@L_A_X So you don't want to set your v-pred model versions to Illustrious since Illustrious LoRAs aren't compatible? Are usual SDXL LoRAs compatible with it? And what about your popular Epsilon-pred 1.0 version?
The general problem I see is that NoobAI-XL follows the general generation rules of Illustrious with Danbooru tags ect. So either it is based on Illustrious with continued training (which then should be stated as Illustrious in the model card) or it is only following a similar training as Illustrious-XL did but is otherwise completely independent. In the second case, it should get it's own tag on CivitAI. It surely does not seem to be based on the SDXL 1.0 base model...

@lizardon1024 They're saying that the two models are structurally different, due to using different noise prediction. Epsilon is standard SDXL noise range, v_prediction is capable of nearly absolute blacks and white ranges and denser noise at the training stage. The compatibility issue isn't simply captioning or tag based, because Illustrious does not use v_prediction in its training, so it's technically a different model due to standard SDXL training with epsilon. The NOISE is different. That introduces it's own issues when layering noise at the point that the image is being decoded at the latent level. That being said, I have gotten some epsilon based SDXL 1.0, Illustrious and Pony Loras to sort of work nice with v_pred trained models. But it's hit or miss and a lot of trial and error tuning the sigma values.

@mewtsy Thank you for clarifying this. So I see the LoRA compatibility problem (at least for v-pred), but setting it to SDXL 1.0 isn't a good solution either. Following your explanations it is not compatible with SDXL LoRAs and it makes it way harder to look for NoobAI fine tuned models. It will also confuse a lot of people who usually use natural language for SDXL instead of danbooru tags.

@L_A_X Make a new model page for the version that is incompatible, then.

From the name and what it seems to aim for, it's attempting to equal high quality generations that can be done without needing to know lots of technical things or needing a book-length prompt for a decent image.

Supposedly it is safe for newbies using image gen AI to use without hiccups, though some past models were pretty plagued with setup issues (Specific samplers needing to be used, CFG needing to be lower, weighing different things like artists in your prompts if they came out "overdone" etc.)

But gladly the devs are actively working it out so far through these Vpred releases, which is easier to use and needs a little setup but it doesn't have you fighting your prompt as much as Epsilon it seems.

Didn't notice that, thanks for sharing!

Can you provide some examples? Thank you very much

It's weird because most generations look great with 0.6 but every now and then you get one that has a red filter over it like the CFG is set to 100 lol.

But it only happens once every 20 generations so it's not a huge issue. Other than that I think 0.6 is an improvement on 0.5 in pretty much every way.

@zeseren eyes seem same or better on average. maybe sakanakochan has wrong sampler/scheduler setup or something. & ihaven't seen red filter every 20 gens either... maybe something up with your neg prompt or settings?

@valazor I have that issue in v0.5 too, look at the CFG. If you do some testing with the CFG you'll see that increasing the CFG will turn things red QUICKLY. From what I'm testing, small steps in CFG change wildly the image, so keep every sampling at the same CFG and use EULER as it's the most stable sampler.

@Daru_22 I know, I even tried a CFG of 3 and it still happens sometimes. Most images look fine, but every now and then some image looks horrible like it's been burned lol.

@valazor 3 isn't in accepted range... only 4-5 and the closer i got to 5 the more i disliked it, 4.2 is what I use and i have never had this issue. try copying the workflow off one of my recent posts. Euler sampler, beta scheduler on 4.2 cfg and 35 steps.

@fizalpher I usually use 4 but the acceptable range is 3.5-5. I'm just trying to say that this is not a CFG problem, it's something weird with either the model itself or reforge randomy messing up.

@valazor according to official (them, literally on this page and the english guide) it's 4-5.5, however on this page it's 4-5 and i find that is a better guideline. even small .1 steps change drastically, you aren't supposed to go under 4 for this model. And again, I have never once run into this issue so it's something on your end/your configuration. Instead of ReForge i'd recommend ComfyUI as you have a lot more control over your nodes.

