Never had that problem with wai. Maybe you can link a sample picture with your prompt and used Loras. In that case, answering the question would be much easier.
Are you using onsite or Offside Generation?

Just put monochrome in negative dawg

Try "colorful" in usual tags

v14 handles some poses / concepts better than 11, but generally, 11 is better overall. IMHO 11 = 12 > 14 > 13 > 15. 15 also handles some poses / concepts better than everything that came before it, but the anatomy seems unfixable from everything I've tried.

Check the site announcements, basically the whole site has become read only for NSFW free users and for NSFW users who can only pay with credit card. The only way to generate NSFW now is if you pay them with other methods.

@dragonfable081551 So presumably @WAI0731 changed the name to make it not appear as if the model is just for NSFW?

Can you explain what you mean? Did I miss something these past few months? 😮 😮

@Eliz99 Civitai now requires you to purchase either a membership or yellow buzz to generate any material that is not PG

Tensor.art , Pixai.art

@Asdfghkkn these are nsfw friendly?

@qazwtf Until they get nuked too, yes

OK, but a migration is dutyfull. Users have no other way to dissent than to not follow

@qazwtf I tried them out. It was possible with Pixai.art, but it was already prohibited with Tensor.art.

https://civitai.com/models/1390184/eye-size-slider

This might help. You can also search for iris and pupil size sliders on the creator's profile

However, this LoRA is not guaranteed to fix the eyes if the character's face is far away

This one is pretty good too: https://civitai.com/models/1433845?modelVersionId=1658812

If you're having trouble with eye colors while using Euler a, simply switching the Sampler to DPM++ 2M Karras might improve the situation.

There are a few options, first off depending on what your using, you may have an addetailer function, WebUi Forge has a bog underneath highres fix for it. to use it you need to download files on here or hugging face for face, hand, body, eyes, ect. if your using ComfyUi they have an Addetailer Node that also use the files mentioned above. Also High res fix can correct many of those issue too so make sure you have that set up right. And if all that has been tried, there are some good Loras out there. one of my favorite is "detailed realistic eyes-gmr" by Artikuz_Ai. Hope this helps!

試試看负面提示增加床的權重 或者檢查提示詞是不是有涵蓋的某些XX動作跟床本身有關係。

ComfyUI is like a workflow and customization powerhouse. Usually has the newest features and is really good for video outputs, but has a sort of learning curve for its node based system.
WebUI is a more straightforward and simple approach, letting options and settings easily change results without much hassle.

If you are fine with simply making images and not trying to branch out so far in making other stuff, WebUI is perfectly fine.
But, if you want the most control, is willing to learn, and want to make unique images that you couldn't on WebUI because of limitations or trouble, ComfyUI is great to try.

(I suggest just try both for a good amount of time)

I switched from fooocus to automatic1111. I've been trying ComfyUI these days. If you don't know how nodes work, it's hell.

seaart确实坑，我唯一能生成的账号是美区谷歌注册的，节点也不挂其他地区的

The explanation has disappeared now, but apparently v2.0 is of lower quality compared to v1.0, so it seems they decided not to use v2.0.

@emmeria765 hmmm...but it is confusing as on Illustrious website they said V2 is better than V1 but the company is not sharing too much because of some funding I guess.

what bad about 15?

I also agree that version 14 was much better than version 15.

@Cat_Hisses If you only use it for girls with standard body types, you won't notice, but there is a lot of inconsistency with other body types or with boys.

9

13

16

I use 11 idk

v14

What other models do you personally recommend?

@DatCivitAIGuy If you're looking for a flat style similar to WAI, the new janku model is nice if you're willing to wait a week for it to come out of early access. Hoseki got an update back in september, and I got some really good gens from that. A 'flat color' tag at a slightly lowered strength made it look pretty similar to WAI 13/14, but it handled hands better in some poses, had better prompt understanding with dynamic angles, and has newer training data. PinkAlchemy by Ramthrust is something I've been messing around with and getting good gens with (though it's closer to 2.5d over wai; I'd still highly recommend it, but it is a bit more different). AdAstra is a NoobAI EPS checkpoint that I really like, but it does have the shortcomings of noob. Amanatsu just got an update recently, though I haven't tried it yet. Historically, I really liked 1.1 when it came out in March. That might be worth looking at as well.

