black image, other models still work fine. Need someone confirm...
im not getting the same results with the model 3.2 beta as the images showned. i tried with diferent clip skip and eta noice 0 and 31337. Do you know what can be happening?
Click on the images, download the image and put it into the inspection tab. Click copy to txt2img and generate the image there. Is it then the same? With Clip skip 2 and noice 31337.
i made a review with the picture because here i cant upload sry im new to this(got it much closer now because i forgot to put width at 448). Btw I don't what is the inspection tab on stable diffussion web ui.
@31sivilao the "PNG info" tab
@Ikena just wanted to give you a heads up, you can still manually set a different VAE in the UI. All models have VAEs included, just unless someone bakes in a working one, it is usually an incorrect or distorted one.
Run it with --no-half-vae
My latent upscalers do now weird block artifacts and white and black grids. Don‘t know if it was the update to windows11, torch update or a update of autos1111 webui. Can someone confirm it or is it only me? latent was my favorite upscaler, also for testing new versions
I have always found latent upscales do the weird blocky pixel thing your talking about, unless, you set a very high denoising strength (usually 0.4 - 0.8 depending on resolution and upscaling scale). I would recommend if you want to keep using latent upscale to go with around 512x768, upscale by 1.4 - 1.8 and set the denoising higher. Then take it to img2img and use ultimate upscale, maybe 3x - 4x here. If you want to take it another step further use extras to upscale again maybe using AnimeSharp4x / R-ESGRAN 4+ OR use a tool like topaz gigapixel. If you look at my most recent review you will see a couple of the ones I posted used latent-nearest go check out the exact settings I used. Hope this helps and keep up the good work :)
EDIT: Look at the second one I uploaded for latent-nearest upscale, it was from 2 days ago
@odawgthat Thanks for the answer. I know of those artifacs now, because i ran my example images again and i have now those grids and weird blocky artifacts in it. I have those not before, my examples are my proof.
@Ikena tex2img with highres fix latent works fine for me with the current a1111 version (on windows 11, upgraded from 10, 2weeks ago)
higher difference between the base resolution and the highres resolution -> needs higher denoising to compensate this
@tawsLoRa Yep
@Ikena Maybe something happened to it in an update?
Hi, appreciate your work.
Want to ask a question, can I use your model to generate images and sell the images I generate?
For example: on patreon or video games?
Short, yes. You can also see that here on civit.
Would it be possible to have a pickeltensor version of this model?
I make a inpaint and a pickeltensor version in the next update, but it needs time. The removel of the cropping issues also change the model to the worse, so i must work further on it.
I love the Blurred image of the NSFW vending machine xD
How can I get high-quality pictures?
I'm using colab and I read some of your comments, but I couldn't understand all.
I really want to know a solution with step by step.
Please help me
i´m using webui and somehow in around 50-80% of the cases my picture has either multiple girls(with 1girl) or i have additional limbs or strange positions.....someone has any idea?
Use prompts be like "1girl, solo" and etc.
Negative promts: (worst quality, low quality:1.4), bad anatomy or something else.
Just check the other AI work's and combinate for what u want.
@Apollynaria tried, still 30 of 50 pictures are BS. Multiple legs, wrong legs, second person with wrong body, etc. Sometimes even black pictures, sethjng about an error with VDE(i'm using none). Had to add no-half-VAE to args.
Are there prompts which are known to add BS?
Should i try to use the 3.1vae?
I have clothes, expressions, positions etc in the prompts. Could that be the problem? Or maybe has the engine a memory problem? Atm i generate the pictures with a nvidia 960M(2GB), had to fiddle a bit around to get it to work though
@Xeltosh Is it only my model or other models also? My model have the same level of BS as every other anime model out there. But 50-80% seems to much. I'm also surprised that 2gb works at all. Look at the examples or review images, do you get the same image out?
3.2 has a vae already included.
@Ikena will try to get the same images as yours after work. Dunno why i havent thought about testing that 🤣 other models produce also bs sometimes but not in that high of a rate.