@fizalpher It says "The CFG coefficient is recommended to be between 3-5.5" in the english guide (https://civitai.com/articles/8962)

But yeah, maybe I'll try comfyui instead.

@Flange Hi, for any style LoRAs it's better not to use any trigger tag. But it's better not forget about turning on shuffle captions parameter. But if you really need a style tag, just use "XXX", not "by XXX" or "artist: XXX".

@TroubleDarkness Thank you very much! I will try it.

@TroubleDarkness Just because I noticed that the recommended style tag in Noob's guidebook is "artist:XXX", so i have this question about how to deal with the tag during training.

@Flange Actually, info about proper artist tags is extremely unclear, older version says just "name", guidebook says "artist: name", anyway. you can just check the prompts with both "artist: name" and "name". constructions. You'll find they're working pretty much the same way. IMHO, single artist tag without "artist:" works better. So, there is 100% chance that it was trained just like "name" tag.

@TroubleDarkness I think LAX was/is a NAIV3 user so he probably just copies the mixes from there which uses the Artist:XXX syntax. It's not necessary in Noob since it's trained with just name but from my tests, it doesn't really matter, Noob or Illus picks it up just the same.

it would make absolute sense for it to get its own tag, but that's up to civit staff

I guess no answer is also an answer... NoobAI-XL is based on Illustrious-XL and therefore should credit it.

The back slashes "\" are used to escape a character. This is because the parenthesis character is used to trigger something else, and the backslash is to indicate that you want it only as text and not to trigger other actions. As for why you need parenthesis, it's likely the tag or character is ambiguous on its own and needs the text inside the parenthesis to differentiate it.

As you may know, anything in brackets in WebUI will be used with increased emphasis. For example (1girl) is something like ~(1girl:1.11), adding a backslash to the backet symbol just breaks the syntax not to trigger emphasis on tag. Originally, characters and other tags are trained without backslhashes, but nature of emphasis adjustments in the interface obliges you to use backslashes.

okay I get it, I think. I just saw it being used for the series name frequently which confused me.

@shubasaur877 It's just because character names are frequently used in pair with series name in Danbooru tagging system. Danbooru / Gelbooru / rule34 are imageboards for different themed media.

shimakaze (kancolle)
asuna (blue archieve)
etc.

Those franchise suffixes are used not to confuse anyone, because, obviously, there is a lot of characters with the same name or even name and surname at once.

Also, if you're new, don't use tags with underscores, like "asuna_\(blue_archieve\)". Instead of underscores use space while prompting for anything. So, this tag would look like "asuna \(blue archieve\)". There is few exceptions like score tags from a different model or specific artist nicknames, for example.

You may want to install booru tag autocomplete / autocompletion extension for WebUI. It can be found in extensions -> available tab or manuanny downloaded from a github page, using install from URL tab, if you're using A1111, Forge or ReForged WebUIs. You're writing "larg" in the prompt box and you're getting examples of existent tags, like "large breasts", "large bow", "large hat" etc. Extremely helpful for anyone, who's working on Danbooru/E621-based models which NoobAI is.

@TroubleDarkness thank you very much! i understand it completely now. that autocomplete feature seems very interesting however i am using swarm/comfyui, ill look into it, thanks again

@shubasaur877 I believe, SwarmUI has it's own built-in autocompletion feature. Or at least, there is a working extension for it. Not sure about UncomfyUI, nerver was a fan of this quite user-friendly interface.

@TroubleDarkness Thanks i found how to get it working, so much better now and this model is really great :D

@shubasaur877 Good to know.

您好，关于文档中提示词规范化那里，我其实一直有个疑问。就是当我使用“颜色+danburoo存在的提示词”的组合的时候，效果是否还能那么好？因为带了颜色的前缀，这个提示词本身就不是一个danburoo存在的提示词了。

另外，指南里只推荐了采样器，是否有推荐的调度器呢？

cool!