I'd say start there. And don't be afraid to just look around. Some of the newly added checkpoints by random people with 1000 downloads randomly go crazy. I randomly stumbled on the AdAstra checkpoint, and it's honestly one of my favorites.

Your in-depth review is really appreciated.

Which GPU are you using? Because, honestly, with my 5070TI, this model is almost perfect.
Hands and feet are nearly perfect 80% of the time, without LoRA.
I'm coming from PonyV6 where 50% of the hand and foot generations were wrong.
I'm new to this model and the anime generation in general, but seriously, this model is almost perfect for me.
It only has some "green and pink" spotty artifacts on a few generations every now and then.
It's hard to tell if it's the model itself that seems almost fried or if it's the negative prompt + LoRA or a positive prompt versus some words.
Although to tell the truth, I'm now curious to try the 14 or the previous ones, since everyone says they're better.
Ah.. resolution always divisible by 64.. Because over the many years that I have been generating I have always seen that resolution can create distortions of the body.

@REgenerationSD I have a 4080 setup and a separate 3090 setup. I've run the model on different UIs, between reForge, Forge, and currently comfyui with a slap into a couple different workflows I alternate between. If you're coming from pony, it'll feel like an improvement. Pony is so much worse than illustrious in terms of basic anatomy it's not even funny. But as for illustrious models, I found this one to be lacking. Please keep in mind pony hasn't been relevant for an entire year now, and WAI 14 is about 6 months old.

My original review was a bit more positive on v15 as I had mainly only done limited testing. 1 - 2 characters max, limited angles and perspectives, etc. But as I widened my use case, I found more and more issues with the model. Again, there are aspects of it that I like, but nothing about it is good enough where I'd go back to dealing with simple issues that most illustrious models don't have. I personally find most NoobAI models to be comparable in terms of quality, but the issue with that is Noob has significantly more positive variety. Of course, if you're getting good gens with it, that's fine. Don't be afraid to stick to it. 15 just lacks a purpose for me when I've found other models that do much more.

i was also puzzled on how its possible that this model has so many tips and downloads but also on how many gens in the recent tab look odd, something aint adding up.

they probably goofed up release v15

Have you tried weighting "(masterpiece:1.3), (best quality:1.3), (amazing quality:1.2)," and (worst quality:1.35), (worst detail:1.35), (bad quality:1.3), and tiring your own quality tags and quality checks weight by most important to least important and ordering them in that vein. warning do allow a delta in weight of your own quality tags and quality checks by at least 0.1, and specific failure modes should be unweighted or at most weighted at 1.1, for example "extra legs, bad fingernails," etc and you should only weight the most common and worst specific failure modes, limit yourself to at most 3. which will likely be a tag to do with hands, like "worst hands"

Also if you are going for a stylized aesthetic you should probably add "photorealistic, realism, " to your negative prompt, also semireal depending on how hard you are pushing lighting and other stuff which would invoke a realistic aesthetic.

And to the person who voted this down the issue of adding too many tags is you are cluttering the latent space leading to dropped tokens, but prioritizing quality tags and quality checks by their importance you make sure that the weighted tags are the last thing to be dropped, and as for specific failure modes you don't weight those highly because the model will spend undue attention on those areas if you do, which weighting quality tags doesn't have nearly the same issue as they are global tags which will effect the whole image. Also really your goal for putting specific failure modes in the negative prompt should not be to completely eliminate the chance of them, but to reduce the chance to a manageable level.

