please never say any of this again

furry dataset tend to disfigure human anatomy. It's best to set them aside for a greater goal.

@HQdesu I guess you are right.

furry and human models tend to contaminate each other

use kohya ss first.
just look at the metadata that ppl use on civitai and import them to kohya
i advise Metadata viewer 1.6

@Le_Fourbe Thanks for the response, I finally got it working with Adafactor I think, had to turn down snr gamma to 1.0 rather than 5.0 with a LR of 0.0001 in one trainer. I just kept playing with the settings one by one until the training stopped burning and it started to converge. I usually look at meta data with notepad but I'll checkout the viewer as well, thanks again.

I guess you mean A1111, but the models works just fine with ForgeUI. Also bigger resolutions are normal for XL models.

Edit: You can get good images with other resolutions but I think the model it's a bit difficult to use Im still trying to getting used to it as well.

you need 1024*1024 and around. forget this 512*512 from SD1.5
as a base model you need to put LOOOOOOONG litterate prompt to get the best of it.
it is still unreffined and undertrain but serve as base for training for much beters merges and finetunes models
i advise NoobAI-XL V1.0 epsilon. much more usable but still unstable.

@avefanes What are you using for your vae on ForgeUI for Illustrious?

@Darkarma  Hi, is just the same SDXL VAE from HugginFace I always use except when Im using Pony. So I did not touch anything else from the settings. https://huggingface.co/stabilityai/sdxl-vae

Also, use "CLIP Skip -2" (or nothing) in Comfy, don't use "CLIP Skip -1".

@nymical no don't... or do... this tips is kinda outdated for SDXL.
it was relevant for NAI leak back in 2022 and it's derivatives due to it being trained as such. this is not the case anymore.
however it might change aesthetic for the beter depending of you prompt inputs. if you get funny details skipping a layer might play in your favor.
CLIP skip 2 is not a rule

@Le_Fourbe I checked before making that comment, I'm using ComfyUI, using "CLIP Skip -1", gives random noise output, nothing resembling an image. But, -2, gives a normal image.

I don't know how it works in A1111/Forge though. I'll edit my first comment accordingly.

EDIT: Just tested, it doesn't matter on Forge.

@Le_Fourbe I have to back up what the user @nymical said, Clip Skip at 1 only gives deformed stuff and it just ignore prompts. Clip Skip at 2 is just way better.

Sorry if I cant not help you with your problem, but It worked fine for me with the SDXL Vae from Huggin Face yesterday while using ForgeUI, so at least the VAE was not the problem for me.

it has a vae built into the model itself. use that vae.

Dont use too many negatives. Maybe 1-3 words. Then, simple prompting - best with booru-tags. Steps: 20, Sampler: Euler a, CFG scale: 7, Size: 832x1216 - For good generations. Use artist tags! Really important.

Try to use artist tags unlike Pony they are not hidden, any artist with more than 100 images on Danbooru should do, dont use a VAE it already comes with one integrated into the model, use Euler A sampler + Simple Scheduler Type. I used 28 steps with 7.5 CFG scale, use XL size images like 1024x1024, 896x1152, 832 x 1216, etc. It worked for me both in A1111 and ForgeUI. Im trying ReForge next.

EDIT: I forgot to say use Clip Skip at 2.

@avefanes @Lavish Thanks I will experiment and play around with it more, I was a bit to quick with my judgements earlier

when using this family model follow the rules :
1) <1girl/1boy/1other/...>, <character>, <series>, <artists>, <special tags>, <general tags>
2) simple negative works kinda like animagine. depending on target, looks at the given sample prompt for starters.
3) this is a LARGE finetune over ALL content including bad aesthetic. as a result, it is very raw and unstable but have the best illustrative knowledge to train.
4) much like PonyXL6 era, everyone use autismMix as it is plain beter. in this case you have to look up to NoobAI-XL eps. this is the autism equivalent. i prefer the StableNoob-ai Mix iterCOMP+SPO

What I noticed by a lot of Illustious models in general, its best to turn down the CFG scale from 7 to like 3 and then experiment by going up one by one. Its a very sensitive model, but more stronger than any PONY model you can use. No need for many words.

There are a lot of finetuned Illustrious models already. They can produce specific artstyle/quality. This model contains everything, so you need to write artist name and other quality tags (pony "score" doesn't work)

@Shio_N cute man but wher eare artists names for this model, readme does not have them

@niccc discord announcement channel of NoobAI-XL finetuned model of illustrious have all artists and characters

@niccc You could also try with any artist name tag with more than 100 pics from Danbooru.

