I generated over 50 images of Alice last night in a beautiful flower-filled garden and she looked absolutely delightful, so I have no complaints. (#^_^#)

(No point in trying to upload anything here, however, as the anti-loli policy is ridiculous on this site. Realistic models I can understand, but flagging a full-clothed character walking down a path with no hint of sexuality whatsoever is quite frankly retarded.)

Long torso can be down to the resolution I've noticed, especially if you go over the recommended sizes. I'm using 'retro artstyle' myself, so I'm not seeing much of a problem with most of my favourite characters (produced some great Sheila images last night, right down to the freckles on her backside). Someone suggested adding 'realistic' to the negative prompt, so I'm going to give that a go later on and see if it makes any difference. (=^_^=)

一方で、線が太くなって色使いが派手になったことは、アニメ的とも言えます。realistic をネガティブプロンプトに入れると、そのことが良く分かりますよ。

"I’ve noticed the same thing. The model seems really unresponsive to lighting tags; it tends to force dramatic lighting even when I explicitly include specific prompts to counter it."

It's as if they are trying to purge furries and monster girls from the model. Every version since V14 gets worse in this regard. V17 doesn't seem to understand the term "digitigrade" anymore

@Hylocereus fr, barely any options besides this one as they also are heavily biased towards human anatomy, I absolutely despise all of the furry models with the generic art styles and stupid faces they create, at this point im just going to try and gather all of the lucky images I got and try to make my own character lora because this shit is getting annoying

How do I use it? I used the "CFGZeroStarAndInit" node and the generated image became garbled

@SomeTimesNaive233 What are your global settings? If you aren't using Tangential Damping, CFGZeroStar can easily burn the image. Also, ZeroStar tends to be less stable with high CFG values. I use the ordering: Tangential Damping -> ZeroStar

I'm new to the WAI IL checkpoint, can you explain briefly why v14 is considered the peak thus far? What's wrong with the newer versions?

理由呢？下结论谁都会下。

whats best for lora training?

v15 is the best, with superior graphics compared to v14.

Yes, if you prompt correctly =)

有加lora吗，会不会是lora的问题，我之前就是这样

@arknightszxy971 No, I have no lora added.

Same here. With some prompts I could get maybe 1 image without text for every 200 with. Best way I found to "fix" it was move to Unholy Desire Mix.

You can try using lazy negatives for general negative coverage.

@RenjiXII 这样啊，那就只能后期再用PS处理一下了

are you using lazy neg/pos embeddings if so try without lazyneg

You're absolutely right.

所以选V15吗

看几率的

加方向提示词

See if stuff like this helps:

"Flat colors"
"rough sketch"/"Simple sketch"/"simple drawing"
"lineart"
"Pencil drawing"
"Black and white"
"Flat shadows"
"cross hatching"/any other shading method

Negative prompt: highlights, saturated colors, over rendered colors/over rendered lineart
There are also lora like "anime screepcap" and "pencil drawing" and others that you can experiment with. Good luck.

@z2944344686 谢谢，我会试试的

@darthmobian57825谢谢你，我会试试的 

Can you suggest the best versions of WAI-illustrious in your opinion?

@iwid224 I've been using for months the v14.0 and v16.0. In my opinion and experience, v16.0 isn't better than v14.0, but it offers a slight variation if you somehow aren't satisfied with v14.0 results, or if you wanna test something a bit different. For the record, in my opinion v14.0 is the best version not only between these two but between all currently available.

I share the exact same opinion. The rendering in v16 was excellent and completely lacked that artificial AI feel. However, I was quite confused when I tried v17, as the coloring looked just like low-quality, mass-produced AI illustrations. v16 really is wonderful.

it's worse than v16 and I use the exact same parameters.

v11 is still best

Let's say that, many model creators neither know basics of AI/ML nor know what they are actually doing... never a good thing.

Try adding the tags "anime coloring, anime screencap" at the beginning of the prompt; it usually works well. If the prompt is long, add weight like (anime coloring, anime screencap:1.2) or (anime coloring, anime screencap:1.3), but if it's short, you can add it at the end, which usually works.

What specifically are you struggling with?

lots of loras...

@Balthasar_  just simply getting the image quality of some of the loras that are posted here, not matter what I try I just can't seem to get the same level of quality and I honestly don't really know what I'm doing wrong or what to change, I've been messing with a lot of settings both here and in comfyui and a lot of images just barely manage to reach the level of quality I see here.

