Wdym?

@qek Hires upscale

@yangyuanxiao22651 what i always do is i use upscaler to produce better result or final result that i want. i disable upscaler only for testing to see if the prompt and the result is what i expected. once i'm satisfied i enable the upscaler for the next generation.

不會，我不清楚這背後是如何運作的
但標籤是否更新取決於底模作者們是否願意多花費心力去處理

he did, old one is in one of his links in the description.
while this version is kind of better quality, its kinda bad for nsfw, and bunch of character loras work worse.

I recommend inpainting for more detailed pubic hair

You can give it a try. excessive_pubic_hair

而且也没有要重建的意思，估计是凉了

@OathOfJudah 啊， 好可惜

V15 experience:

I put the artist as the first prompt before the quality prompts, unless the model has issues, then I put the artist after the quality prompts, and if that does not work I give them a value of 0.8, and that has tamed things.

i also do

You aren't alone. It's very noticeable when the skin and fur color is purple.

@fallendragonwolf115 It seems to be somewhat better now. At least for the models I use. Any idea what the heck happened?

@Nolovebutlover probably the models having a freakout. it isn't on the fault of the loras in any way. I tested such.

I was getting similar issues with reds.

I was wondering which one you would recommend more, v16 or v14. I started using AI-generated images relatively late, and when I first heard of waiillustrious, it was already at v15—so I never had the chance to compare it with v14...

@Aetherlyn that was also based on preference, but if you new i recommend using V16 because there has more many integrated anime character, that also still works well if you have long prompt and more easy to control if some piece didn't work well

but you can also try V14 or V12. maybe you will find some interest in that model as well

I have the same issue. When I upgraded to version 16, I often get incorrect hand results. Is there any solution?

Based model

......,

masterpiece, best quality, realistic,

......,

....., aged_up,

.....,

Hope that helps!

Does this work because the thumb is not considered a finger?

@Dark_Yami79 the odd thing is that I don’t use positive prompts like the one that person mentioned, I only use embeddings that fix hands, or LoRAs that improve them

however, using too many embeddings, such as neghands often causes the image to change slightly

I recommend using embeddings like neghands, or IL / XL hand LoRAs, to fix the issue.

I personally add "five fingers" to the prompt. As a side effect, the character might try to show the hands too much, making weird hand gestures. So I would often omit it from the prompt initially and add to the prompt when inpainting hands

@Dark_Yami79 That is my experience so far, yes.

@sheevlord i made the experience that the thumb is not considered a finger, so trying this gave me 6 fingers most of the time.

我也喜欢😍

@WAI0731 please look into this.

@krigeta I have so many complex scenes and poses I want to try

maybe in a few more months. Since the image size is a bit larger than a normal IL model, some people have cut it down into a turbo version, and it runs faster

there are already some existing LoRA files that give images an "anime-like art style" or cel shading, but it’s not 100% ready yet, but almost though.

maybe use “1girl, solo” or add weights, such as “((1girl))” or “(solo:1.5)"
this often fixes the issue, you can also use a LoRA made for a specific number of people.

@HHNation Yes I use specific number like 1girl all the time (can't use solo if you want the 1girl interact with male(s)), the problem here is even when I specify something like 1girl and 1boy often they appeared multiple times in different angles as if the AI is trying to draw a manga with multiple panels. That should be something you'd set with LoRA or specific prompt, not default. It doesn't happen every time btw, but noticeably far more often than older version or other model.

bless you

yea its something i noticed only started past couple days, big brownish smudge on the character's body

I think you can try using other SDXL VAEs? For instance, I've tested and the smudge is not there with, say, TAESDXL.

the solution is to close civit and touch grass

832x1216

You can keep 1024x1024 but since illustrious v2 is merged you can have consistent outputs at 1536x1536 but not for every characters.

you should add into negative prompts:
more than five fingers per hands

Also in positive prompt, it helps to specify "hands have five fingers" which will makes the model focus more on hands.