You mostly dont need Loras, unless its super specific. And you can use SDXL Loras, just test around

In my experience, most LORAs work even better than for the models you use them normally. I have to turn the LORAs down by around 0.2 or 0.3 (so if you use 0.7 normally, use 0.5 or 0.4), for it to not go too deep into it. So no, you dont need to change the LORAs.

yeh this is the one but ALSO the effect is more/less dull on other models depending on how they trained/merged.
it's about data matching.
all illustrious models are using danbooru data so this translate well overall. NoobAI-XL however added an heavy amount of e621 data.
on NoobAI-XL the V0.5 adhere more than V1.0 to Illustrious 0.1 lora.
for best matching using Obsession, train obsession.
but my instincts tells me that NoobAI-XL V0.5 eps is also a good base to train illustrious model due to enhanced knowledge and close matching.

either look at noobaiXL controlnet (beter) or the previous xinsir suite of controlnet. preferably merge that model before (itercomp + SPO model out there fix lot's of controlnet issues)

I use control net normally like any other model, I only use Control Net Union ProMax and it works fine on Reforge.

this is a symptom of a GOOD BASE MODEL, the approach is to not skip on many divergents and less quality art data.
if this is hardly predictable, therefore need more literal prompts, but have knowledge, it is for you to find or train it to your will.
merges of this model tremendosly cover that for you. i can advise PaSanctuary, NoobAI-XL right now. the later have it's own contrlnet suite.

as fourbe said, yes, this is a good thing, if negative prompt isnt effective it basically means the model doesnt know what it;s NOT supposed to generate

I use like 1-2 negatives at most and the generations are 9/10. The model just does what the user wants. :)

Try a negative prompt like this:

flat color, blush stickers, multiple views, (bad:1.05), error, fewer, extra, missing, worst quality, jpeg artifacts, bad quality, watermark, unfinished, displeasing, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract], (blurry, lowres, worst quality, low quality:1.1), ugly, (big head, bad anatomy:1.05), artist name, (watermark:1.1)

Most probably because the best images in the trianing data are always NSFW.

Have you tried putting "general" in positive and "questionable, explicit" in negative?

basically the ratio of woman under the word "masterpiece" is super high.
most artists makes nudity their best work.
so any quality tag is bound to bias over lightly clothed bieautiful womens.

Yes

@zy02 thanks!!

i start to suspect that might be a fluke. it might be too costly to train 1536*1536 10M let alone 20M.
2 months no updates/news.
we can just hope that they are still in training due to limited GPU number.

@Le_Fourbe I reviewed their paper, which includes the results for version 2.0 and 1.0, and I assumed they had already trained the new version.🧐

probably not, if it was ever planned to be released publicly it probably would have released by now.

doesnt help that they gave bs excuses in some old civitai comments about how releasing a model takes alot of time because of precautions and safety and blah blah blah (the usual excuses, it does not take 2 months to upload a model)

fortunately, LahXar lab have our back !
NoobAI it is now !

@Le_Fourbe Noob is as*

Did you consider the fact that this 0.1 version was leaked against their will? It was not an "advertisement".

If it wasn't for a leaker, you would still be stuck with Pony.

@upscaleanon537 maybe, but i dont think so, i think they were genuinely planning to gradually release the better versions but the company made them backtrack.

also, using a leaked model for commercial purposes may lead to legal issues, but they chose to release it anyway, which wouldnt make sense if they wanted to keep it proprietary,

Yes, it can use the same LORAs and tags, but you dont have to use "score" tags and you should reduce the CFG scale to something like 3 for the start to test out the mode. LORAs should also be reduced in terms of strengh. If you have used 0.7 with a LORA try 0.4 with this type of models.

This model doesn't hide the artist tags unlike Pony that's the biggest difference. They use the same booru tags but If you are more into furry I guess it's better to stay in Pony.

@anotherstabledif1714  ok thanks, no wonder the results seemed so stiff and unresponsive.

@norim Even if someone is into furry, it's better to use NoobAIXL instead.

According to the arxiv documentatio, yes Illustrious can take natural language.

@avefanes all of it or just v0.1?

worst quality,bad quality,bad hands,very displeasing,extra digit,fewer digits,jpeg artifacts,signature,username,reference,mutated,lineup,manga,comic,disembodied,futanari,yaoi,dickgirl,turnaround,2koma,4koma,monster,cropped,amputee,text,bad foreshortening,what,guro,logo,bad anatomy,bad perspective,bad proportions,artistic error,anatomical nonsense,amateur,out of frame,multiple views,weibo username,

I personally use this (thanks to Nubby)

flat color, blush stickers, multiple views, (bad:1.05), error, fewer, extra, missing, worst quality, jpeg artifacts, bad quality, watermark, unfinished, displeasing, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract], (blurry, lowres, worst quality, low quality:1.1), ugly, (big head, bad anatomy:1.05), artist name, (watermark:1.1)

Thank you both, I will try it later!!