@Balthasar_ I wish we could post images because even while following the suggested steps that are on here for this checkpoint the images I make still look off somehow, either too bright or the faces seem washed out and even using detailers or upscalers I can't seem to fix them.

@Mytei what loras would you recommend?

Careful LoRA application, and figuring out prompt terms that give certain finessing visual effects. Just gotta play around until you start seeing what you want and like, that's what I've done. Also a lot of images here you can see peoples prompts and various setting, some even have their comfy nodes available and you can literally copy and paste their image workflow, and tweak/experiment with it.

@bruh7554 dm me and we can talk over discord.. i was where you are xD..

In my experience, my first piece of advice would be:

1. Reduce the size of the LORA files you use. The recommended value if you're using more than two is 0.8.

2. My second piece of advice is to try to use as few prompts as possible, or only the ones you need.

3. Try using a simple or white background, as using more than two LORA files tends to generate artifacts and errors, and a simpler background will give you better results.

4. Don't use detail enhancement LORA files when creating an image; it's better to use them in the "img2mig" section. I hope this helps, and good luck

@sentecia will give that a shot

@Proboot sent you a quick dm

@bruh7554 
style:
- Hyperrealistic_illustrious
- CunnystyleV10
- 3d loras for different skin etc.
lighting (if you want bright/dark images):
- Dramatic Lighting Slider
- lightingSlider
Character loras:
- Body Type Slider - Illustrious_alpha1.0_rank4_noxattn_last
- Skin color slider - Illustrious_alpha1.0_rank4_noxattn_last
- waist_slider
- distance_slider (close/distant shots, can make batch image generation consistent)
***
- mimimeter
- heightRatioSliderIllustrious
2 of the above if you want to get chibby style characters
***
- StS_Age_Slider_Illustrious_v1

Also, Dont ever listen to comments saying to use less loras, its BULLSHIT!, If you want your images to look exactly the way you want and imagine, you ABSOLUTELY! NEED to use every lora there is to better up the chances of model generating them. You remove 1 lora, bang, image completely different.

I mainly use SDXL, but you get the idea what loras to look for, you can try ANIMA, ZIT, or FLUX depending on type of picture you are looking for out of the box.

why the hell are people following me LMAO

To be honest, in my opinion, the sweet spot is version 14; there's no better one. Versions 15, 16, and 17 went from bad to worse. I recommend you stick with version 14; for me, it's the best one so far.

@dupe_900 Thanks for the tip. I'll go and try it out.

so mang people say this,I will try,thanks

17 еще не пробовал, 15 да хуже 14. а вот 16 как альтернатива под другие арты

Используй Lora, и будет тебе счастье

use v14 deffo

@tobycortes true and v16 is the limit...

diffusion_pytorch_model_promax

Yes, but it can be somewhat inconsistent depending on the combination of characters and/or loras used etc.

@HarblFarmer can you give me an example prompt? I've tried it before and it always ends up duplicating characters when I try to make them stand side by side though for some reason it can handle two characters standing back to back really good

@bruh7554 https://civitai.red/posts/28908979 This is a super basic example.

do this :"masterpiece,best quality,amazing quality,very aesthetic,absurdres,masterpiece,best quality,highres,beautiful,scenery,2girls,shoulder-to-shoulder,looking at viewer,

1girl,ganyu \(genshin impact\),blue hair,shoulder-to-shoulder,looking at viewer,in the right,

1girl,midori \(blue archive\),blonde hair,shoulder-to-shoulder,looking at viewer,in the right next,

school uniform,school,"
but it not good at three or more

@HarblFarmer

>it can be somewhat inconsistent

It will be, i'd say.

well earlier,civitai did not allow model names containing 'NSFW' to be published.
anyway,i also suspect that some training sets may be contaminated,but v13 and v14 are still the two versions with less contamination

set clip skip to 2 is what works for me!

if u are using comfyui, stay in the range of -1.5 to 1.5.

@NeuralKnit Not sure what is the difference between 1.5 and 1.5.
I'm aware that usually these kind of tag are supposed to be weighted on positive values, like on that description said the creator suggesting 1~2.
I'm just goofing around, thinking the possibilities of that tag might be adjusted to be a built in slider like those slider LoRA.

Who knows right?

And look at those salty people still thumbs down my comment haha. What a sight~

@zetsubou I believe it's because NeuralKnit gave a range from -1.5 to 1.5. You prolly just missed the minus symbol^^

@mysdproject377 no i'm not. He edit his comment apparently. With or without the minus, I already understand the context back then.
Thanks for telling tho.