2GB works because i made it work with a registry-entry and the arg --lowvram
As long as i dont generate a picture that is too big(a 768x1024 pixel picture needs around 1GB of gpu-ram), there are no problems. Sometimes the cmd of the webui "freezes" but you only have to press enter inside it to let it resume working. I have a clicker for that active, that presses enter every 15 second for me.
The only problem here is the 2-second-timeout the OS gives the GPU when calculating stuff(that GPU is slow af for this type of work, at 25 steps i need around 6 minutes for a 768x1024 picture). The command to calculate a picture to the cmd can sometimes take longer then 2 seconds to get processed and the OS terminates the process of the webui because it doesnt get an answer in time and the webui crashes as a result of that. A registry entry can disable this OS-query but it is not recommended from the manufacturer. Found it in a GPU debug forum for testing purposes. Its my old laptop, so i dont care if the OS is slow or running instable. It is my only nvidia system though and amd-gpus dont support stable diffusion properly so far
resolution works
dont generate above 768 resolution. use upscale instead. I had the same issue
@Xeltosh wait wait wait...what?
you can generate a 786x1024 sized image with only 1gb vram?!
I can only generate a 512x512 image with 6gb vram...
@Sl1pp1 yes, i had to tweak a little around though.
First off, i had to deactivate the timeout-query in my windows-registry, every 2-4 pictures i got a time-out-error because my GPU took too long to answer back to my system and killed the process for SD, so i always had to restart webui to continue before implementing the registry-key.
Then i also use the argsuments:
--administrator --lowvram --always-batch-cond-uncond --disable-nan-check --disable-safe-unpickle --opt-split-attention --no-half --precision full (atm i try "--upcast-sampling" instead of "--no-half --precision full". so far it works out and is not really faster/slower
I upscale a 384x512 pictures by 2, up to 768x1024 with active hires.fix. it takes around 9-10 minutes per picture, but it works. I can even run a batch count of 100 (batch size 1) if i want to. takes forever and i have to let a auto-clicker running, which presses a button every 20 seconds to "unfreeze" my webui cmd, but it runs.
@Xeltosh thanks for the info man!
@Sl1pp1 yeah, but apparantly one of the last webui updates broke my configuration. Most of the pictures i get the out of memory again.... Still trying to find a solution
Hmmmm.., why remove the lemon model? 🤔
My patreon use it now for lemon tea
I have a question. I like the model, it's very versitale and can produce almost everything including hardcore NSFW, but 1-2 of 10 generations it tries to make a multilook, like comix character sheet, character + same character from other angles or with closeups. Like this - https://ibb.co/JnLrdcy I tried to pup all kinds of "multiple people" tokens, 2girl and all such stuff, even with emphasis, but it still generates some this way. Any suggestions?
„multiple views“ in negativ also doesn’t help? Sadly with 80k images on danbooru, it is a thing, anime models try to do sometimes, without intention. My model is no exception.
I understand that, just was curious if you encontered and countered it. :-) Any more tokens that are useful besides "multiple views"? I'm just playing right now with wild cards with outfits, places etc, some generation are fantastic. :-) And also trying to figure out what words work better here and there. So I make a big 50+ batch of generations with hires fix over the night then finetune the overall prompt and wild cirds itself. And about 10% show multiple views. Sometimes it starts to generate man behind girl or it tries to stick dick between tits. :-D Naughty naughty anime images. :-D When I will be done with simple wild cards, I'll try to make NSFW ones. Also tried to make different colors or textures of skin from simple real color tones to fantasy ones, managed to make 5-6 simple ones from pale to dark and colored ones. Could not manage to make animal or reptile skin. Just to clarify boundries of the model. But I'm still trying. :-)
@Ikena and it also alway tries to do that - https://i.imgur.com/NYfWFbL.png Even if I try to put paizuri and all possible "penis between breast" tokens to negative even with emphasis 1.5
@drakmour Ah sorry for that xD my model aims to be a hentai model, if you try to do more sfw stuff, it try's anyway to be naughty. Don’t know which Version you are using, but V3 throws also with sfw in positiv and nsfw in negativ shamelessly a dildo in the middle of the room. V4 is in that aspect better.