Am I against weighting heavily in the core prompt, oh hell yes, max 1.2 in extreme circumstance, and the vast vast majority of your core prompt should be unweighted and you should use descriptive language instead, ultra, hyper, best, detailed, exquisite, being examples of descriptive language that generally works well rather than stealing the models attention with weighting.

You can use UltralyticsDetectorProvider with Eyeful_v2-Paired.pt in ComfyUI, for example.

you will need to use ADetailer with eye detection to redraw eyes in higher resolution.

@lk74 Correct! I don't know why HarblFarmer gave you a thumbs down.

It is a general problem of 1024^2 resolution. Take your pick from: 1) standard high-res fix features, 2) Kohya Deep Shrink and, 3) ADetailer. I love Kohya because of the balance of quality and efficiency. By the way, putting "empty eyes" into negative prompt makes eyes more starry. "Detailed eyes" does not work on IL-based models, basically.

@empanada_48 Will it work in WebUI Forge?

@Aleksey035 I dunno, i used A111 and ComfyUI and they always worked well with Eyeful_v2-Paired.pt.

imagine using Pony in 2025 😂

it's difficult to say. if it were me, i ended up downloading several checkpoints and ran the same prompt on a few different versions. compare and contrast, then keep the one that has the best results to you :)

may i know ur workflow ? please

Danggg you just gave me a reason to download this, i was like mehhh another anime checkpoint that i don't need but saw your comment and instantly hit the download. lol

Just throw old_man or ugly_bastard in your neg prompt, it normally gets rid of that

@kotsarelos62662 mature_male, elderly

So using youth or young prompt. Opposite is "mature male".

You can create whatever you want with it

I also feel that the same cue words perform very well on v15 and v14, especially in light and shadow

Nope, it will do that from time to time

Use inpainting.

Fair enough. It doesn't bother me as such, I was just worried because some images got flagged as "sisters" and "twins", which wasn't my intention. I guess I could try regional prompting too.

@Afgmc Regional prompting is just a shittier and more complicated version of inpainting. Trust me, it's really easy and you'll probably kick yourself for not trying it sooner

@magicballoon, I was having problems with Img2img in general messing up the image colors on A1111 and Forge but it works better in Forge Neo so having tried it a bit, inpainting is clearly the way to go. Thank you. [EDIT] The root cause was the gamma and srgb chunks that Paint.Net was adding to the SD output files when I edited them before using Img2img, so I wrote some code to strip those out and put the original SD text chunk with the prompts back in.

if u use A1111 u can try BREAK word, google it

@evilstormhot109, DAMN! Been using SD for years and never had much faith in BREAK, but just saying something as simple as "dark hair BREAK blonde hair" with 2girls seems to work very well indeed.

Try add clone in Negative prompt.

v14

ALso has a lot more watermarks... v15 sucks compared to v14

Been trying v14 now as I started off on v15 and it seems much more sensitive to prompt weights. I was having to put some very high values in v15 to get the effects I wanted, but the same prompts on v14 make it produce nightmare fuel.

@haenlesn937 I still haven't had any watermarks since I started using it. I was surprised. I upgraded from Pony to this version V15. No writing, no watermarks for me.

@REgenerationSD yeah, watermarks aren't a big issue.

It's because CivitAI new policy of course.

Yeah, I found WAI's answer on the question, thnx. Looks like it's #1 FAQ

@evilstormhot109 What FAQ?

@evilstormhot109 can you tell me where to find the answer or give a tl;dr? would be much appreciated

@madaraxuchiha88 Frequently Asked Questions = FAQ

@IPrometeoI I get that. I'm asking where is it? I can't find the FAQ that has WAI's answer that @evilstormhot109 is referring to.

@madaraxuchiha88 My bad. I think the FAQ can be found on the CC's discord server

To be honest i d even say that to this date v11 is still the best.

which workflow can anyone help me?

Sad but true...

Extra limbs: The same seed and the same prompt often produce multiple limbs on V15 while V14 is fine.

