thanks for the answer

@WAI0731 can you give me refrences website that can fully ultilize capabilty of WAI NSFW Illustrious model

@WAI0731 thanks you

I downloaded the whole thing and is running it in stable diffusion on my own computer.

look like issue on your pc, I have same speed in 12 and 11 version

Yea, probably...

https://civitai.com/articles/10962/illustrious-prompting-guide-or-v01-or-generate-anime-art-with-ai

Give V12 a try since the description says V1-V4 are outdated. Not sure about the others, but V12 is perfect if you're using the right prompts and art style LoRAs. If you're looking for an anime-style shading LoRA, look for one—it's often a flat anime art style LoRA.

Most Illustrious models can do anime style just fine if you prompt for it properly.

I usually default to these in my prompt for anime style: "anime artwork, illustration, anime screencap, anime style, key visual, vibrant, studio anime," + the respective recommended quality positives of whatever model you use.

Personally, in case of Wai-NSFW, I recommend V11. To me, it feels like V12 is less adherent to prompts. I also prefer the style of V11 more.

V9 is the most anime-like and responsive to artist styles, in my experience. It generally doesn't require style loras.

1536x1536 drools

which version

What do you mean by that?

Just prompt what you imagine :)

If its bc of resolution etc.... use minimum 1024x1024 / 832x1216 or 1216x832. Steps 15-30, GFG scale between 5-7 (your choise), Sampler Euler a (karras). Use Quality prompts : masterpiece,best quality,amazing quality (at the beginning) and start prompting.

Okay, a short example, just copy it:

masterpiece,best quality,amazing quality,1girl, solo girl, white background, cowboy shot, red hair, blue eyes, bright smile, summer dress, summer hat,

As negative use : bad quality,worst quality,worst detail,sketch,censor

Thats very very basic, i hope you get the gist.

@whity00 my dumbass thought a prompt generator was the actual program. i got that part but im still having some issues. 1. idk how to make the images look like the models 2. i tried to generated 5 characters and it only gave me 3

@hentaigood137997 Honestly, your best bet is to look on Youtube for a bit of training. Learning to prompt isn't something that can easily be conveyed over a simple chat like this. Also, you need to specify what UI you are using - comfyui is very different from Forge/A1111, and results can be very different. Understand that there is nothing intelligent about AI, you can tell it all you want but it will never be even remotely perfect as you've noted. Also the demo models are cherry-picked from hundreds of generations usually. To view what prompt was used, click on the image - if the creator opted to share the prompt it will display there. Play around, remove part of the prompts, change them, you'll get the hang of it.

@sildenafil928 got a low end laptop gtx 970m. tried forge first and now trying comfy

It would suck if v12 goes away for a while because v12 was my favorite

He/she did it already,look at his profile

@P_Universe and where on his profile do you think you see a mix with 1.0?

@m4rcus the model is called something like fck illst,just read

I think it's just you.

Which god you had in mind 😂

There are a couple of tropes, as well as anatomy tags that can help with this. Adding "bishounen", "bara", "femboy" and a few others will consistently generate men. Tagging male anatomy also helps (the tag "pecs" will give a rounder look to the chest, and may sometimes create breasts, but "pectorals" will not).

But yeah, overall I agree. This model is fantastic but it has a difficult time with men and androgyny.

@henriqpass Have you tried to generate male bisexual pictures? I got nothing on this topic, just futa or yaoi, but no luck with actual male bisexuality.

I've just tried "2boys, shota, anal, yaoi" prompt and got exactly what it means, and still yaoi picture even with "anal, yaoi" removed from the prompt.

@empanada_48 What do you mean with male bisexual pictures? Like a guy doing it with both a man and a woman? haven't tried it yet, but given that it's a complex prompt I would believe it needs a lora

@henriqpass Well, i mean homosexual-and-heterosexual intercourse involving at least two males and at least one female at the same time. I've tried some loras, but they either affect the style a lot or just ruin the picture by turning it into lovecraftian horror.

@empanada_48 You can circumvent a lot of those issues running it local with extensions like lora control, that allow you to time in which step the lora is going to act and leave, but besides that yeah, I don't know

点开webui右上角的扩展
选择从网址安装
把https://github.com/lanner0403/WAI-NSFW-illustrious-character-select
这个网站复制进”扩展的 git 仓库网址”
安装完后重载ui

@cAzic 谢谢大佬！

Personally, I like SHUFFLE more, cause v-pred.

check the illustrious licence to get an answer of range of use.

copyrighting is a case by case basis. if your end output doesn't make a "close" copy of an existing copyrighted picture, then there is no infridgment.

a diffusion model is a transformative work of the original art data feed into it, the model nature is not an art piece and is fair use still today.

know that depending on your workflow, you may even copyright an image that have used AI. txt2ilg only won't be copyrightable, you have to photobash and do post process to inject human input to generated parts. even if you don't create all the visuals, the placement and tweak you bring to your picture might just suffice to make your assisted AI image copyrightable as a whole, but not in parts. you can still rip some part off it's fair use.