@Ikena I use last one 4-th. And about my recent post about paizuri, I guess it's a "default" pose if you don't put any pose in the prompt, it puts it in paizuri. :-D It doesn't heavily tries to do NSFW stuff, I get mostly what I prompt. Just happens sometimes.
@drakmour whats the resolution you were generating images at? I had the same issue with 768x1024 resolution. Reducing it to 512x768 instantly fixed the issue. I suggest you try it.
@AlexSpirt this worked for me, thank you!
The best ingredients often require simple cooking!
This model can work amazingly with simple promopts, which doesn't mean it cannot work even more amazing with complex prompts.
what a bullshit language
perfect model thank you
wanted to test the difference in realism between 3.2 and 4.0 so I did a few generations with pretty much the exact same setting, just alternating between models, samplers and denoisers: (downscaled the images by 2 because I wasn't sure how big the limit is on imgur).
tested with a fairly low detail prompt:
https://imgur.com/a/mVusIAL
and one with more detail:
https://imgur.com/a/5ss2VJ0
Overall I do think they are about equal when guided by celeb names to reference a realistic face,
From this generation I do think that v4 is on average a bit better at proportioning more realistic faces when not guided.
In term of background they're equal and the background doesn't really benefit from things like 'realistic background'.
added two images made with more weight to green hair and 'photorealism' and 'ultra detailed' to the high detail prompt and it's not much more realistic.
(it's annoying that we can't link images in comments)
Just out of curiousity, but i see a hypernetwork related to hands in there... you dont happen to have a link do you
@paranoianl824 I found it on one of Corneo's comment under one of his TI(after checking it was on the widowmaker one)
It's not perfect but I do think it reduces the overall amount of times I get bad hands(they still happen tho)
Here's the file:
(I put it at 0.75 strength since that was his own setting, but feel free to try at other values)
https://drive.google.com/file/d/1fToRYs_ssc36DLlVFMsYA_tRz4weDtgH/view?usp=share_link
If you're trying to get a view from behind, "ass" will often give you absolute dump trucks, and "from behind" will often be too strict; try "looking back" instead. And you should probably remove breast sizes from your prompt when doing it.
Was looking for this, thx
Any one use web ui from Stable Hode like https://aqualxx.github.io/stable-ui/ ? The results I get are so different with even a short prompt like "solo sitting on a chair". When it is suppose to be V4 of this model.
The UI shouldn't matter, I'd hazard a guess that you're not using the ubiquitous negative prompt "(worst quality, low quality:1.4)"
How can i get this model to work better? or is it my stable diffusion version? i get very good pics but not as detailed as the ones on this comment section or in the post's summary.
I would use the newest version 4.
with and height at least 700x, DPM++ 2M Karras as sampling method, sampling steps 20-30, (worst quality, low quality:1.4) in the negative prompt, (masterpiece, best quality) in the positive prompt. if you get a good one send it to img2img and generate at like 0.7 denoising strenght, and DDIM as sampling method. Then send to extras and then upscale with the ESRGAN_x4. that is what i do at least
@neyney It is better to run SD_Upscale script in the img2img with a low noise on like 0.2-0.35 instead of sending it off to extras.
@DenpaDenpa Now that i got deeper into it i agree with you
Can't quite get the images right, could it be the eta for DDIM?
use higher resolution at least like 700+
Does the v4.1 have a vae baked inside ? If so, can you add a not baked vae version ?
You can use another vae with it, it override the baked in vae
@Ikena Ok, but since there is some debate about the bad interference of the baked vae models with LORAs (I don't know for sure) I prefer the not baked vae models. Just asking, if you can't, no problem. Great blend anyway.
how come the newest update has almost 2gb less ?