I don't use anything. :D

@duskfallcrew I would think this is the way to go. Less is more.

That didn't work for me, I tried about everything with the clip skip, removing it, adding a node to direct it, different checkpoint loaders, can't get a single thing to work.

It’s typically 5-20 images minimum but idk if this model is different for some reason

It seems to vary, depending on how obscure the series is on the net otherwise. Here's a post using only prompts for Illustrious. I may have gotten even better results if I vetted it from hundreds of images rather than using the on-site generator. Meanwhile villainess-isekai heroines or even characters from 2000s shows which didn't have a massive art-making fanbase would regularly fail to produce results. At least that's the impression I got so far.
@yiun 

Good day sir, could you elaborate?, does it work with natural language?, does it follow the prompt format of base SDXL?, can do any type of Danbooru tag at all if needed?, I know flexible is a broad term but could you give me a quick example as I'm really curious about fiddling around with it?

@cookiechainsawxx I think it's designed to be used more like Pony with tags, but I've heard some natural language can work too. And by flexible I mean it's better at poses/composition and style. Pony tends to have bland poses, and because it's not primarily an anime model it isn't as naturally good with anime styles like this is. But ofc it's a bit more complex than that, this model tends to look kinda washed out with some loras. I think some people are still learning how to train for it.

@justaguy111 Interesting, I mostly use this merge: https://civitai.com/models/857327/ponyillustrious-xlmix and achieve some dynamic poses without controlnet, but it needs to be pair with another CKP because make all character look "really" Asian.

@cookiechainsawxx Search for NoobAI instead, it is continued 34+ epochs more off of this on all of danbooru/e621, it's easily the best and most flexible anime model far surpassing Pony and even NovelAI. NoobAI-XL (NAI-XL) - Epsilon-pred 1.1-Version | Stable Diffusion XL Checkpoint | Civitai

@fizalpher And who you may be kind stranger?, thank for the heads-up, onward then to new horizons!

@cookiechainsawxx just someone who notices a lot of people still haven't heard about the successor to Illustrious & where all the development has gone towards after release of 0.1, Lol. It is a different company behind it but they threw like 32 h100s at the problem and it's effectively beyond what Illustrious V2.0's was at this point (the closed source one).

@fizalpher I'm so far behind the news, I thought Illustrious was an university project and to my knowledge there was no "closed source" version of 1.0 and 2.0 mentioned on the paper as those where never trained... My only concern is how much this new recommended model is trying to be more than a merge, to be frank Ponyv6 merges including PonyIllustrious have served me well on all SDXL related environments, from natural language comprehension to Ipadapter to Controlnet. For example I trained 2 models on Ponyv6 with OneTrainer and those work flawless on perfectpony, PonyIllustrious and even vanilla Illustrious. My true concern is to need new resources to do anything with NoobAI-XL the same thing that's gonna happen when Ponyv7 comes out as a "none stable diffusion based model" (according to the author). I know NoobAI is SDXL based but the author claims things have changed.

@cookiechainsawxx The paper did actually show v1.0, v1.1 and v2.0 in the "Model Comparison" it's quite a bit beyond that now for their closed source project but I believe that NoobAI surpassed it around 0.75 epsilon mark and they're quite past that at 1.1 with epsilon & 0.65s v-prediction model(eventually v-pred should be superior but right now eps 1.1 is more stable/better quality for results).

@fizalpher wonder if they are gonna keeps releasing a new model version each week? somewhat gives me that "fear of missing out" vibes and makes me distrustful (also Blue Archive imagery reminds me of gacha hell), that said thank a lot for the time and proper words, I'll try the 1.1 model 👍🏻 cheers mate

@cookiechainsawxx they are, they have multiple discord communites where things are discussed & are pretty consistent with weekly releases around fri-sun based on the latest releases/experimentations. not much FOMO really, some people prefer older models to the new ones, like some people who used 0.75 epsilon even while v-pred was coming out and 1.0 epsilon was out. all depends on which you like best, but objectively 1.1 is the best right now.

@fizalpher illustrious is definitely does better in some aspects, like in the paper they showcased an image generated at 1248 * 1824, noob cant do that.