That happens when you set the resolution too high or a wrong aspect ratio, the model tries to add something to fill the empty space (in your case a duplicate character). Lower the resolution or if you're using A1111 or Forge, use something called Kohya Hires.fix. Look for tutorials about it on the web, it helps with generating stuff without duplicates at different aspect ratios and resolutions. Hope it helps <3

https://civitai.com/images/65990815 idk if this can help, you got me curious to try. worked well, just not too big i'd say for me saying i wanted them large.

@SneakyPisces that's not good enough 😢

@evilstormhot109 I think I see what you mean. To be honest all the more 3d models/styles do larger fangs better. 2d needs to catch up.

https://civitai.com/images/66146670 I think it can

@rivertree i don't want to be like all those guys and high likely you won't reveal it but i have to ask; what was your original prompt to get the image looking so clean even with upscaling to 4k?

illustrious is SDXL, you can train illustrious as following by sdxl lora training tutorial etc. just use weight of illustrious instead of sdxl.

In v12, for mostly full body prompts, I use 42 steps DPM++ 2S a Karras (txt2img) and 26 steps DPM++2S a AYS (Hires, 0.54), all cfg 5. Consistently found it to be much more logical and better at following prompt than Euler a with different configurations, and the finger consistensy is just impressive. Also painfully long sampling though, 2s a is double per step.

CFG of 3-4, dpm++ 2m, 30 steps + 10 steps of hi-res scaling 1.25, at .60 denoise.

These checkpoints don't look that great without the hi-res.

Since moving from 1.5 to Pony way back in early 2024, I have found that DPM++ 2M (Auto) with CFG 5 and 28 steps strikes the perfect balance between perfomance and looks. I do 28 steps of 1.3 hires upscale with 0.5 denoise on top and if I like the output, I'll top it off with Adetailer. So 1.5 minutes as is or 2-2.5 minutes with Adetailer.
I genuinely don't see the difference between 28 and 30 steps (no cherry-picking), however DPM++2M is vastly superior to Euler(a) even on models that say you should use Euler. I may try DPM++ 3M SDE to see how it affects the composition. With the latest IL models (this one and NTR) composition is the only real thing that AI struggles with. Multiple characters, anatomy and (lately) backgrounds have all passed the "as good as human" treshold for me.

@Noise_Connoisseur With DPM++ 3M SDE, one thing I'd suggest is bumping up steps to at least 40. I would get occasional visual artifacts when using 30 steps or less

You could try adding description for only that body part and nothing else. Also you can type the name of a body part, followed by the word focus like this:
ass focus, breast focus, leg focus, face focus.

You can also use multipliers like this:
1girl, aqua \(konosuba\), (ass focus:2),

i figured that "closeup something" is mostly the best tagg to have a good focus

@QH96 I'll try that, ty

use patreon username, copyright, wordwork and likewise tags as negatives

No offense or anything but I havent seen any patreon logo, random logo, watermark, signature or distinctive mark since I started experimenting with these models since 7.0. Its probably just your negatives.

@WAI0731 I use artist tags often (daily lol) so I'm thinking basic/good negatives can negate them no problem.

Thank you for all your hard work. You are my hero. <3

For me, they pop from time to time, but you can put "patreon username" in the negative prompt to fix it.

@rivertree i see, thanks

@WAI0731 大佬，我现在把ControlNet模型拿掉了，结果连一般生成角色都变成机械人了，不知道这是不是BUG，我就算重新生成工作流程也一样的结果。https://imgur.com/a/BSkj1Hv

Nevermind I get it the second KSampler is used for the Hiresfix

Good

TIL. Thanks kind stranger!

il 0.1 i talked with creator

The prompts may be demanding that this model is not learning much. Too many negative prompts may also degrade the quality of the generation. If you cannot identify such prompts, you can try lower CFG Scale.

@Hackgets what is that? sorry im a real noob

@Temure its in the advanced settings. Right below the negative prompts. its the first slider above the Steps and above whatever Sampler you're using.

Bad anatomy is a placebo, regardless of model.

If you use negative tags to fix deformities, idk what to tell you. This is not sd 1.5. Use negative tags to remove items only. If you get deformities, I'd have to guess you've overloaded Loras or overloaded prompts.

your loraS bad.

first of all Illustrious, but if there is no Ill Lora then try to use NoobAI Loras

You should be using loras that are trained on Illustrious v0.1

There's a lot of cross compatibility between the two really

noobai is also trained from illustrious so do what you will with that info