Weird anatomy: The body in V15 is bigger, longer and more mature overall, and it's much less responsive to artist tags compared to V14.

Try this amazing LoRA: [for illustrious & noob] Reduce Wrinkles(Lines above eyes)

Any workflow Text2Image should work

@victorhugodias2001189 *for SDXL

Maybe try using BREAK to define each character?

if you gen locally I use this with comfy https://github.com/Danand/ComfyUI-ComfyCouple. Works pretty great and managed to make it work with wildcards for random prompts on each gen. There's might be a bit of style leak sometimes but characters dont bleed to the other character most of the time

Try using the BREAK command at the end of a line to seperate persons from one another.
Also if you want better control over their position in the resulting image, try using a regional prompter.

@PossibilityWeird5153837 @Mytei it isn't what it's for and it doesn't work, IllustriousXL is enough good to make separate characters and you can inpaint if it doesn't work

@qek Thank you for your response, i wasnt aware illustriousXL couldn't make several characters. I will try inpainting. I hope to find an more advanced version. thanks

Use both. (refining)

both

V 14 is peak

How can an AI model influence your hands, butt and genitals?

I use ComfyUI, so this might not fit your environment. I found best results if you run a lower strength in the LoRAs (.5-.7) and prompt the model more "(like this:1.1),like this"; it can steer the gen towards the char you want! Seeds are also kinda important, so fixing a good image seed rather than a running numerous random seeds can help you poke the gen towards your vision and desires! I haven't really messed too much with character or style LoRAs yet, so I'm more familiar with prompting. 💋

I just made one for ComfyUI! It's XXX /Hentai Gatcha style - https://civitai.com/models/2166861

I made a "bare" version of my workflow which removed LoRAs and Custom nodes, for your ease of use - https://civitai.com/models/2168743/hentai1911-freyja-pixel-bareback-edition

It's just renamed, it's the same exact checkpoint as before the rename.

I disagree, V15 is fantastic. I can do so much with it. I can regulate how 2D to semi-realistic it can get, it's very rare when I need to use a lora. I haven't used V14 once I made adjustments of using V15.

BTW I'm doing some behind the scenes testing on a Dual model Wai-Illustrious workflow (v. 12, v. 14, v. 15)! Like if you're interested, and I might calibrate and release it 😘

GOAT MODEL

For real? I stuck with 14 instead of 15 but I'll give 9 a try.

I'm going to be doing some research with older Wai models as well! 💖

I tried v90 and I'm hooked

@renderguy I have v10-15 downloaded for my workflow, I'll give v9 a try! 👩‍💻

@pixelladies45 That's cool. I Have only tried 14,15 and now 9 its pretty good for what I do.

I can only assume it has to do with the PG version of Civitai.

832x1216 is pretty common one i see ppl use with XL models

Here's a cheat sheet for you:

640 x 1536: 10:24 or 5:12

768 x 1344: 16:28 or 4:7

832 x 1216: 13:19

896 x 1152: 14:18 or 7:9

1024 x 1024: 1:1

1152 x 896: 18:14 or 9:7

1216 x 832: 19:13

1344 x 768: 21:12 or 7:4

1536 x 640: 24:10 or 12:5

730 is not /16, make sure that's always the case, doesn't really matter otherwise, it does well until 1536x1536, I usually go with 1024x1344 or 1440 or 1536, they all good

Is 1056x1536 crazy town for initial generation size?

Totally serious! Happy to give back to the wonderful community - LoL

your not wrong

Yeah I agree my personal favorite and V11 too

Same

I disagree.

@GM_MD And what is your measurement methodology? I took 10 different seeds, generated the same pic with them, with the same prompt, with V14 vs V15. How did you measure?

i was on v12 for a LONG time and the last months started using v13 and v15. for using "cocktails" combining lots of LORAs i've found in my experience that v15 is a lot more stable.

@SillairArt thanks for the heads up, i'll pick V15 and see how it goes

i'm on v14 right now. before that of course v12. i already testing several image in v15 with many loras combination but i that return too "cook" somehow.