@Hugs288 Was it a lineart or a full color+BG? A native 2K output sounds crazy D:

@cookiechainsawxx illustrious does natively generate up to like 20mp i think was in the paper? but it's pretty far behind in knowledge/aesthetic (the v0.1 version) which makes it not really good to use as a base model right now since it seems like the versions after might not come.

@fizalpher https://www.getyarn.io/yarn-clip/8d10445d-5692-4e12-8674-6ce8d65530e8 (I never read the paper Lisa, I'm sorry)

Go to danbooru here and choose character categore and order by count. The more posts there are, the better character knowledge. There is one coveat, though, Illustrious uses this dataset, so any characters that appeared after 2023 wouldn't exist in base model, but maybe in some finetunes of it.

I've only been using IL for about a month, but I'm now fully converted from pony to IL- I'm still learning and certainly no expert so take with a grain of salt, but here's what I found helpful starting out...

If you're prompting artist styles, then checkpoint doesn't matter as much (compared to what you're used to with Pony), but checkpoint does certainly make a difference. I like the Nova merges (Nova Orange, Nova Anime, Nova Unreal), but tbh that's just because I always liked the Nova pony's.

Prompting artist styles is imo the best approach to create totally different styles. (Yes, Pony also knew many artist styles, but IL seems to have much more extensive knowledge and better adherence to an artist's style). Here is a map of artist styles for reference https://satisfying-trusted-pruner.glitch.me/artists-emb-2048-v2.html (this was linked from an article here on civit but I can't find it now- credit to op)

You can also look at my showcase images on my IL loras for examples of drastically different styles using same checkpoint. You can also mix artist styles of course.

As far as loras, much like XL vs Pony, you can sometimes get pony loras to work on IL if you turn the weight way up (like 1.5+), but don't count on it.

But, IL is much better with prompt adherence and seems to have much more general knowledge (including extensive NSFW knowledge, better even than Pony imo) so I've found that many loras that were needed for Pony aren't needed with IL- characters especially, but also many concepts, poses/positions, clothing, anatomy features, etc are already well understood by IL.

Downside is realistic IL checkpoints are nowhere close to realistic Pony at this time, but there are promising ones in the works.

All this is anecdotal and ymmv. But for my use case and personal preferences I find IL is all I need anymore.

@icelouse Do you also have experience with NoobAI-XL? Because I have read people talking about that too, but I have no clue how that compares to IL or Pony

@ralkey None unfortunately... I haven't even got my feet wet into it. It's on my list along with a deep dive into Flux and 3.5 😂 But I too would like to know how it compares to IL... Hopefully a NoobAI evangelist will respond like I did

@icelouse I did a bit of research.
IL got a dataset of 7.5M, NoobAI-XL has a dataset of 13M. (source)
I'm not going to say that NoobAI will be better.. But I can assume that NoobAI is at least not bad.

@ralkey I recommend trying out NoobCyberFix v3.0. This is the latest noob version with some anatomy fixes. You may be able to use some pony lora with it but at very low strength. Also the prompt adherence is so good you don't need nearly as many lora for the same results with pony. Anything noob related is really good and it can work with other illustrious stuff well. https://civitai.com/models/913998/noobaicyberfix

@icelouse amazing comment! Thank you for that wonderful insight! This is why the open source community is the best!! I wish I could give your comment 10buzz :P

@HaloSkull thank you for the recommendation. God, I have so much to try out, and so little time.

@icelouse thanks for the write up! Makes exploring a new checkpoint much less daunting!

Through my own testing, I found that IL does not really like the pony style / character Lora's I like to use, sadly enough.
I hope the other checkpoints recommended can do that, as I have no alternatives to these Lora's.

edit:
NoobaiCyberFix ended up working the best with my lora, doh sadly still not good enough, with any weight.

@ralkey The character may work with just tags if it has them in the booru data set

@HaloSkull That's the difficult part.
The character I like has only 4 entries in the danbooru DB, with only 1 entry being properly colored.
Not only that, but none of them are in the style from the character's original series.

Use a plugin or try to lower CFG

Before I'd get a ton of weird over saturated images but after restarting my SD webui most of the art looks good.

I also just found out.

Reading up on it now.

Just started looking into it a couple days ago. Honestly I've been liking it to far, I don't think I'll be going back to pony for anything anime based. I'd suggest you start with this merge tho. https://civitai.com/models/835655/illustrious-xl-personal-merge-noob-v-pred05-test-merge-updated

(get the v3.0)

This is the result I got, Pony vs IL, incase anyone wants to help me.
https://civitai.com/posts/10036715

You definitely want to prompt an artist name if you're using the base IL model here.