Model pruning. It removes weights that aren't necessary. If you look around you can find some models that have both a pruned and non-pruned version. That would let you test to see how minor the difference actually is if you're curious
Pruning a model takes it from 7 GB to 4 GB. The real answer in this case is fp16. People used to save models in fp32, but with modern nvidia cards (16XX, 20XX, 30XX) all calculations are done in fp16 to make it much faster and so the bigger model makes no difference.
Not only pruned down from 7 to 4 but also removed ema data from it, that is useless for image generation. Only 2gb of absolut pure image generation weights.
@HollowStrawberry hi i am not expert but i thought nvidia adopts mixed approach? just being curious
thx
Another protip:
A big problem with doing multiple characters is you often get arms melting into eachother, try adding "holding hands" to fix this, if you can stomach the absolute DEGENERATE act of handholding.
(btw if you're doing a lesbian wedding scene, put "male" in the negative)
And a bonus, if you're doing trans girls, don't add futanari or newhalf to the prompt, just put "penis, erection" in. might need to (emphasize) erection.
full prompt:
standing, outdoors, 2girls, multiple girls, yuri, wedding, wife and wife, (red hair), medium breasts, wedding dress, elbow gloves, holding hands, eye contact, smile, masterpiece, best quality, highres
neg: male, braid, railing, loli, blurry, censored, border, multiple views, lowres, low quality, text, signature
(for some reason, "masterpiece" insists on doing a leaning against a railing pose with every novelai based model i've used, so i added "railing" to the neg)
The 4.1 inpainting model has the exact same filename and size as the main model. Was the same model uploaded twice by accident?
It is a different file. If you look at where the download button is, 5 rows down is "Hash". The files have different hash values.
filename is changed by civit, pruned models go down to somewhat the same filesize. You can see that on other models, they have 7gb, 4g or 2gb files. 4gb is pruned, 2gb is pruned + removel of weights that are not necessary for image generation.
@Ikena Interesting. I'll just have to rename it after downloading then.
What is the difference between the inpainting model and normal one 4.1?
inpainting is for img2img -> inpainting. SD is switched SD 1.5 to SD 1.5-inpainting: https://huggingface.co/runwayml/stable-diffusion-inpainting
Lately I get constant error messages when using this model, only with this one...
Could anyone help the difference of inpainting version, does it bring better results?
Hello, the model looks great, I already downloaded it, I'm dying to try it, to use it in google colab it is used as if it were a .pt file, right?
just follow these tutorial for colab - https://saashed.com/how-to-run-stable-diffusion-for-free-on-google-colab-in-less-than-10-minutes and this one to download civitai models to colab https://saashed.com/how-to-easily-download-use-custom-stable-diffusion-models-from-civitai-in-google-colab/
This looks awesome! Although the shading feels too "shiny" if that makes sense, i wonder is it possible to produce more anime looking images with less bright shading using this model with the right prompts?
Play a little bit, reduce (worst quality, low quality:1.4), to like (worst quality, low quality:1.2). Put realistic into negativ. Model can do more anime looking images.
I didn't quite get the worfklow of this model + the inpainting version.
are you supposed to use the same one in both times? or the inpainting is done from a different starting image? can you include a quick example of your workflow that you used to generate one of these images? (given that they used inpainting) Thanks!
Do inpainting model has the exact same name as normal one?
Hello author, first of all thank you for the model you provided :)
Then I want to ask, in the version 3.2, 4, 4.1, you said in the introduction that version 4 has a lot less nsfw than 3.2, and then version 4.1 has more nsfw than 4, so version 4.1 compares with 3.2 in terms of nsfw Would it be better?