@yiliankexi 好问题，一般都是可以的。

模型简介里的“NOOBAI XL 快速指南”文档中提到，艺术家标签需要加artist:前缀，但你的文档里又说艺术家标签不需要添加任何前缀后缀或修饰，以哪个为准？？？

@1519190067765 不需要添加任何前缀。

快速指南应该没有提到需要添加前缀。

@Euge_ 快速指南中“总结”部分的注意事项第三点：提示词使用D站标签，艺术家标签需要添加artist:前缀。我测试的情况是加了artist:效果貌似稍微好一点，是错觉吗

@0xSeiunSky 感谢提醒！以用户手册为准，不需要添加。

其实因为没有直接更改标签文本，因此添加与否不一定影响效果，放心用~

me too

me three

Me too.comfyUI can work.

我也是

all I get is some strange noise, blurred image and oversaturated dots. Using A1111

same, i though my lora do it but nope, run it raw and still only jumbled noise... all other version work totally fine hmm

1.5 embeddings don't work with SDXL

what? you do realize this model is 100% different architecture, right? this is SDXL/Illustrious based, and you don't even need embeddings for 99% of concepts that any normal person would want to use, only if you're trying to use the absolute cutting edge newest anime char/artist that got popular in last 2 months.

我就觉得大概率不会，不可能一出什么新底膜就马上训练，特别是像flux和sd3.5一类的。sdxl经过几年的各种人的训练才出这样一类实用模型，没有稳定生态的情况下，这种行为是浪费资源的。况且这些模型对显卡要求高，就算有训练出模型，也只不过是少数人的玩具。

lora训练我个人是会使用艺术家名称直接训练的，效果确实会更好

比ponyXL在还原画风上的效果好其实很正常，因为模型本身就保有了对画师图片的训练，lora下的二次训练自然效果会好

据测试群大佬所说，sd3.5为dit架构，相比xl的unet训练所需的算力需要10倍。

champion's tunic \(zelda\)

In a word, label normalization is divided into two steps: (i) replacing underscores in labels with spaces and (ii) adding a backslash "\" before brackets.

first you also drop the underscores for spaces. next you escape the parenthesis with \ on the left side of each one.

disable all noise related things like noise offset, min snr gamma, multires_noise_discount, and add the follow args --v_parameterization --zero_terminal_snr --scale_v_pred_loss_like_noise_pred --debiased_estimation_loss, the rest of the settings you used to train on other loras like dim etc should be fine.

@Darkwen ok, tks, i'll give it a try

@Darkwen hey it worked, yet i don't know if my configs were wrong, but the Lora did not perform as good as in pony

@Darkwen what about huber schedule = snr?

@Nephilim good to know, what settings are you using? remember that illustrious, and this one specially is very different from pony, pony text encoder is burned to hell.

@Nephilim i dont think huber scheduler matter that much, i didn't change it, you should only use that args and check other settings like the optimizer settings. 

@Darkwen dm'ed you

@Nephilim huber is effectively reducting the effect of "crap in, crap out" effect of training (it's an alternative to L2 loss). the paper says it skip bad captionning of subject when it occur in favor of other subjects beter represented. but also it hurt accuracy an convergance.
by affecting it as a schedule, you can effectively dampen the effects of huber loss depending of what you want to train. more detail but more defect vs less details but less defect. depending on your dataset quality this might be relevant, especialy for finetunes which include automated dataset and poisoned pictures.

@Le_Fourbe I actually curate very well my datasets(it's just for LoRA's), so keeping it on must be a bad idea correct?

@Nephilim well it might or might not.
the schedule is by default fairly low to the point you "might" not notice. i didn't run comparision tests myself as i struggle with other parameters too. i keep it on for whatever case one poorly captionned picture creep in.
you might find value in curating just a bit less by using huber loss schedule and find a good good enough result.

Pony was never good... Without LORAs anyways. Pony is still a great LORA engine but it is by far the worst anime base model that exists... Even Illustrious V0.1 (the base of this model) is far better than pony IMO. Pony had a run of 1 month before getting beat by everything under the sun practically, especially novelAI diffusion and now this which surpasses NAIdv3.

