They got freaking flux man, why push realistic stuff on noob?

Still doesn't know Burnice...

nope. i tried generate shorekeeper and it's not shorekeeper tho

@RavirKun try shorekeeper (wuthering waves)

Illustrious based loras compatible with eps one, but not fully compatible with v-pred one I remember

Depends on lora and noobai version

i haven't met the problem your mentioned on generating, 0.9 can just use easily with default workflow (1024*1536 30steps euler AYSscheduler), no oversaturation, no colorful filter. Maybe your should try different settings?

And i think the LoRA trained on Epsilon can use on v-pred simply, instead of V's training, there's no bias on the final image.

You shouldn't expect lora training to require the same settings with vpred as with epsilon, it's an entirely different noise prediction type. Noobai didn't invent vpred, there's no reason they should need to tell you how to train your loras with it, even if it'd be a nice guide to have. It's a learning process and you'll either have to learn new training parameters yourself or wait for others to share their experiences and knowledge.

That aside I agree it's gotten increasingly fried and I'm guessing they pushed the LR too hard or without ideal testing because they wanted to fit too much into their limited compute timeframe.. eh, I'll just use the earlier vpred models I liked my results on better.

0.9 becomes very bad...I am afraid that adding the real person part has obviously polluted the model. The LORA trained with 0.75S is okay, but the LORA trained with 0.9 is basically very bad.

They should make the realism part into a separate model, like how Pony has Pony Realism

@And233 No, Epsilon LoRAs don't work on V-pred model. It is real to make pic ok but only with rescale cfg / latent modifier.

@hasoo lora training is not so far from finetune. So I hope they can explain what they changed for their v-pred in comparison with epsilon.

@Enigmata don't use lycoris, the standard LoRA train on epsilon can use directly on v-pred, I have used that for a month.

they just mean to say that the full release aka 1.0 is coming

That's a question you should be asking at the community of whatever program you're using to run this model

there is no base 0.90

如果不正確啟用v预测,會直接鬼影

LORA則不需要特別處理相關參數,模型能正常生圖就可以

決定生成是否使用V預測的是模型,而不是LORA
而V預測模型也可以使用非V預測訓練的LORA

@NTR_BLACK 原来是这样，非常感谢！

I'm also using eps 1.1 in Forge, I'm really amazed by the quality, but for some reason I can't get consistent result when I use a lora, every seed gives me a slightly different artstyle, this never happened with pony, does this happen to you?

its not hard to train lora for v-pred

just disable noise offset, min snr gamma and multires noise

and add --v_parameterization --zero_terminal_snr --scale_v_pred_loss_like_noise_pred --debiased_estimation_loss

@Hugs288 I train with Onetrainer. Not interested in messing with kohya

@wayva no. but this sounds like a problem I used to have wayy before NoobAI came along 

@wktra You can research the relevant training parameters on your own (onetrainer ui does not support but still supports the above commands), or you can ask the author of onetrainer to feed you.

also back image generated

solved, the issues is caused by neg prompt

Danbooru, I guess. The inverted faces appear much more often in drawings than on photos, so the model retains facial structure much better in odd poses because of that.

Practically any anime model can do this

You can try my 3d art style lora. It has realistic influence, Especially for girls from RE, The Witcher, mass effect and other franchises.

@Enigmata Thanks, I'll check it out! I am looking for something cosplay-like as the R-version shows in the preview. A good 3d lora is never bad to have though!

P.S. I see you have plenty of NoobAI-XL loras, you might wanna mark them as Illustrious instead of SDXL since NoobAI-XL is Illustrious-based finetune.

You can try my 1.0 detailer. In case no masterpiece etc or loras or artist used it shifted style heavily towards realism, but results were meh. On the other i never really dug that way

@average_ai_enjoyer I'm very interested too. Lots of the realism mixes will throw off outputs badly, but maybe the developer can do it without altering the essence of the checkpoint.

Still massive asian bias and very bad anatomy. I am not asking for Flux.1D level of performance, but that's worse than SD1.5 IMO. This is how a european flight attendant is supposed to look like according to the 0.9R with your prompting style... https://imgur.com/QR5Jz19

I am not sure what's the problem, could be forgetting photography altogether, and in this case some base model remerge would help a bit, or it could be bad annotations for the photo training, but it surely doesn't respond to tags like lightroom or photographer names or genres, and it doesn't sample good images in general for stuff like mixed media.

@Erilaz Dude it's based on cosplayers, 99% of professional cosers are either chinese or korean.