@Edd08 not only you disliking my main comment, you also disliking my reply huh.
How about you dislike this one too and my other reply so both of us can get more blue buzz?

@zetsubou oh ok np^^

all danbooru tags +

check this out it, you can deep dive into this an have clear view of tags

https://en.wikipedia.org/wiki/Danbooru

deepbooru (xxxx - 2020) is the old version, this illustrious models use danbooru wd14 (2005-2023), danbooru includes 3 datasets, 2005, 2017, 2023, :)

I am confident this checkpoint can do that since I saw a post similar to that months back in the gallery posts. I don't have it saved so I'd say your best bet would be finding an illustrious lora with the pose you want, and an illustrious character lora for the character you want and then generating those with this illustrious checkpoint. Also if you can find a post with something similar to what you're looking for [remix] is always a valid option for prompt.

why do you need this💀

dude

maybe a control net of ice hockey players hip checking each other? (half serious)

1. Take a screenshot of that vid
2. Upload it to deepdanbooru
3. Start from there and start tweaking tags
4. After you get a good result go touch grass ASAP

I’m genuinely curious—why is everyone praising version 14? I’ve tried it, but I still find myself preferring version 16. I mostly stick to basic text-to-image workflows with the occasional LoRA, so perhaps I’m just not utilizing it to its full potential. Could someone explain why v14 is so highly regarded?

@rihabnvm626 I'm speaking from experience: from the details of the eyes (even when not prompted for 'perfect_eyes' or 'clear_eyes' or using ADetailer), v14 is probably the best model on this site. It handles fingers very well too. A few negative prompts for extra_fingers or fused_fingers go a long way. I rarely get generations with four or six fingers. Even at an angle, when the fingers aren't the focus, the checkpoint does an incredible job of guessing. Your mileage will obviously vary, but mixing in four to five LoRAs (as my workflow requires) has given me excellent results. I suggest checking whether the LoRAs you're using were trained on v14 or something similar. v17 is close to v14, but not quite there yet. v14 is lightning in a bottle.

v17 backgrounds are way way better and also the new characters are awesome too

@rihabnvm626 V14 just makes strong 2D anime style without extra prompting or Lora, solid stable anime model. Anything above it tends to add some kind of realistic 3D shading and effect that tends to end having 2.5D bodies and background on a 2D face and it looks off. V17 is a very good model anyways but without prompting or Lora it looks like a semi realistic model instead of anime.

The answer to this question is trickier than it may seem.

On technical level, you just... send a 192 token prompt and it automatically gets split into chunks. There's no upper limit. However, you get diminishing returns the more chunks you have working on a prompt, it makes things weirder with CFG, generally needs more steps to converge, on top of already requiring extra compute.

Not to mention, since each chunk is kind of a separate entity, which tags go to which chunk make a huge difference on the result, but there is no real "rule" here, just experimentation by gut feel. Sometimes you get a good result by splitting chunks by categories, like a chunk to quality\style, a chunk to scene description. Other times, you get better results by... weirdly enough, making a goddamn mess of your prompt with even distribution of categories among chunks so desired concepts get reinforced. It's a rabbit hole you can drown in. And it's purely from the limitations of the old text encoder of sdxl era.

I usually see strong degradation starting with 3 chunks and not even bother with 4, but sometimes 3 is still acceptable, and in rare instances, the only real way to get a complex scene right. So, my advice is... Don't make complex scenes and stick to 2 chunks!

And, yeah, the separator tag for manually splitting chunks is BREAK - no commas required, neither before nor after. Example: tag1, tag2, tag3 BREAK tag1, tag2, etc...

I would write the prompt in a text editor, like obsidian, and strip the prompts into different types of segment, then break that one down into smaller groups of more meaningful "tags". avoid repeats.
that's my normal process when I see massive prompts, normally it's just unskilled people typing lots & hoping, this helps you prompt far more properly as well.
also break it down into "style, scene, character, score/ quality improvers"
that's what I get the best response from normally.
keep in mind I artifically induce token attention segments using the BREAK command

Thanks sires and i did my research and i made my own UI whenr ei run modles on my system in python so i went on AIs to learn an dmany vidoes where i find that we can divide the cokents in tho chunk chunk 1 will have 77 tokens and remen f will go in th other chunk one chunk can hold max 77 but i still think some token got ignored by the AI Modle whne generate images