Or, try one of the other IL checkpoints, which will give you a certain style without prompting an artist (similar to what you're used to with Pony checkpoints having a certain style)

I usually do both- use an alternate IL checkpoint and also prompt an artist (or multiple artists)

Related: here's that article I was trying to find in my other comment. I also recommend epiTune's other articles for IL help

https://civitai.com/articles/8454/illustrious-artist-text-encoder-umap

He quotes from IL paper which basically says you need to prompt an artist or results may be "dull or unclear"

You might not be able to perfectly match your favorite pony style lora, but with all the artist styles IL knows, you can usually find a close match (by combining multiple artist tags, style tags, etc all in prompt). It's definitely something new to get used to but I like the flexibility.

Here's an example of your prompt with just a couple changes: several different artist tags and style tags, also some of these use other IL checkpoints so check the meta data:

https://civitai.com/posts/10040143

Also re: your other comment about loras, you can always ask the creator to re-make it for IL, I've been doing that with a lot of mine.

@icelouse I see.

I can understand why you would like it, but I find it rather awkward.
As I do not know any artist's name and their quality out of my head.
So I spend too much time finding an art style I like :p

(I will definitely save your post to use as an example later)

I used NoobaiCyberFix (also an IL checkpoint) and that has a default style I can use.
Which worked quite wel, doh it could be better.

@icelouse I'm going to shoot you a random question, and hope you have the knowledge.

I want to make a Lora of an artist, should I use the base IL or an alternate IL checkpoint?

@ralkey Base model probably. Not 100% sure for style training, but I know for objects I'm using base model IL

@icelouse approximately how many artists are in the database if you dont mind me askin? i assume that ILL cant know everyone.

@3th Well I assume it knows every artist on danbooru. How well it knows their style probably depends on how much content the artist has.

In epiTune's article I posted above there are several different size maps, small map is the top 512 most prolific artists, the biggest map is 27,982 artists.

When I'm looking for artist styles I use the very last link in that article which is 2,048 most prolific artists

They aren't released

@munchkin thank you, have they said when they will release the models ?

@mabaguettegrandi990 Indefinite, only that they do plan to release them, which was a long time ago now

There's a model that can be an actual "PDXL Killer": NoobAI-XL.

@CuauhtemocI5MAL I'm so confused how this all works. What's the difference between Illustrious and NoobAI on SDXL? I see the name "NoobAI" on plenty of Illustrious models too

@enigmo NoobAI-XL is responsive to prompts made with Danbooru and/or e621 like Pony Diffusion V6 XL tags, meanwhile Illustrious is only responsive to prompts made with Danbooru tags.

NoobAI-XL can replicate the style of several artists, meanwhile Pony XL can only the style of a tiny amount of artists.

I saw someone said it uses Danbooru tags. O:

@LemonSparkle ye they seems to be working but its nice to have just cheetsheet with all examples for styles

@ohellno1 I'm not having a great time with it lol, pony does a lot better for me, and even some SD 1.5 mixes, but I got a little better results with it using an artist tag, and exact match tags and neg tags, I just started poking around danbooru to find stuff that looked okay.

@LemonSparkle tbh i was mostly using some nova models because they switched from pony to illustrious and they are quite decent

Most of the time it works poorly

No, sdxl lora's wont work as intended with Illustrious base models

Hit and miss. Sometimes yes, sometimes no, sometimes...kind of?

There are a handful of loras that work between SDXL/Pony/Illustrious. Mainly sliders, and they usually require higher or lower strength. The majority don't work, though.

The stuff on danbooru that is AI, is tagged like that (mostly), so yeah it works as a negative, I've used it before.

@LemonSparkle Oh, I thought they had a large sd 1.5 dataset separate b/c I can't access very much AI art on Danbooru and filtering out Nai from stable diffusion didn't seem to do much in the negative prompt. It seems like the only ones available are pending approval. 🤔

@Aki_dayo The ai-generated on Danbooru is around 6.9k images right now, bigger than artists, so it should work, if you're looking to avoid common issues with it you can try these negatives too: "artistic_error, extra_digits, fewer_digits, white_outline"

@LemonSparkle Mhm. I should be seeing them, but they aren't showing :< It seems only Safebooru shows AI-generated images (500). I can't seem to access other images on Danbooru, even if it says 6.9k, I can only see the pending ones (~3 pages) out of the 344...

I wonder if they aren't actually allowed on danbooru (https://danbooru.donmai.us/wiki_pages/ai-generated: "Uploading fully AI-generated art to Danbooru is against the upload rules; see forum #223144.") (perhaps why we have AIbooru) and perhaps it's training data from the AI-assisted tag? idk. This is odd.