@undef01 No. That's because of limited dataset and bad annotations, or forgetting the real faces to begin with, I'll die on this hill happily.

Most of fashion photography is female models, and yet you can prompt any half decent photo model to output a man, it will do just that. Most of animal photos are cats and dogs, but models can do other animals no problem. So why should that be a problem here?

Cosplay isn't exclusive to Asia, there are lots cosplayers from other places, and there are tons of characters who designed with western look in mind. Safe to say, if a model can't account for that, it's a flaw, and this model doesn't respond at all. I am confident in that because the regular model does respond well to the ethnicity tokens when you ask for painterly output.

I am not doing my critique because I want to roast the devs, I actually like NAIXL models a lot. I am doing that because I know this can be done better. Illustrious might not be the best base for this task, but people managed to get it working with SD1.5 (NED) and even Pony fine tunes, so it worth trying to leverage that experience.

As a side note on the prompting,

we shouldn't use photo-referenced tag in the prompt. It introduces a major contradiction because on danbooru, that tag represents drawings using a photo as a reference. So it's for drawings, not actual photos.

photo \(medium\) SHOULD suffice, but all the half-assed sketches scanned with a phone have be pruned from the fine tuning dataset for that to work better.

@Erilaz Test the photo-referenced tag. It does look worse in the other sections, but it helps cleanup tags if left at the end in the section for descriptive tags. I think the dataset could be enhanced with more hand-tagged cosplay photos, but its a delicate balance to keep the tags vs enhancing realism and anatomy representation. For example, hands kind of work ok, but certainly not at Flux level, whereas ears just wont work unless you mess with tags, like making elf ears or something very specific, to give the model a shape to aim for from the beginning.

Any damage introducing a properly tagged realism dataset does is a sign the model has been taught something contradictory and was hiding it.

I agree that faces trend toward asian, but its a very soft bias that it can be judged from.

@Erilaz Dude do you follow actual "professional" cosers? Actual western cosers are the bottom of barrel of quality not to disqualify western cosplayers, but most actual professional individuals lean into being a professional model, that's how it's in the west, we don't get a market like in japan/korea/china where professional cosplayers are an actual job, western cosplayers of quality are a very few number the general cosplaying scene in the west are simply "bad" to "horrible" in matter of quality because most are just hobbyist enjoying themselves that's why cosplaying sets are easier found for asian individuals.

@deepvision I think negative prompt is much better suited for solving the signature issues than a tag that is supposed to have a huge stylistic effect. As of the Asian bias, idk... For me it's significant, because it's uncontrollable and noticeable enough. Not an issue in most cases, but definitely worth mentioning and fixing. Which brings me to another point.

@undef01 Why should we even care about the industry? The original (non R) model understands the concept of ethnicity rather well, the same applies to the non-photoreal images with this model, but as soon as we ask for any real faces, this model doesn't follow the prompt at all. Ciri turns out to be Korean, Princess Jasmine turns out to be Chinese, and Loid Forger is a freaking feminine weeb! That's a straight downgrade for the same reason nobody expects to see Ryan Gosling's face when you ask a model to generate Martin Luther King.

I am not advocating for removing the asians from the dataset, far from it. I merely saying that it should be expanded and annotated better so the model starts to activate the weights associated with different ethnicities as well. We know different faces can be sampled because it works for traditional media styles in the very same model. Wider diversity in cosplayers, better annotations and more photo media training should help with the issue. You can use danbooru tags to describe any photo, always make sure photo media tag isn't forgotten and start training. No need to do just the cosplayers. In fact, I am not sure why cosplay is the main trigger here, when it should be photo media instead.

@Erilaz "Why should we even care about the industry?"
It's not quite "caring about the industry" but how the actual state of cosplaying works, the dataset in the model is based on asians/leans towards asians simply because it's the largest number of photos of professional cosplaying that can be found, which happens to have the highest quality, that's unfortunately how cosplaying goes, if the dataset did not just focus on cosplaying but also any professional model/instagram influencer then it would be able to do western faces too, but like you already know the model focus only on cosplayers photos in this version, so you won't get references of westerners in the dataset.

Yes, SDXL and PDXL Loras work with Noob, however their strength is variable and won't be the same as if you were to use the Lora with the right model. Not to mention this only really counts for the EPS models, as Vpred Loras require different training to work.

Imagine SDXL as a foundational coin, representing a core dimension of value. The chains, like Pony and Noob, branch off from SDXL. Over time, as these chains connect with more models, they grow stronger and more complex. Eventually, they evolve to hold significant value, becoming some of the most valuable coins within the ecosystem.