Normally you don't need those depending the pose of the character but these prompts help having less often those extra fingers. This works for foots too.

@InoSim i will try it, thank you so much

Lower your CFG, this model recommends 7 but it will make extra fingers more often and CFG 5 is more than enough

When I look at the image, my first impression is that Lora is also a problem; it seems overfit. How large is the base graphic? In other words, what dimensions did you use to generate the first sample? How high is the denoise setting on the detailer? It might help to increase it to 0.7-0.8 if you decouple the detailer from the main positive prompt and put your own one upstream. How many steps do you run the graphic through the ksampler the second time? And which denoise do you use?

[removed since poor and confusing advice]

@Wesl801 By all the good gods, what did poor ximtars ever do to you that you would subject him to this monstrous workflow? :D

I work professionally with Comfy, and this thing raises more questions than it answers for me. Maybe it's because I have no idea what the Clown Pack is or what you do with it, or why it's supposed to be better than classic samplers? But why are the hands and face detailed twice? What on earth is Controllnet doing in there? Why? Why x4 upscale? I think this workflow is clearly designed for photorealistic images, and slamming it on poor Loony is like going rabbit hunting with a sledgehammer. If I understand him correctly, he wants Loona to have better eyes. And she would get that quite significantly if he turned up the denoise from the detailer and put a support Lora like “Better Eyes for illustrious” (there are several recommended ones) in front of it. And in any case, the strength of the character Lora should be reduced a little. If the results then become worse or more inconsistent, a better Loony Lora is simply recommended.

This is not primarily meant as criticism. On the contrary, I would be really interested to know what you are trying to achieve with this workflow and, above all, what the advantages of these samplers are. ^^

@graurabe Haha yeah I'm still learning, you gave them the correct advice. Didn't intend to confuse. Those are very good questions that I should have been sure of before posting. I was just trying various Illustrious workflows and noticed consistent detail with that one before any detailers, but it does seem to be geared for realism than anime. If you want to dive into these crazy samplers there's an explanation workflow you would understand better than I do, that the author included on their GitHub: https://github.com/ClownsharkBatwing/RES4LYF

https://civitai.com/images/121429524

Here's an example. These are Loony and Mills from a Lora I made myself. The image is the raw version, which I created here with the Civit sampler in 30 steps. Of course, it would be even better with a decent Comfy sampler. But that's not what this is about. The image is the raw image, which means it hasn't been run through the face detailer or sampled again; it's simply the most rudimentary image possible. And yet the eyes look better than in his example after he ran it through the detailer. Often, the quality of the Lora is crucial. You can tinker around a lot with images from poor character Loras, but they'll never really look good.

Depending of the expected style you can add tags such as "flat color" to your negative prompt to avoid getting pics with... flat colors. Set a higher score to the tag when you keep generating flat color pics if adding it to your neg prompt isn't enough. The same goes for other style/coloring tags. So use the appropriated Danbooru tag for what you try to generate.

glad to see i'm not the only one, it affects a lots of stuff inlcuding LORAs

I don't get green on the clothes, but I do get weird orange and fushia as background sometimes.

you can change your Negative Prompt,When high-weight negative prompts are used, they can causeColor compression、Yellowish shadows、Abnormal contrast、Greenish patches in certain areas、Muddy or dirty-looking skin tones

@dosjimmybrand you ! are ! my ! god! You have successfully solved my problem!I always get purple or red tint,it's so weird!but you!OMG!

The models are trained on tags and images. In the early days of the original Stable Diffusion (SD) models, they were trained on existing artists’ works, which were essentially learned and baked into the model. Tags are also used during training, and each model often has specific trigger words depending on what you want to generate.

Some newer versions of models, like ZTurbo and Flux, can understand natural writing better, but they are not 100% accurate, more around 85% accurate. However, models like IL, SDXL, and Pony typically rely more on tagging.

Some of them can understand simple prompts or phrases, but it really depends on the specific model and the metadata that was used during training.