@Aki_dayo Oh, you probably need to look at "AI-generated status:deleted" to see a lot of it, O:

@LemonSparkle woahhhh :d Thank you!

I'm a dog person, I require a trigger warning before being shown a cat.

From weeks trying it out it sticks more to your prompts and it knows characters just from prompting them, pony models aren't as good generating characters unless you add a lora, as for training I've found it easier to train than Pony, at least for styles.

It does not have the artist names censored nor hidden with tokens unlike Pony, that's good because while there isn't an official list I can test thousands of styles or fuse them without relying on LoRas just by looking up names on Danbooru if they have enought pics, I can get up to 3 characters without need of regional Prompter and prompt adherence is also better than Pony. Results are more natural like. Also it knows far more characters than Pony, the dataset of Illustrious is just far bigger 7.5M images against 2.6M images (Pony), unless you want furry content I guess Pony is better, but there is also NoobAI for that.

In my limited experience with it so far, using just basic prompts, it defaults to a 2d art style a bit like a manga or cell shaded anime illustration with more fine lines, and it tends more toward white or simple backgrounds though it can create some crisp backgrounds when it does.

While pony by default tends to produce more of a 2.5d style, that when lined often looks like it was done with a (dying) medium Sharpie marker. Pony makes backgrounds easy enough but they have a tendency to look a bit like an impressionist painting, and a lot of the time Pony images can have a sort of muddy or undersaturated quality to them.

That said there are also drawbacks for each of them too, while pony has some, albeit very limited, ability for composition control and can tolerate a mix of tags with short few word captions for interesting combinations, Illy has very little ability for composition control unless it is something tagged on Danbooru, and doesn't tolerate captions, it only likes sequences of Danbooru tags particularly in series with artists that are tagged on Danbooru, and Illy can output a lot of garbage far easier unless you give it the right prompt (like prompting a specific artist), while pony seems a little more forgiving at least to me.

Both can be adapted easily enough if someone is making a lora, and for a 2d character I can see why someone might go to Illy though, or if they're really into stuff for particular artists or characters on Danbooru.

you can prompt a character name and an artist and it will just do it, nothing else needed. pony needs loras and other bloated tags. in general this is a much easier to use model

@Hugs288 there are some very thoughtful reasoned useful responses above, and then there's yours. Do better.

I don't have to waste 21 tokens in my positive and negative to fix a "it's a feature not a bug" in the training dataset (ponyscores). Beyond that prompt adherence is just incredibly great if you understand how to use danbooru style tagging. If you want to write midjourney style, great American Novels then you'll join the chorus of "OMG Illustrious sux" people. But really I cannot think of a single thing that I've noticed Pony does better as long as you understand how to engineer prompts properly

@aikage Score tags are the worst, I can't stand them, and I avoid using them when I can, they don't even work besides bringing up some bad prompts to maybe just mediocre, the aesthetic scorer they use doesn't know what makes a good image it prioritizes busy images with lots of random nonsense, mistaking that for detail, over actually good images (also it just really likes pantyshots for some reason, and no I'm not kidding xP). Scores is half the reason why there's so much more noticeable AI slop when you start looking through a lot of pony pics.

My steup is: RTX3070 8GB and 32GB RAM and I generate first-pass image in about 20s.

Ksampler settings: 20 steps (10 steps for second-pass/hiresfix), CFG 4, dpmpp2m karras

When generating I'm using whole GPU VRAM memory and more than half sometimes even 3/4 RAM system memory

"I am generating on a PC environment similar to yours, and it takes about 22 seconds to generate one 800x1200 size image. I am using Forge. If your performance is close to this, it is probably normal.

Your RAM is too small,recommend at least 32Gb

As morenatsu wrote, you could use more RAM. Especially operations like changing models or adding/removing LORA will take much longer with little RAM.

If you use an addtional lora and get slower speed, you should go to setting and enable "
Lora/Networks: use old method that takes longer when you have multiple Loras active and produces same results as kohya-ss/sd-webui-additional-networks extension"

It uses danbooru tags, which has most of the same tags, though there's a few differences here and there I think. I don't know any offhand that are gelbooru trained specifically, if I'm ever not poor maybe I'll start training them into something though lol x3

@LemonSparkle there are artists who are not available on Danbooru but are available on Gelbooru. So, when I tried to use the tags, they didn't work. Gelbooru also offers more freedom with most of the artists banned on Danbooru. I hope the model will get trained using Gelbooru tags

@monicalucci Well, I'll put that into my basket of ideas, so if I can ever afford a new computer to train on maybe I can do it, I've already saved up half a million images+ from the boorus, gel, and others, some that don't exist anymore lol xP

@monicalucci Gelbooru does have a lot more artist, but I think checkpoints avoid using Gelbooru due to being less moderated + allowing realism and more open to 3d. Danbooru rules/moderating tags and posts just helps a lot with keeping a better dataset. Would 100% try out a Gelbooru checkpoint though.