Sorry for my trouble, because my hard drive is not big enough to add multiple versions of the model orz...... so can you recommend me the best version for nsfw? ww
Use V4.1, if you have a kink for gaping and don’t mind, that many images get cropped, use V3.2.
perfect
I hope I can finish my landscape batch before the next version hits. 😓
It seems many misunderstood that licence stuff. With my model, nothing change. You can use it like before, merge, sell images out of it and have simply fun with it. Only other image generator sites can't use my model anymore. That is the only thing that change. I have all rights to my model and it remains open source in that regards to the community. Simply enjoy grapefruit, i provide updates to my model and upload it like before to civit :)
"Fantasy.ai is the official and exclusive hosted AI generation platform that holds a commercial use license for Grapefruit, you can use their service at https://Fantasy.ai/. Please report any unauthorized other generator sites which are using Grapefruit without autorization."
We understand very well. One party is paying you so that it can market your model's name exclusively. And then you want us also to work for that party without being paid to report on unauthorized use, all while the original models used were made with the intent of being free. That's your prerogative to sell out, advertise, and try to make $ here, not anyone else's.
@TonyDPrime It is either every corporate can use my model for money making and i and the community get nothing from it (mostly without my knowledge) or only one corporate can use my model and i get some from it. I choose the latter and ensured that my model stays free to use here for everyone. I personaly don't want that someone work for me and i didn't choose that wording. I already spending earnings back to the community/modelcreators and tools like URPM, everydream, myne factory and civitai itself. The rest goes to equipment for my training endeavour. Like a new power supply for a bigger GPU later. Because my 12gb GPU is on the border of training capability.
@Ikena You do know. That you legally probably aren't able to either enforce that or even be able to change the license from which this all stems from in a way that clashes with the already stated terms. That's a slippery slope on an already legaly grey area of AI Art generation monetization. // Edit: Did you ever fully read the SD license? You would need to make your own full model .. which probably costs around 20k~200k dollars of gpu time to compute(based on the dataset). Before you can change the license to whatever you like. The way how to do it is openly viewable and doable by anyone. That's how most generation sites "should" work.
@Greenfox I can't enforce it, i can only request it, fantasy.ai propably can. To contrary believe, the SD licence is changeable if you don't change paragraph 4.a. of the licence that points to paragraph 5 Use-based restrictions, that say that all restrictions in Attachment A must be followed. But i'm not a lawyer and i can be false in that, but that is what Dreamlike Diffusion do.
@Ikena "4.a. - for use, reproduction, or Distribution of Your
modifications, or for any such Derivatives of the Model as a whole,
provided Your use, reproduction, and Distribution of the Model otherwise
complies with the conditions stated in this License."
@Ikena Yup, that's what i am pointing out. No lawyer myself, but i read enough patents and make them with patent lawyer types. So that sound alot like "Yeah you can add to your license, that you don't want people to mess around with your purely landscape model. As long as they only use it for landscape generation changes. But they still can sell or do whatever they want with anything related to the model otherwise or with the generation outputs. Like training a model to make nice landscape backgrounds, but add characters into the generation output etc."
@Greenfox ehm that is not 4a you know. Read the sentence before the text you copyed out of the licence
@Ikena I cropped that one wrongly. I just wanted the text after the 4.a. What i meant is, that right after the 4.a attachment hint for use cases. It says that you need "complies with the conditions stated in this License". Which basically says what i meant. You can change certain things, but you can't revoke parts of the License that has been aggreed on before. Like everything in there, except the Attachment 4.a to add stuff to it that doesn't conflict/clash with the License in itself. Given that it basically says you can't mess with the conditions of the License. Sorry for my badly worded example, you would add "- This model may only be used for landscapes, i gotten rid of everything else. It might be able to do other things if forced, but i don't want you to do that with my specific model." into 4.a. That's how i understand that license. Because you can't change coniditons like 2. " ... Subject to the terms and conditions of this License, each Contributor hereby grants to You ... " , which guarantee that its an open source file for anyone to use and meddle with. As long as they hold to everything else stated. So if someone comes along, takes my "Landscape Only Model" and mashes it into something like "Spaceships Rule". I could theoretically get them to take it out of their model or atleast point at them while doing an angry face. As it might be hard to track that, i got no clue there actually. You still can give fantasy.ai your newest model version a few weeks/ years / forever before takeing it public, nothing in the license says you need to provided your models to anyone. But once someone gets hold of it, they can do whatever they want. As long as it doesn't break the License. For example, i actually use this model together with another one in Automatic1111. To make nice anime-style monsters and animals. Aside from the spontanous anime-girl convertion that happens sometimes. It's the best output i have found for what i want to create. Also quite funny, when one seed gives you mostly what you search for and then there's some random seeds that add stuff you didnt ask for. // Just noteing. Thanks for the model. But the whole license stuff has a reason its open source one that makes you stick the same license to everything done with the SD baseline or any work based on that. Atleast that's my understanding of the text.