你的想法很不错，但这取决于现有知识架构以及物理规则，我们/他们已经尽可能做的很好了，无需再次提高要求

ComfyUI Image Workflows

But the best illust model is noob.

This seems to be common with 1080p generations. it seems to struggle with high resolution 16:9 ratios and compensates by stretching the character across the width and introducing artifacts rather than creating an appropriate composition. if it wasn't for this issue I think this would be the best model hands down

@Fealow Ah... Well then i suppose i couldn't cheat out quick generation without hire,fix on huh for this, thanks for the info

"shortstuck" or lora?XD

95% ill loras will work good on this model

Just an other name, same Checkpoint.

我也在等17版本，不知道明天会不会出，等太久了

Positives: (aged up:1.4)

Negatives: (petite, loli, chibi, child, aged down, minigirl:1.4)

you're not doing anything wrong, the author just probably used a ton of loli images and didn't tag them as so so the checkpoint thinks lolis are 'default'. And the downvotes are legit just from pedos

Telepaths are on vacation. Provide the prompts you use.

Try adding "adult" to the prompt. Negatives in my experience don't help much here.

@protectron ok i try, though 'adult' is not a real tag

@Kasper369 Again, anecdotal, but I notice it makes a clear difference in facial features and body proportions for the same image with and without the tag. "mature" tag amplifies it further.

@protectron thanks

https://civitai.com/models/1025710/age-slider-lora-or-illustrious-xl?modelVersionId=1150298

Age Slider LoRA

True

是的，我还以为我的问题，看来大家都是一样的

try to use parts you want to see in the positives, so, for example: thighs, legs, calves - this can sometimes help

full_body as a positive or out-of-frame/out-of-frame_<bodypart> as a negative can help as well if you're trying to avoid mentioning a body part so it stays clothed.

I typically just use 832x1216 or 1216x832 for landscape orientation. The issues you mention arises when using an unstable resolution on which the model hasn't been properly trained on.

Resolutions that should work;
768x1344, 832x1216, 896x1152, 1152x896, 1216x832, 1344x768.

If you still have problems maybe it's the prompt itself? using "solo" can help sometimes but also more importantly I've noticed is to try and not mix prompts that would contradict each other since that tends to mess up anatomy.

Trying to prompt for too many actions at once like if you want the character to hold something or do some other arm/hand action tends to generate multiple limbs.

You'd think that simply prompting for two different hand actions the model would understand that you want "this hand to do that and the other hand to do this" but nope, still get 3 arms once in a while!

This problem is not exclusive to WAI of course, happens with every model I've tried.

Rule of thumb for prompting is to ask yourself "what do I actually want to be visible in this image?" i.e. prompting "feet" or "shoes" with "upper body" will tend to try and cram the former prompts into your composition which may end up looking odd.

But I'd try out those resolutions above first since that seems to be your issue - if all else fails you can look into Kohya HRFix.

It does say to use larger than 1024x1024 above, and indeed, I've been having few problems with 1080x1440 with Stable Diffusion / Forge myself. However, without knowing your precise setup, LoRAs, prompts, etc., I can't really add much more than that unfortunately. (^u^)

As all things should be

This is a common issue with pony/sdxl models, unsolvable just by prompt (and on civitai generator). If you're using Comfy on local - you can try ComfyCouple or anything with regional attention/guidance.
Also, Invoke is good at this, it's less techy and has a canvas where you can draw a rectangle and prompt a character to appear there. (doesn't work 100% correctly, but far better).