@monicalucci Same, gelbooru is superior, you also don't need to pay for membership to view certain tags like on danbooru, I don't understand why danbooru is used so much.

@AIexperienceguy gelbooru sucks

this is the base model used to train lora models, if you use other merges your lora can get overtrained easily.

Probably not because they leaked this one 3 days before the official release. Such a shame, but at least we still good NoobAI-XL project, Still I think Illustrious v0.1 is quite good by itself.

@norim I'm completely out of the loop. What's the story about leaks and no further updates?

@norim so no more further updates?

I'd recommend just using other Illustrious-based models. Similar to Pony, I've never been able to get any good results from the base model.

Hi, the model does not have a base artstyle. so it encourage people to experiment with different art styles, you just have to write at least one tag art style to get good results. Or you can mix up to 13 artist names without breaking the image. It does not even need a VAE. It should recognize and mimic the style of at least most artist on Danbooru with 300 pics or more (dont expect much from those with below 200 pics). Clip skip at 2 and you might want to lower the CFG scale like between 5 or 3 but I have got good results even 7,5 (but they might be a bit burned)

You absolutely want to use a Illustrious-based model instead of Illustrious 0.1. The 0.1 is a great technical foundation, as I understand it, but is woefully unoptimized. There are, however, a lot of really refined models based on Illustrious you can use instead.

The base Illustrious model requires you to define a art style.
Use another model based on IL to avoid this headache,
like https://civitai.com/models/827184?modelVersionId=1283437

Have you tried giving it an artist style tag?
This Model requires you to specify an art style.

otherwise, you can use NoobaiCyberFix, which is what I use.

Illustrious LORAs usually use other models based on Illustrious, like WAI NSFW Illustrious

no,or you can try noobai-xl

Go for NoobAI if you want E621 characters+ artists from that site, this model is just for anime.

From what I've seen this is a major improvement over Pony, just be aware that you have to prompt as if you're doing a search and not as if you're prompting an ai.

@Immac Interesting, yea ive been doing research on it and played around with it a bit. however idk what im doing wrong. my pony prompts lead to much better image generation, its something i think i will have to play with.

what do you mean prompt like a search? ive see some prompting from sample images like this

masterpiece, best quality, <lora:TCV_illusXL_Incrs_v3:1.0>, 1girl, red eyes, long hair, black hair, shirt, shorts, confident, outdoors, tan, hand on own hip, (realistic:0.4)

this one came out really nice but is prompted like i would my pony

Yeah in the same boat, Don't see flux anymore I'm waiting to for some loras I like to work with illustrious before I compare.

@Deterna Ive spent the better part if a day playing with illustrious. And i can say it’s probably the best so far out of the box almost no loras needed maybe a character. It just listens to the prompt. Its actually quite amazing. But i would assume its checkpoint dependant. Idk if i just found the cream of the crop or what.

@jarmankell987 I'd like to check Illustrious out, can you recommend a checkpoint you're using?

@Dzban In general WAI-NSFW, NTR, Prefectius and Nova are the big ones for me.

@Dzban Well yea. Are you wanting NSFW? Because i would be willing to share some generation data to get you started. And i even if not you want sfw i can recommend a prompt. But the checkpoint is NSFW

@Dzban It’s literally the WAI the guy above me mentioned.

This is a lot better but still has some flaws like everything is so bright/shiny. I highly recommend using artist tags to get around this.

@notmyproblem324 yeah I noticed that and even with the checkpoint I'm using I'm seeing ppl use tags like shiny skin and wet skin. It's plenty wet & shiny without those tags lol

@jarmankell987 Just stop using "masterpiece"

@jarmankell987 imagine you're looking for a picture in a booru, you wouldn't search for "master piece" or "absurdres", you would search exactly what you're looking for, and try omiting as much as possible, for example, "long hair, t-shirt, skirt, solo" certainly implies 1girl, so don't prompt 1girl. as for negatives, leave it empty or just use "worst quality". Also you need to find an art style you like, that part might be biggest issue to overcome to start with. As an addendum, this is a base model, think of it like SDXL base, so look for refined models with an art style you like, that might be easier.

yes. IL is a massive improvement from Pony.
I recommend to use WAI-NSFW-illustrious-SDXL v9.0
or Noobaicyberfix

I’ve been away for a while as well, came back, and something is definitely wrong.