V Pred 0.75s or EPS 1.1

Epsilon 1.1 seems to be the best mix of quality and concept understanding (artists/poses/clothes...etc.)

The v-pred models are good but have some issues in terms of image quality. Maybe v-pred 1.0 will solve the issues.

v-pred is more strictly adherent to your prompt. it follows it much more literally. you didn't describe a background? you don't get a background. epsilon is more creative for things like that.

@fizalpher I do describe the background, but at times I will only get one part, ex: I will write "beautiful garden full of roses" and I will get rose on the feet of the character and everything else a plain color. But not in all generations of course, but it will happen.

@darionk The model doesn't use natural language, never has. It'll give you garden and roses, not a garden full of roses, unless you happen to get lucky with the seed.

@upscaleanon537 it is sdxl based in the core, so it does. Especially short ones. But it does not understand "full" for example, you should raise weight of flowers to get more of them. Or find corresponding tag like park.

0.6 is the best one

eps 1.1 imo

seconding vpred 0.6. Epscels are blind

0.9R or eps 1.1. 0.9r has far more training than eps and produces better quality but some people like the creativity of eps more.

Version 1.0 is purely an anime model, while 0.9R was just a test by the developer.

I feel like the overall quality of images improved a lot because of it, even going strictly for non-realistic stuff.

Prompt order has a crazy impact, especially the “1boy, (x) character, (x) series”. I’m actually doing some testing to find the best order for images containing more than one character right now as simply moving a tag like “trio” one spot changes results and adherence. Feel free to look through my latest posts for some basic prompts, more to come soon.

You don't seem to understand the "potential" of the model.

It's true that you need to write a lot of prompts to use this model as is, but by showing the direction of noise removal with LoRA, you can easily create beautiful illustrations.

If you don't mind using LoRA, I hope you will refer to my recent post.

I use Forge and I don't do anything more than the prompts.

For example, you can create a perfect image just by loading a few LoRAs and the recommended prompts (+1girl).

By the way, almost all LoRAs are published on civitai, and the rest are my own original LoRAs trained on my own SD1.5 images.

Prompt order in noob in particular is more relevant than ever it seems, obviously earlier prompts always have had higher impact, but now it really is more of a wild difference. That being said, the only thing that matters is prompting in the order of importance and impact of the tokens. In my experience with any model really, but particularly these ones, well known characters goes waaaaaay back in the prompt if you're trying to go for anything more advanced than "whatever picture it ends up being of the character", like specific poses and concepts and stuff.

@mega2ton2 "You don't seem to understand the "potential" of the model."

No, i dont. What gave it away, that i tested only 2 days and had no idea about it before? Seeing images that so far dont reach quality of other XL based models? Couldnt be that. Ok but seriously, idk what people expect from beginners, to just know how to perfectly use a work in progress-model this complicated, with every released version being different? Noone so far has given me a detailed description of what exactly the "potential" of this v-pred model would be, ive just heard keywords like "better colors, lighting" which doesnt tell me really anything, since IL and even pony models ive used so far had absolutely no problems with those things either. If you could give me like an explanation what EXACTLY v-pred does over all other models so far, id be actually truly grateful. Because obviously i WANT to use this model and try out what its capable of, i dont go through all this trouble for fun... well, actually i do, but the fun part comes later. But i, and others, need help and explanations to get there. The description of this model doesnt really have a "This is why you should use this model with the new v-pred method over all others, and this is what it does!" -section.

Also i do seem to have misunderstood the role of this model a bit, i thought especially this one was supposed to be able to rely on loras less? Id have to find it again, but didnt the creator say somewhere that loras would have to be trained in a different way then before for this model? At least this article has following quote:

https://civitai.com/articles/8962/noobai-xl-quick-guide

"It uses the complete Danbooru and e621 datasets and should be able to complete generation tasks excellently with as few lora models as possible."

Every person using it so far doesnt seem to be able to generate much of anything without loras, not even the article as you can see further below! So where is that quote coming from, if i see people use loras, and now get told to use loras?

Almost noone seems wanting to be critical about this model. I do, because i care. I want this model to become as good as its aiming to. And you know the only way it will get there? Critique. If everyone just tells them "Oh this is perfect, so good!", they wouldnt have any incentive to improve. I cant wrap my head around how many people dont seem to understand this simple little concept. Praising the things that are already working is good, and so is mentioning the things that dont work yet.