Check my profile - I have some pics with 2 characters I've made using Invoke with regional guidance feature

actually, there are few prompt techniques that increase the chance that characters won't swap clothes/colors. One is to put a character description in parentheses, and also separate characters by BREAK word. Also adding "duo" and action description usually help too.

duo, dancing,
BREAK,
(1girl, red dress, blonde hair),
BREAK,
(1boy, black suit, blue hair),

That definitely helps a little, but there's still some bleed over, especially with NSFW interactions if you know what I mean. It also has a tendency to reverse the roles and no reordering of the prompts you type makes any difference. Of course, you could get around this by simply having both characters completely naked, but it does kind of defeat the point. (^w^)

@sherriterri320 there are few tags on danbooru that might help with clothing: clothed_female_nude_male, clothed_male_nude_female, clothed_female_nude_female, clothed_male_nude_male

you can try one with higher attention like: (clothed_female_nude_male:1.5)
might give additional 5~10% of success

reversing the roles is a big issue... :( I get that even using Invoke with regional prompting.

you can try describing with natural language like: boy is submissive, girl is on top
or using character names instead of boy/girl

@sherriterri320 Try removing quality tags (masterpiece, highres...) and leaving only relevant tags to what you want. This makes them have more weight and tends to make the result more accurate. Than do image to image with quality tags.

Using all these tips, I got something sort of close to what I wanted, but the swap-over of characters still persists. For reference, I wanted a futa Haruka (Pokemon) with another girl, but no matter what I did, it would always switch them around. I'm probably going to stick with single characters for now and use PixAI's new Tsubaki.2 model for pairings since it's a lot more consistent overall. (^_^)

You can start by adding "Mature look" in your prompts, or add several ones in the negatives such as "Young, loli, childish" etc...

Yeah if you think they look too young for your liking, just add things like "age up, mature female/male, adult, older, etc." and then add to negative "loli, sh ota, child, minor, etc."

@Lain32_ it doesnt fix it, i am still getting crude, young characters, what is happening

It does indeed, although I'd be nice to know precisely which characters are included and whether or not more are going to be added. (=^_^=)

also

Every Illustrious based model is trained on Danbooru tags.
https://danbooru.donmai.us/artists?commit=Search&search%5Border%5D=post_count

Alternatively, you can use Style Lora.

https://civitai.com/models?tag=style

Up though march of 2025

@uels321559 You've been a great help, thank you very much

@GM_MD For v15 yes, but for v16 nope

@Haiishou  V16 How to find the style list

@1132470273558 for artist and character tags, use this https://www.nep6.com:8806/
for artist mixed style, use this https://dskjal.com/spa/noobxl-artist-blend/ (preview was generated on noobxl, so it may differ if you use wai-il)

add "mature woman" to your prompt. A lot of these models lean moe - setting (mature woman:1.2-1.4) can give you a pretty good-looking result while still retaining the model's features.

Just add "loli" and "young" to your negative prompt.

suck to suck

hah

add watermark to the negative prompt, if on a webui can use the extension Negpip to make sure it doesn't show up.

Even with negatives, it will generate a watermark/logo occasionally. Flux2 Klein can easily remove these (or nanobanana if it's SFW image)

the logo usually coming from artist styles or lora, try to remove it one by one and see the difference

You need to use adetailer extension for that (if you use Forge/Forge Neo webui). https://github.com/Bing-su/adetailer

If you use Comfyui, use these extensions: https://github.com/ltdrdata/ComfyUI-Impact-Pack and https://github.com/ltdrdata/ComfyUI-Impact-Subpack (to use other variants of yolo detection models)

Use the Hire/Fix option then put in the recommended upscaler and settings

@springmushroom_86 Which ADetailer detector are you using?

and what do each of them do? For example face/hand/eye, they fix faces hands and eyes or they make it more hd and detailed? and if they fix them then how exactly do they fix them? I heard that for example face_yolo and some other ones at the top of the list were used back in the day when the ai was generally bad, but now it's pointless to use some of these adetailers since the ai does a great job already. is it true?

face_yolov8n.pt

face_yolov8s.pt

hand_yolov8n.pt

person_yolov8n-seg.pt

person_yolov8s-seg.pt

yolov8x-worldv2.pt

mediapipe faco fult

mediapipe_face_short

mediapipe_face_mesh

mediapipe_face_mesh_eyes_only

@glockkatana Adetailer is still very useful to use, for example if you generate full body, model won't be able to add more details in regular resolution (SDXL is only trained at 1024px).