thanks for keeping it open source
Hello author, first thank you for the model, I downloaded and added the model to the Stable-diffusion folder, then I applied the model and used the same parameters but could not reproduce your example diagram, the generated picture is of very low quality, I don't know what went wrong, help
What prompts did you use? Seed Number? Did you change the selected checkpoint/model you use in Automatic 1111? That's probably the main problem most often. People forget to pick what model to use and just use the basic SD model xD.
why many peopel use SD1.5 instead of 2.0?
I think because you can train better on 1.5
2.0 is simply worse then 1.5. You can use 2.1, but only if you create complete new models.
Because it's much easier to train NSFW models off of the 1.5 base.
lemon is not available anymore?
It's exclusive to his patreon members now.
Which lemon model are you referring to?
@JosukeTOR understandable, have a great day
@es69 the sfw version of this model
@nolanaatama @es69 lemon is the complete opposite of sfw, it is a mix with max nsfw in mind.
@Ikena what? I thought it was the safest version.
what I remember was I love grapefruit artistic taste but don't want the nsfw stuff and read briefly that lemon was the safe version and keep in mind I'll try it later.
maybe I came from different dimension, love your model btw mate!
@nolanaatama In my experience, putting “nsfw” in the negative prompt makes very unlikely to get nsfw content unless you force it (and is more likely that you get an abomination). I use this model a lot for sci-fi character ideas and I don’t get nsfw images unless I really want to.
Really great model!
Why am i always stuck on 95-97% on loading picture? and the time just keep adding from ETA 1second to forever
did you agree with this?
No, I already send a take down request. But no response (completely normal phenomenon).
@Ikena lol
@artwork_dalcefo
Oh, what a delightful website! Pray tell, what is your noble purpose in gracing us with its promotion here? I daresay, the Civitai users who post are such clever chaps and lasses that they could easily create their own images with the data model. Why would they waste their precious time and money on a site that takes ages to produce a picture?
Ah, but you're such a kind soul to inform me of this website. I've gone ahead and reported your post, just to show my appreciation.
@chinpako what do you say? This is a notice to the creator about commercial use without consent. I think you've misunderstood something and are angry with me.
@chinpako that is not a promotion, it is to inform me about a commercial use without consent of me.
@artwork_dalcefo
I am so sorry.
I understood to my big mistake.
I hope you will forgive me.
I thought that you were indirectly promoting me.
Of course, I am very much against unauthorized commercial use of models.
@artwork_dalcefo
https://www.linkedin.com/company/yodayo/about/
It is surprising that they don't hide their whereabouts....
I'm going through a similar thing with my mix, fuck this site and their predatory practices.
crappy model, crappy ethics, crappy all around
poor you
?
Can anyone tell me how to get the "Hires upscale", "Hires steps", "Hires upscaler" parameters. What is needed for this and how to get it ?? thanks!
You need to check the Hires. fix box, it's located below the sampling method. See image for details.
https://ibb.co/8drKY3h
If you don't have the Hires. fix box, do this:
Update your Stable Diffusion. Install Git (https://gitforwindows.org), right-click in your stable-diffusion-webui folder, select "Git Bash Here", and type "git pull", without the quotation marks. Hit enter and you should be good to go.