1.1 coming in 3.5 weeks, and 1.2 soon after. They are just milking it by releasing shittier versions first. And trying to get you to pay for each of em

Oh, no, but they require to unload prompts! (That's sarcasm based on some discussions)

They want to monetize the model, it seems.

Any recommendations? I can't generate anything even decent with Noob, I'm interested in opinions, since seems I won't be able to train on Illustrious no more due licensing.

What that means? No more 0.1 finetune?

@damaribryen693 It currently means nothing, as there's still no legal precedent for actually enforcing any AI model licenses.

@damaribryen693 First of all even if they could enforce the license, you can't just retroactively relicense something after you've had it in the wild. Everyone who obtained the model before the new license was added is still bound by the status quo before it.

The ToS seems entirely innocuous, effectively granting you ownership rights to all generated content.

set clip skip to 2, google how to do it in the software you use

@disastinator how do you do it on civitai generator?

@bougyakumahou comfyui is more complicated but have more detailed controls.

What licensing thing?

try to avoid the word "general", it's considered an age rating tag (meaning no NSFW). Also try putting "censor" in negative
also try RouWei, a retrain that targets both sfw and nsfw while prioritizing prompt adherence

Forgot to update, but the issue resolved itself when i tried again a week later. Honestly have no idea why lol.

You really posted this everywhere, didn't you? All the Illustrious models on Civitai, Tensor.art, Reddit.

@ClownReplica Well, apparently it's a scam and he's butthurt about it, what did you expect?

@herkerp123759 I'm gonna withhold judgement on that for a few more hours.

@ClownReplica Yes, because I want people and developers to pay attention to this. I want to bring this to the attention of the developers so that they know that there are people who paid for the model and are waiting for it to be released right now. I also want people to pay attention so that they go to the developers' website, read the phrase "We promise to open model weights instantly when reaching a certain stardust level (The stardust % can go above 100%)" and see that Illustrious v3.5 v-pread has already reached this level. The more people know that this model should be released now, the more problems the developers will have if they don't do it. Knowing this, I hope that the developers will be more motivated to keep their word and release the model as soon as possible, as they promised

@ClownReplica So, when people with all the money, scam others, we should just shut up about it? If it was a brokie scamming others out of necessity, you'd probably be screaming from the rooftops about it, and crying into your 10$ starbucks coffee..

@MindInTheDigits Nobody should be paying for models, there''s plenty of free ones for local use. Even Flux, and qaunt versions of flux for lower end PC's. The only sense in which capitalism still works, is that if people are willing to pay for something, it will get a pricetag. So, keep dem pricetags off our models, don't support people that sell them.

It runs fine on A1111, though sometimes the wrong combination of negatives makes it produce bad output

i'm trying too and i seach some tips to improuve my works.

I think they are trying to say you can do things like collect ad revenue but to not paywall with it and to share it with others rather than horde models for yourself.

even nAI and pony (partially) work with Illustrious, seems like you are doing something wrong. try another Illustrious Checkpoint like https://civitai.com/models/835578/pasanctuary-sdxl-illustriousxl

@iJWiTGS8 yeah I tried another IL checkpoint and it works. It's just this one that doesn't.

you're missing a lot fren

noob

same here. the results are just crap. probably I'm doing anything wrong. all the other checkpoints I've tried work great except this. nevermind

this one is best for lora training. stuff generated with it is subpar more often than not, but loras trained on it work great on other illustrious checkpoints.

Exactly the same problem, dunno what happened

Was this on the Civitai onsite gen?

@Honest_Abe yes, all my images are done onsite

You meant "Base Illustrious"?

train on this one for maximum compatibility, then, use whatever illustrious model you want (results may vary because each model has an inherent style).

Of course, Illustrious XL 0.1 was made a long ago, waiNSFWIllustrous got a new update a few days ago

this is the base model you fucking idiot

@moratono99980 wai and wainsfw is the same model, just an older name.

this gives, I'm sticking to being a catholic instead of a christian, vibes.

@Meowzilla 3 months later and there's no finetunes on the site.
guess you were the fucking idiot.

@IHateThisSite 

Unfortunately Meowzilla is correct.

All Illustrious models are weighted versions of this model. This model will most likely forever be the most versatile, other models are specialized in what they do and give up their ability to do other image types easily in exchange for a better ability to do one thing specifically.

I personally prefer Hassaku and Plant Milk.

@IHateThisSite there are? every "illustrious" model is a fine-tune of this.

But what about L-A-X and NoobAI team?

2

none. SDXL does not use clip skip.