I use face detailer model most of the time, I tried hand detailer before and they just enhance more details on hands part only (it won't get rid of multiple hands etc you need to inpaint them). I use Anzhc's face segmentation model, unlike yolo face models, they are using segmentation mask detection for the face area only.

@glockkatana for the rest models I don't really use them, last I know about person yolo model it's only good for realistic gens

@glockkatana it fixes those by detecting the face first (in a square) and then rebuild this square by regenerating this place which will be more accurate

If you need details, you need to use a higher resolution or use a new magnification node.Simple prompts will not provide better improvement for visual details.

yes, crazy how we still need loras for popular recent gacha characters Q_Q

The character knowledge even went down from v15.0 to v16.0...

@lizardon1024 ye... q-q

@RizzieLove I don't know if you are interested in this, but I've created an article comparing the character knowledge of different models :)

Why should models waste any amount of parameters on gacha games, just use a lora, crazy how many anime models are so focused on a billion mobile game characters and cant do anything else well

@DiffusedSignal Yeah right, crazy how people are obsessed with Anime characters when using Anime models to create images of Anime characters... /s

@lizardon1024 lets remember its just a 6B model, if I have to choose between one that can do complex angles, poses, has good prompt responsiveness and knows artist styles well, I'll take that over one that has trained on a myriad of gacha characters, but can't do anything outside of ''1girl, standing''.

LoRA training has come a long way and I dont think all anime fans even care about the latest gacha/shows being released. Theres like 40 years of anime history at this point, trying to include it all or stay up to date is fruitless, better to focus on a strong base model first, just my 2 cents.

@DiffusedSignal Illustrious models are trained mostly on the Danbooru database, so if you are unhappy that this database does not include enough pose tags, you can probably suggest new tags at Danbooru (or use pose LORAs for specific poses that are not included yet).

The reason why there are so many "gacha characters" is that they are incredible popular, which is reflected by their immense presence in those Anime databases. So the Danbooru database contains whatever people like and love to draw. And people seem to prefer to draw newer characters (which are often from gacha games) instead of older ones. The tag system then tries its best to describe what is shown on those images, which then of course also includes the character tags, of which a large portion today are "gacha characters".

The biggest difference between all those Illustrious/NoobAI finetunes is:
1. Their art style (or responsiveness to art tags, but this is slowly decreasing compared to the base models)
2. Their prompt adhesion (which becomes worse if incompatible models are mixed)
3. Their knowledge about new tags that were added to the system since the release of the Illustrious/NoobAI base models. And the only new tags that are added to Danbooru are almost exclusively new character tags.

And you seem to care only about 1 and 2, while others are also interested in 3. That's all.

If you are unhappy with the general tag system but still want to create Anime images, I can highly recommend Anima: https://civitai.com/articles/25969
Anima not only understands the same tags as Illustrious, but also natural language. So you can simply describe in words whatever pose you want if you cannot find the right tags for it. And Anima is even smaller compared to Illustrious but much more capable, so the size of Illustrious wasn't the limiting factor.

After The Elders Scrolls 6

@zethrasgorgoth772 哈哈哈哈哈

"Monster" monsters I assume ? Because for humanoid monsters, that should be very good if you can manage your prompts correctly

So which one do you recommend?

Not just that, when I prompting Ronnie Anne Santiago, the style of drawing is on anime--not from the cartoon itself.