To the other 2, thank you already, i had a bit of time to try moving prompts around, mainly the "quality tags" from the front to the back, and i think i already saw a difference. Now i cant say 100% how much, since i have to do a lot of other things to make images even acceptable, but its very clear that it does something in general.

What i want to mention is, i can test only very, very slowly, because i use AMD. (The story about that would literally, and i mean literally, fill a multi-page article, what ive tried already, learned so far, what alternatives ive looked up, what i plan to do, etc. Until early next year, i wont be able to change this.) One pic at 832x1216 (like recommended), at 30 steps, which i also still test around with, takes like 3-4min. Which is acceptable for me, but compared to the what, few seconds any nvidia card needs by now? Ye, testing takes some time. It will probably take so long for me to get results, that the v1.0 of the v-pred will come out, so my plan now is to learn as much basic things as i can, play around a bit, but only really try in-depth things with that full model, since i expect that to already be quite a bit different to the other versions we have rn.

What i got so far is, slide around prompts, somewhere ive read to alternate using tags from danbooru and e6, cause for whatever reason for someone that seemed to increase the prompt understanding, and use loras. Id be very thankful for any general lora recommendations, like quality wise, what do you all use? Ive tried a few based on XL and pony, as some work, somewhat, but im sure some people have already found some that work quite good.

Lastly, it might be more difficult for me to make the images i want, because i only make furry art. Which this model can do. But just as any other model, nothing of that will never be as good as a simple "1girl" image. So all the people having no problems making those... ye, i believe you lol. But one big positive i can already mention is, artist styles do work perfectly without loras, just from a single name prompt, even all furry artists ive tried. The pics i can make arent good yet, but i can see instantly that it fully got the style i wanted. This so far is the biggest plus for this model imo, and i really hope i can get it to work much better later.

It might become obvious how much i care about this model already by how damn much i write.

@Bigboyblaziken 

It's important to try out different things.

I hope you'll consider it. Have fun.

@mega2ton2 I will never understand the potential of the model if noone tells me.

... Do you understand it? Again, id love to know, but so far noone has given me a reason to believe they know. Still no cohesive explanation has been shared, by anyone.

@mega2ton2 "You don't seem to understand the "potential" of the model."
This was a little arrogant of you to say, considering his first post pretty much said he underestimated the model. He even asked for tips and tricks to get better at working with NoobAI.
Your post contributed nothing new and made you look like a 🤡

I just wanted to tell you that noobai is a model with potential, but I'm sorry if my English is poor and it sounded like I was mocking you. That was not my intention.

I hope you get what you want.

@mega2ton2 Ok fine, but you know what i mean yes? Its confusing to hear from someone that i dont know the elusive potential of something, without that person knowing the potential either, not mentioning any information on it.

So i just have to use loras to get acceptable images, im fine with that, i needed them for pony too most times to get good images. Its just not what my expectations where at when hearing from other people how good THIS MODEL is. This might be a me issue, but how good could you judge a models performance while using 3 loras with it? Im testing the v0.9R model right now with 3 detailer loras, and yes... the image gets a lot more details with them lmao. But still not as many as other models would while using them too.

If you need to use loras for acceptable results (so far, maybe not later anymore), then please, i would love to read that in the description, or anywhere for that matter. "able to complete generation tasks excellently with as few lora models as possible" I might have misunderstood this then, i guess it basically is saying "You have to use loras!"

I really want to stress again, that i have nothing against this model. Otherwise, i clearly wouldnt be here, and wouldnt write entire articles of what i think. I WANT this model to succeed, want to use it, want to get good with it, but that will take more time. And from what ive seen, that will take more time for everyone else too, even the creator. I can see the problems it has, but i also can see some strengths already as i mentioned, strengths that CAN make it better then other models, ive just not yet seen all the strengths other people mention. I just got to be honest here, so far i do not see the better lighting, better colors, better composition, compared to honestly any other XL based model. Not from what im able to do myself, and not from one single image in the galleries of any version of this model, even with all the loras they use. That doesnt mean it cant be better overall, it just doesnt show it yet, or noone knows how to use it fully. The main thing i dont understand is, that some people claim they can make amazing images with this model already, and i dont see it. Do they just not post those amazing images? I have ALL filters off, i would see those posts.