@fizalpher 100% agreed

@fizalpher I'd say Pony was too heavily fine tuned so loras didn't work so well. Like if you prompted for realism, it was almost deterministic with the style. And creating realistic loras always converged to the weird 3d aesthetic.

With NoobAI, concepts doesn't converge so well to one path. That makes NoobAI feel more random and chaotic but same time it provides more space to express different concepts.

我的经验是，这个模型喜欢详尽细致的提示词，提示的越细致，对图像的框定就越趋于稳定。就是烂手太严重，多废图。

Just train a lora

@klikkeri1 2 years later since the release of SD 1.5 and still struggling with pole dance. No matter what lora you use. You can't make the model to do it properly. I tried it many times first hand, i saw people trying it many times second hand. And this is just a grain of sand on an entire beach of horny world. Some stuff just can't be done with a lora.

@Ringakiseki "can i fine-tune this model to be more nsfw focused, i want a gooner version."
you don't need to fine-tune the model, you can use lora

@Ringakiseki To be fair, there are so many possible positions in pole dancing

prompting issues

@RavirKun cum issur

@smugfh  thx after writing cum my prompts got better 👍

@Ocean_lake you are wellcum.. : >

牛逼

What is SMEA?

@anifibous That is for NovelAI, not for this module.

@WallBreakerNO4 sadly

Is this on v-pred or eps? v-pred sometimes has color errors on higher CFG values, which can be alleviated with CFG rescale or APG.

@momoura vpred. I use a CFG of 7, but I will try CFG rescale! Thanks for the suggestion. It seems like it happens after a few generations. I'm fascinated by Illustrious so far, other than the more difficult prompting vs. Pony. It's nice to finally have a model that didn't hamstring itself by obfuscating artist styles, etc.

@waifuliberator 7 CFG is way too high for the current v-pred checkpoint. I would recommend starting with lowering it to 4-5, then trying out the other options in tandem to reel in any weirdness. Rescale and APG are both available out of the box in reForge or as nodes in Comfy.

Some tags might make the subject(s) more lizard like. Like prompting for "raised hands" might render your waifu to a fish.

@momoura Thank you, I will try that!

Ah, I see why "muzzle flash" adds an actual muzzle to my character's mouth....XD

e621 tags are really well curated and consistent. Danbooru has tags like "voluptuous" "hourglass shape" and "curvaceous" that has no real difference in practice

only IL will work. and you cant use them here in civitai ONLINE generation because OP didnt want to put his model with Illustrious Category.

it's up to a1111's developers, model it self can't do that

@nipotto Ah, I see. That's unfortunate :(

It is, I addressed this earlier but op doesn't seem to care about it. Illustrious section exists for a reason. I believe OP is trying to hound the "SDXL" rank ladder instead as Illustrious doesn't have any rank section. Illustrious section bug has long been fixed yet OP still insists on putting it on SDXL section while all the LORA is mostly incompatible with normal SDXL.

This is a retarded argument that stems from a profound lack of model understanding. By the same logic we should rename the illustrious category to "kohaku", because that's the base for illustrious. The vast differences in training time and dataset compared to illustrious absolutely warrant a separate category and having it as base sdxl to at least communicate the architecture is perfectly reasonable until then.

@scythesaint99 kohaku isnt modified as much as this model. It's that simple, there is a reason why AnimagineXL didnt have its own category. It's not modified too much that majority of standard XL LORA is still compatible with it. Illustrious however, has been modified too much that NONE of SDXL lora will work properly anymore, heck the compatibility is worser than Pony is. Which is why Illustrious category exist. Why is it so hard to put it on Illustrious category. just click and done.

The whole category system is very arbitrary anyway. It should be like HuggingFace where models can branch off from any other models, eg. NoobAI Loras/Merges branch from NoobAI Checkpoint which itself branches from Illustrious and so on.

The more important aspect is that the v-pred model should be separated, as that's the one that is completely incompatible, while e-pred is compatible with not only other illustrious loras but even pony loras.