I use v13 mostly

I switched to the "wai Nsfw Branch Rouwei ePred1" version xd Much more stable and no less colorful, although a bit old. V15 may not understand some tags and there will be a conflict, on "wai Nsfw Branch Rouwei ePred1" there is more stability, better at dealing with conflicting tags.

Can you retry that exact same example but instead of "pin up" try "pinup \(style\)"?

Cuz I think that's where those random medals and emblems are coming from.

@Starvader Sadly, couldn't recreate the exact same image - turns out it was an inpaint, tried to remive that stuff. But there were other images:
https://civitai.com/images/113497137
https://civitai.com/images/113497131
https://civitai.com/images/113497282
https://civitai.com/images/113497284
And it indeed helpeed! Thank you so very much!

@JxC Glad to see my observation ended up helping, I just cross referenced that particular tag in your prompt on Danbooru, since everything looked good except for that, and seeing how there is no "pin up" tag over there only "pinup \(style\)", my guess is the model read that as "PIN on the UPper portion of the image/subject body" through Natural Language instead of ignoring it, causing the problems you experienced.

@Starvader always more to learn about Illustrious, it seems x) Pony or SD1.5 weren't as picky about the prompt, but I can't deny the quality upgrade!

@JxC Oh yeah if you come directly from Pony and SD1.5 you gotta be aware Illustrious/NoobAI models are primarily trained on Danbooru tags (NoobAI is also trained on e621 tags), with just a bit of Natural Language to varying degrees depending on the model/model mix used, so to ensure you don't fall into issues like this again I would recommend making a habit of double checking any tag you feels is not working right with that website to ensure you are using the proper tag for it.

@Starvader thanks for the info! It sure is useful! 🙂

@Starvader so to try different Styles om Illustrious it's better to use "artist-name \(style\}" without just using "artist-name" on Promtps?

@TatsuyaAI that depends on the Base Model/Checkpoint actually, some models like this one WAI-illustrious, only require the raw tag of the artist as you see it on Danbooru and nothing else, some models like the ones mixed with the RouWei Checkpoint will require adding a "by", so it ends up looking like this: "by artist_tag, " when you prompt them.

Just be sure to check the description of any new Checkpoint you try to see if they handle this in a particular way, if they don't mention anything at all on artist tags, it probably just means using the tag 1-to-1 with the tag on Danbooru will work just fine.

When using danbooru tags and seeing weirdness, check the tag history, as danbooru is changing tags for various reasons. The model has a time when the data set was created, and those tags may have changed after that date, so the older tags will be needed.

You're probably using a style and a detailer that also has a style; try changing their order to see if that solves the problem.

@Qweser I actually swapped UIs to forge neo and that fixed the problem :)

@AkaliLovr Cool, I also switched to Forge and noticed a huge difference.

I found the fix, when it asked for file path, like C:\user\models as an example, add 1 extra line like so C:\\user\\models, Enjoy!

@ThatSoKitten I tried it, but it didn't work.

Chek is tutorial https://stable-diffusion-art.com/controlnet/ . Ensure that you select the Allow Preview checkbox. Once you have selected openpose as the Preprocessor and the corresponding openpose model, click the explosion icon next to the Preprocessor drop-down menu to preview the skeleton.

try getting a regional prompter when working with 2 characters. i dont usually make 2 character pics bc they're super hard to get but the rp should be an ez fiz 4 u if u know how to use it

Doesn't "futa_with_female" work?

futanari,sex,yuri, or a futanari, a girl,yuri

using the tags "2girls, futa with female" usually helps on that, but there will always be a bit of RNG to it and occasionally you'll still see both as futas or something else, only other extra step you can take to reduce that is Regional Prompter unfortunately.

Yes, it was. You can see old name in the description and links

@eevqd Thanks. I'm re-downloading everything now because of new PC and was wondering where it went :D