It must be a glitch for this checkpoint.

because pony is designed to make furries better, illustrous is better at humans

That's because Pony is trained on by a bunch of furries with a bunch of furry pony stuff, that's why the model is called pony. You're asking for a healthy meal at McDonalds

after v17 and before v19

Hi,
https://danbooru.donmai.us/wiki_pages/tag_group%3Aimage_composition#dtext-styles

and https://civitai.com/articles/7819/illustrious-xl-v01-visual-dictionary, but I recommend that you ask Grok a question like: Help me find the nodes that are responsible for the danboory tag tooltips when you enter a hint with the function of automatically adding tags and a loaded library and access to the site when you click on the book icon inside the hint. It should help you find those Nodes making it easier to work with danboory trained models.
and.....It's best to use styles using Lora, it's much more stable

And... I literally just found out about such a wonderful thing as a translator for Danboory tags. He understands natural, narrative, literary language and translates all this into a tag system.
https://anythingtranslate.com/translators/danbooru-prompt-translator/

Don't go too mad with the initial resolution and particularly the denoise. I've been experimenting with this myself and anything over 1024x1536 is almost guaranteed to cause problems. Adjust denoise accordingly, having it higher for lower resolutions, such as 0.4-0.5 for 1.5x scale and 0.25-0.35 for 2x scale. (^_~)

@sherriterri320 Hi, I'm aware that resolution, denoiser and scaling are the causes, but why is it so bad? Im using 1152x1536 at x1.75 with 0.42 denoise, and sometimes (quite frequently) it causes many of the problems mentioned above... But, when it works well, the quality is amazing.

With my initial question, I wanted to know if there are any alternatives to avoid these problems, either by using Hires. fix with some tips that im unaware of, or without using it at all. Currently, I wouldn't know how to reproduce the quality and sharpness of my images without using it.

@NobodyANY Your starting resolution is pushing it a bit... I mean, I was using 1080x1440 at 2x and 0.35 denoise and I kept getting doubled-up body parts, especially with partially naked or naked subjects (not so bad when fully clothed, even at 1536x1536). The other thing is any LoRAs you're using also have to be able to handle the resolution otherwise issues are going to arise no matter what you do with Hires. fix. One guy on here recommends 832x1216 at 1.5x and 0.5 denoise, which does give reliable output, though I have been trying to extend that to at least 1024x1536 at 2x and 0.3 denoise with mixed results. The upscaler is pretty important too, with RealESRGAN_x4plus_anime_6B producing the best images (Lanczos is fairly good as well). At any rate, you can't get a good clean image without Hires. fix, simply because the model itself is only trained at 1536x1536. Any image generated at 2048x2048 say, merely culminates in doubled up characters. (^_^)

@sherriterri320 I've tried quite a few resolutions, and it seems that for my setup (neoForge), the most stable base is: 1024x1360 resolution, x2 hires. fix, 0.35-0.4 denoise, 5-6.5 cfg (depends a lot with what LoRa im using) and RealESRGAN_x4plus_anime_6B upscaler. Its not completely consistent (75% +-), but im quite happy with the result.

Prompting order and danbooru tags also play a big role, specially, i try to not overdoing it with quality tags and negative prompts (im mostly using the ones that WAI recommends + those that i really need for my specific type of image).

Note: Funny part is that pretty much all this settings are the ones recommended in the main page by WAI.

@NobodyANY You're not wrong about that. I was playing around with generations of Valkyrie from Valkyrie no Densetsu last night using all manner of resolutions and I ended up back with 1080x1440 at 1.5x and low denoise, which is very close to what you're using (I'm on plain old Forge myself). You can certainly get away with more, but the risk of extraneous data creeping in goes up significantly the closer you get to the model itself's limit. It is frustrating though, to get a beautiful image of Flora from Jayce and the Wheeled Warriors out at a massive 2048x2880, only for the very next one to be deep fried in some way. (>_<)

I notice this occasionally too - smaller versions of characters in the background. It is usually a sign the denoise value is set too high, which introduces things into an image on resize that simply shouldn't be there. (^_^;)

If they are on danbooru website then it is possible just by using their name tags, if a style isn't in the checkpoint, then you are required to train a lora on illustrious V2. Then use that lora with wai illustrious to generate the image.

what v25? I just see v16 now.