This is probably the only model I've used that imo is on par with NAI. Especially in terms of replicating artist styles, and for characters and details it's far superior. As you said, inpainting is the one area where NAI is still so much better... but that might just be a skill issue for me. I've always had trouble getting good results inpainting locally.

i would suggest exploring x and danbooru, find the images you like and use the artist tag,

There are some wildcards out there which contain lists of artist names.

Personally a fan of illustrators like Nagu or Ui

https://civitai.com/articles/8962 There is a Quick Guide with links.

https://docs.qq.com/sheet/DZGxRSXhvcmNmeHFv?tab=q5lnh5 Artist database with examples

Pony 不是不更新了，作者在用auraflow二次训练一个大模

@aquarium_pixal 我當然知道,但是auraflow大家都不看好,
我預計他訓練完也會發現效果不好,大概又準備轉向後備方案FLUX.1的中等版本,然而FLUX.1 的非收費版本要大量調整很困難(從作者的很多舉動上看,他有轉向收費的傾向,所以即使那個一百萬美金年收都可能會觸發法律問題),我是覺得今年都看不到PONY 7的測試版出來,而這半年就等同沒有任何更新---但好在PONYXL有大量愛好者持續推出微調版本

auraflow底模能力就那樣了, 完全開源模型當然好...但是代表訓練的錢也是有限的
能力遠比不上FLUX.1, 而SD3.5 已經改了授權, 年收入低於一百萬美金都能自由使用, 更是把auraflow的生態空間壓縮到0...生態起不來了就沒有人願意花錢和捐助更進一步的訓練(算法和人工訓練就那幾十篇論文了,沒多大的改變,到最後就是錢的比併,誰多素材誰多H100 H200,然後拼時間)

Which do you think is better, 0.6 or 0.5? I'm a little hesitant...

@08560825741 you just need a ab test

v-pred不能用webui的主分支跑，请阅读⚠️ IMPORTANT NOTICE ⚠️

@0xSeiunSky 好的，谢谢

Yes. And it works only in V-Pred

Forge can be used, you just need to git clone the latest master branch. However, it may be difficult if you don't know how to configure the environment manually. But overall it is not difficult. I used conda to install the specified versions of python and pytorch, and then pip installed the configuration file directly.

@special_offer_ubik I'm reinstalling reForge (right now) and setting up. I'm on linux using virtual environments (venv) which does something similar to conda. Is there anything in particular that I need for the environment setup?

@Pokefan You should just follow the README prompts and then pip install the environment files. However, this is probably developed using a specific python version environment. I have not used venv much. I feel that it should use the local python interpreter version. There may be warnings at first that the python version is not compatible, but the problem should not be big. You can try it first.

@Pokefan Hi, sorry for the complexity. Did you update your forge/reForge using git pull?

@Euge_ Yes. I have the latest pull from the master branch of reForge.

@Pokefan The main branch of reForge can automatically identify the prediction type of the model. Could you please describe the problem you encountered in detail? Is the model unable to load? Or can only generate pure noise? Or other problems (image oversaturation, etc.)?

@Euge_ Whenever I gen I get what people call "heat maps" and very noisy images regardless of step count. I can show you if you'd like. I'd need to figure out how to send you the output images so you can see what I am seeing.

@Euge_ The images are either blue noise or red noise or some color of noise, but it is always pixels and noise. Most of the time it is colored. Sometimes it is black and some color.

@Pokefan It still sounds like the model is being inferred as epsilon prediction... Strange. After you run git pull, is the file updated correctly and not blocked unexpectedly? (For example, you have local modifications before git pull)

@Euge_ I blew the whole install (I was previously using forge and auto11 before that, my reForge install was new, so I didn't hesitate). Then I reinstalled it again and still nothing. Git clone straight from the repo and I did a git pull on the master branch for good measure.

Same here. I only get noise.

@fantasygensokyo821 

Oh, I wasn't the only one!

I'm using comfyui and I'm stumped by the lack of improvement after applying updates.