Before you show me any pics in the gallery now to tell me theyre amazing, if theyre older then a day (and younger then like... a year), ive seen them. I originally really didnt want to directly compare another model to this, but i want to provide a source showing what i mean. The showcase pics of this model are hires upscaled, as the creator mentions, and some very likely do use adetailer, so theyre not pure either, but they dont seem to use any loras. No quality loras even, not even embeddings. And already those pics for instance DO kinda blow me away.

https://civitai.com/models/827184?modelVersionId=1183765

Again again again, this doesnt mean the model IS better. It just means that right now, the pictures that model can make LOOK better. And for people that cannot get this model to work even remotely close to that... it doesnt matter how strong this model itself is, i can totally understand them using the "weaker" model anyways. I can tell you some weaknesses of that one right now too, it cant do styles well, has strong pre-determined styles, likes to replace furry characters with humans, is very female biased, and more. But just look at the posts anyways.

"They used adetailer!" Most probably yes, just like in the galleries of this model. "They used inpainting!" Yep, just like at least some pictures here, which still dont look anywhere as good. "They used loras!" Yup, almost every post, just like almost every post here. No matter the excuse, nothing stopped people from using all those tools with this model too, they all work. Loras more or less with noob, but thats also the case there, almost none of those loras are made specifically for IL.

All i want tbh, is for one person claiming they can make amazing pics with noob, with any tools of their choice, to MAKE one with it that can reach that level of quality. I dont even need to be convinced, im already sold on noob, the ability to actually use the styles you prompt is unmatched. As soon as i can use it correctly and some things are fixed, i will use it. I just want people to either back up their claims, or admit that they cant make pics like that yet.

@Bigboyblaziken +1. You dont really need loras outside shifting base coloring. Of course there are certain bloated styles or concepts like popped collar that dont really work, but you can do that with basic inpainting. Check my guides for general settings and useful extensions on Forge. And gallery for inspiration ;)

@Volnovik Hi, alright thanks, ill send you a dm so dont fill this up even more.

bad prompt of bad loras used. I see a lot of people use xl detailer and it just slaughters hands. Also you can add artist tag that has a lot of good hands in portfolio and it will improve your image.

Not possible, v-pred would always work differently from epsilon - that's the whole point. A1111 and ReForge have their own ways of using it through dev branches.

I suggest moving to Forge since there are almost no downsides from switching from A1111 except dang hash calculation. It even is as stale as A1111 for past couple of month =)

@munchkin I see, guess I will stick to epsilon for now until changing to forge thank you for the clarification

@Volnovik didn’t switch to forge because I have zero information about how it works or even how to download it not like A111 which i’m used to it but after seeing that forge is better i will switch to it after some learning

@1wkk It is basically the same now, with few added quirks for flux. I advise you to start with StabilityMatrix and check author's guides for basics and my guides for specific parameters and extensions.

@Volnovik I’ll probably go and check on youtube this stability matrix seems complicated for me but saw your guides and will try the settings you mentioned thank you for the help

@1wkk Stability matrix is just to install UI's and some models and loras. It makes life waaaaay easier since it does all venv and tricky stuff for you.

@1wkk https://civitai.com/articles/7492/ai-image-generation-for-complete-newbies

Btw, maybe this guide will also help you

@Volnovik oh it has a guide on using stability matrix, ok following the guide I started downloading stablediffuision webui forge I think after that it should be like the normal A111 also really thanks for your help !

Bar censor, mosaic censor, pixelated, inpainting

use the correct tags of danbooru for censored stuffs

Try adding some of these to your negatives:

https://danbooru.donmai.us/wiki_pages/tag_group%3Acensorship

Could be loras messing it up. Do you have "uncensored" in your prompt? Add it near the end of your prompt if you don't have it.

Also check if you have "nsfw" or "sensitive" in your negative promt. Remove those if they are in there.

@valazor I do have "uncensored" in the prompt. I also have "NSFW" and swap between "rating_sensitive" and "rating_explicit" also in the positive prompt. there's nothing in the negative prompt that should attempt to censor the image. I even have things like "safe, safe for work, censored, bar censor, mosaic censoring," etc. in the negative but I'm still getting small artifacts

@valazor I think I narrowed it down to the sampler causing the artifacts. I like to use UniPC with certain style mixes but it generates artifacts and a weird grainy texture that underlays the whole image. but none of the other samplers handle the style as well as UniPC does T_T

@Digitalize Yeah noobai is a bit finicky when it comes to samplers :/

Glad you managed to figure out what the issue was.

Thank you, could U pls post a sample ?

it means "arti waifu diffusion model", it works like the score in pony you can check it in this document https://d0xb9r3fg5h.feishu.cn/docx/YpOQdtHTDoetcZxIO9fc33onnee

AWA

@derpmagician Thank you so much!