@renari
Thank you for good advice!))
how to keep the SAME CLOTHES, with different posture? Thank you
請問用什麼方法可以保留同一件衣服, 但不同動作?
謝謝
use a lora
controlnet
controlnet can help you
Does this model do men too? I don't see any in the examples.
yes, my model do also man, see the second last image of V4
Such a great model, wow! Thanks a lot... but im having some problems with it:
While using automatik1111s highres fix option, it very often adds multiple "holes"? Is there any option/promt to reduce this?
Would you be against training a checkpoint on 1995 anime series "Neon Genesis Evangelion"? It's a very popular series and a similar style checkpoint has already been done for Studio Ghibli. Would be amazing to see it for 199s Evangelion. Thanks!
Sad to see this model getting repressed by civitai now, it was always in the top downloads and now its not even on the page. Them adding the filters fucked shit up, I got everything turned on and still can't see anything NSFW... even searching for this model doesn't work, I had to go through my comment history on my account to get to it.
click on setting, and mark all flags to see the content
I think there is a bug since the personal filters feature is in Beta. To quickly fix this, press on the eye icon in top right and select "Everything" as Browsing Mode. You should see everything now, including NSFW stuff
Can you change the name so the inpainting model don't have the same name as the normal one?
Are there people who make animated female characters?
I get less saturated results when I use this model with Automatic1111 versus other apps. How can I fix this?
@JanKowalski it is possible the other apps are automatically using a VAE, if you don't have one selected in automatic1111 the colors won't be saturated.
This has been happening to me too and I have no idea why. This should have the VAE baked in, and changing any VAE options in 1111 changes nothing.
@Riyu25 oh my bad I didn't realize you were using version with VAE baked in. I'm not sure what is happening.
Sometimes if you're specifying an artist/author, their works will be muted and no matter what you do they always will be. I tend to try not to define an author/artist, models can also have muted colors even with the VAE baked in.
Something I haven't tried is combining models, 60% with rich colors and 40% with muted colors. I'm here learning from all the great artists here by using their examples and then tweaking them. It helps me learn What does and doesn't go together.
This always still suffers from the character faces "in shade" regardless of position, where the sun is, and regardless of prompting. Seems to be an epidemic in the Anything based models. Is there any surefire way to ensure the face is evenly lit with the body? have tried multiple types of prompts and methods
This model is pretty useful, I think it is a bit better than Grapefruit for most of my prompts. Would you mind if I made a mix with this? I'd like to share my results here in a review if it is okay. Thanks for your hard work! (:
I have no problems with that, do what you like :)
Released just as I'm almost done with a batch using grapefuit, I'll test this one with the "same" prompts once I'm done for comparison. :heart:
Thanks! don’t overwork you, it is simply my first attempt to train.
Nah it's fine since I was planning on testing it when when it'd be out, to see if I could find points of improvements(as I couldn't give many when you were doing the switch from mix to trained model)
Details
Files
hassakuSD15_v1.safetensors
Mirrors
2583_hassakuHentaiModel_hassakuv1.safetensors
hassakuSD15_v1.safetensors
hassakuHentaiModel_v1.safetensors
hassakuHentaiModel_v1.safetensors
hassakuHentaiModel_hassakuv1.safetensors
hassakuHentaiModel_v1.safetensors
hassakuHentaiModel_hassakuv1.safetensors
hassakuHentaiModel_hassakuv1.safetensors
hassakuHentaiModel_hassakuv1.safetensors
hassakuHentaiModel_hassakuv1.safetensors
hassaku_v1.safetensors
hassakuHentaiModel_hassakuv1.safetensors
hassakuHentaiModel_v10.safetensors
hassakuHentaiModel_hassakuv1.safetensors
hassakuHentaiModel_hassakuv1.safetensors
hassakuHentaiModel_hassakuv1.safetensors
hassakuHentaiModel_v1.safetensors
hassakuHentaiModel_hassakuv1.safetensors
hassakuHentaiModel_hassakuv1.safetensors
hassakuHentaiModel_hassakuv1.safetensors