Because it sucks

give it more steps bro!!! needs more processing to get to get it properly, tbh give it 50+, i havent had any problems the model is well trained and fine-tuned on hands by now just give it more room to cook :)

i did some digging up in the generation consistency at lower and higher steps, what i saw was that indeed it has good hands trained in, but also in some positions the middle and ring finger get lost because of the subtle thin outline in between gets lost in the denoise because its almost the same color as the skin itself, so it just fuses it in later steps, so YES the aproach would be little help from ((lazynad:1.4)), and also more steps on Karras schedule type, if your composition is simple you can use regular steps 20 - 30 cooks really nice!!!, BUT if you got some complicated pose or more tags you deffo need 50+ steps to let it cook properly, another thing i found was that food get really well detailed at 50+ steps, so maybe does little details in the hand get resolved even further on later steps, this is just my opnion tho, your sampler+schedule recipe need to have strong game, i really dont care about generation timings if the results are good, remember scketh/test at 15-20, fast cook at 30-45 steps, and let it simmer with 50+ steps and BATCH COUNT 8 youll have enough options for the result

@tobycortes Adding 50+ steps really helps with finger separation, but more often than not, V17 generates a sixth finger, which even tags like lazyhand and others fail to fix. V17 has indeed made a huge breakthrough in background quality, but characters now suffer from more model glitches. Therefore, I consider V14 to be the best choice; the sixth finger almost never appears in the images, even with a minimal number of negative tags.

@alex2402200334417 v14 deffo !!!!!!!!!! above all

i can confirm that. It's reached almost the pinnacle now. Don't know how it can do better.

Illustrious model works especially for anime/cartoon/comics characters. Asking for 3d games characters is pretty difficult without lora because you need to prompt "every" aspects of the character without specifying the name.

This LoRa: https://civitai.com/models/1446362/eve-stellar-blade-7-outfits

Works well with wai illustrious. The difficult point is if you want an anime/toon version of her, you need to hard weight in negative prompt "realistic" and "3d".

Just add the "2d, anime style, lineart, drawing" with a weight of 1.2-1.4 and add 3d in negative prompts. This should do the work.

Actually... And where to hurry and why?I think updates come out exactly when it takes to keep up with popular characters and new data to learn

@RussianMan_BRUTAL that's not really an answer

tell us why you need 18 so soon?

@Ayayatou rage bait?

I didn’t know that people used AI graphics in this way O.o". It’s shocking! Even sickening... but if it amuses you.

RusseMan_BRUTAL is right. It’s a lot of work and he is not alone in working on files that come only from him but also from other people’s data, logically it doesn’t depend only on one person (at least not in this case).

As long as I leave this comment, I take the opportunity to tell the mod uploader how his work (mod series) is flawless, functional, awesome, intuitive, complex but easy to use, etc... (and so on!).

Hoping wholeheartedly that the quality will continue to follow, thank you so much for this real pearl allowing us to design true masterpieces!

---

Je ne savais pas que des gens utilisaient les arts graphiques IA de cette manière O.o". C'est choquant! Écœurant même... mais si ça t'amuse.

RusseMan_BRUTAL a raison. C'est du boulot et il n'est pas seul à travailler sur des fichiers qui ne proviennent que de lui mais également depuis des données d'autres personnes, logiquement ça ne dépend pas que d'une personne (tout du moins, pas dans ce cas précis).

Tant que je laisse ce commentaire, j'en profite pour dire à l'uploader du mod à quel point son travail (série de mod) est impeccable, fonctionnel, génialissime, intuitif, complexe mais simple d'utilisation, etcetera , ... (et j'en passe!).

En espérant de tout cœur que la qualité continuera à suivre, merci infiniment pour cette véritable perle nous permettant de concevoir de véritables chefs-d'œuvre!

@Sylenzyo howly japp session

Try "lying"

Also "lying on side" or "lying on back"

I'm noticing it too

Usually the male focus tag works well

Personally, I have no problem using it for male characters.

If I'm generating a single male character pic, I use "1boy, solo, male focus" at the beginning, prompt. Most of the time, the results are excellent.

Anima

you add the step and use scheduler for highstep such as heunpp2, or DPMPP 3m sde ithink, and if you want to complex poses i think base model "anima" great option

Anima does really well with anatomy and natural language out of the box, although it does require some setup. Finetuned models are even better and handy for specific styles. I recommend RDBT Anima for speed and flexibility.

There won't be a v18. Anima is the new hotness. Illustrious is dead.