Euler ≠ Euler A

@mega2ton2 same with regular Euler

add "sketch, lineart, flat color" in negative.

@Enigmata still trash, maybe its because of Forge?

i have same problem,using forge too @jaimuh731354 

@jaimuh731354 Maybe. I use reForge which has auto-support for V-pred models.

Try euler 25-30 steps, cfg 4.

If picture is oversaturated try to use dynamic thresholding (cfg 5 and mimic cfg 4), rescale CFG or "APG now your CFG".

Also I think you don't need brackets for quality tags (XL models don't like high weights for many tags, it is not SD 1.5 which doesn't care). High weights can break generation from noise.

Try adding an artist tag to see if it helps. Adding a tag like (artist:ke-ta:0.6) adds instant quality in Epsilon. I assume the same is true of v-pred.

super

Sport.

Dunno overall, likely is super but i'm about to test it to find out lol.

it mean "resolved the saturation issue" model ,and 0.65 is here https://huggingface.co/Laxhar/noobai-XL-Vpred-0.65 .

Submissive

SEGS

0.65s = 0.65 slightly merge with another model from an experimental noob branch.

smudged background

sepia, they merged a yellow model

yes. if not, look up forge dual prompt extension to manually make it work

https://civitai.com/articles/9070/noobai-xl-nai-xl-on-forge-is-a-different-beast-with-generation-parameters

Yes, but I suggest sticking to eps till v-pred is ready and fixed. Not just for forge

I dont use a1111 so i can't say if its the problem, but the description says it can be unstable, i recommend you to just download updated forge and test it, it's faster and works perfectly here. Also use Rescalecfg on at least 0.7 which forge has built-in, it fixes colors explosions etc.

if you followed the instructions properly it will just work.

I mean, i did the dev pull stuff (it literally only generated noise before that), and tried CFG 4, steps 30 and sampler euler, i am not sure what there is to do besides that, my settings are all default because i used the same setup to generate with pony beforehand.

I will just try a clean install of forge and see what happens

Edit: still get melted lines and colors with forge, i have no clue what im doing wrong, if anyone is willing to help i can send my images maybe

Same here

prompt S/R in X/Y/Z plot script

Yeah, something like this 1girl,estelle bright,artist:naji,artist:kasumi \(skchkko\)

@plan_truster oh I don't need any extension for that? I thought I do haha xD
what prompt do I put in the prompt box? :o

@Melty1989 ah I mean like to try each artist that is listen in danbooru in order, so like artist 1, 2, 3 and so on. if its the one you mention that would combine the style from both of them no?

@Ainokura in your main prompt you put the first artist you want to test, for example:

asanagi, 1girl, standing,

then in prompt S/R field you put the artists list:

"asanagi, wlop, as109, ishikei"

First, you put an artist in your main prompt that you want to replace in next gens (asanagi), followed by artists it will be replaced by (wlop, as109, ishikei) each new tag separated by a comma. so in an example above, you'll end up with 4 different images in 4 artist styles

@Ainokura in case you dont want to waste a few weeks on non-stop image generating:

https://huggingface.co/datasets/Panchovix/noob_top35kartists/tree/main?not-for-all-audiences=true

it gives a pretty good idea of what to expect from a style

@plan_truster whoa this is so good! that is so kind of you! thank you so much 😭😭 but outside of those artist I think I will try a few that are outside of those list haha
cheers mate! ✨✨

Yeah, the model is getting progressively worse since they added e621 to their dataset in 0.75eps. Sure, now it recognizes more characters, poses and styles, but what's the point if it looks like shit

@plan_truster I get what you're saying, 1.0 eps indeed has worse aesthetic than 0.5 eps. But above i'm talking about the new vpred (0.65s). 0.65 version (without the s, on huggingface) seems MUCH better and actually feels like a proper continuation of 0.6 vpred.

If you use it properly with zero SNR + CFG rescale, it is better than all previous checkpoints. This is a skill issue.

@metal22 I do use it properly (you don't need cfg rescale since 0.6 fyi) and i've used all the previous Noob models (eps and vpred ones). Sadly, you're either blind or don't have much skill to have issues with it.

I bet you're using Forge

Preliminary testing shows the majority of what I tried to be MORE reliable, not less reliable. The complexity conforms to the requests closer than before with less fiddling, which means your prompts for 0.6 probably don't work as well as before.

Fiddle with your prompts, because this thing is clearly better.

@AbstractPhila I fiddled with it all day, now i'm convinced that both 0.65s and regular 0.65 are worse than 0.6. 065s is just insanely bad in comparison and 0.65 has really apparent artifacts problem.

好吧我说的有点过了在0.65s上使用Euler a CFG++并将CFG设置为3仍然可以得到很好的质量，并且在画手上有了明显的进步，不得不说每个新版本始终都是在向上提升的

i start with just doing tags, then try adding a sentence more like natural language, and it can be value added

and it doesnt choke as much on basic adjective-noun combinations like "daisy print" that aren't official tags

good!

That's a Civitai issue. All models are allowed for on site generation but Civitai only loads a couple of models for generation services.

@Jemnite man civitai is so fucking lame for that, they don't give a damn about people who solely use their site to generate images, not everyone has a high end pc. I do have a gaming laptop but its already a jet engine and i don't want it to explode

VRAM costs money

@Jemnite if tensorart can do it, so can civitai. I doubt its an issue of resources, they are just hurting themselves by doing that

The official recommended CFG value is between 4~5.Try it.

@openmn793Still not much difference between generation... I have never had this issue with other models.

@Thijs I don't know the specific situation on your side, but I can be sure that your situation is an individual problem. Maybe you can test the following with the official workflow.

@Thijs Another possibility is that your prompt words are written very accurately and in detail, so naturally the results of different seeds will be relatively close. I have also experienced this, but this should not be a disadvantage of the model, but an advantage.

@openmn793 I see, thanks for trying to help though! 😁

If using fresh version of reforge, check "Use old (incorrect) CLIP G load and ztSNR application" in settings.

@onianiki Using swarmui and also tried the built in comfyui.

@sdfioajdilj nai did not implement "artist:" it was people using it,

for aesthetics noob uses worst aesthetics and very awa

@sdfioajdilj I really hope they stop doing it it just seems to kinda ruin the style of the artists a lot. Probably from blending or something from the artist: being in all the artist tags.

@sdfioajdilj yeah am aware of that, well clip is dumb so it wont degrade or improve results drastically

Yeah noticed it too, this seems to distort to sketches a ton

"boku_no_pico, official_art"

"touhou, anime screenshot"

https://docs.qq.com/sheet/DZGxRSXhvcmNmeHFv?tab=9v7zie

https://civitai.com/articles/8977

EPS 1.0 for now imo

Works just fine on Forge. Just remember to use the right sampler, the right amount of cfg and use artist tags aswell. It takes more tags to get a ''decent'' picture.

https://civitai.com/articles/9070/noobai-xl-nai-xl-on-forge-is-a-different-beast-with-generation-parameters

For Danbooru artists yes, for e621 artists no

"Both methods have been trained, and natural language has also been trained, so all kinds of usage methods are possible"

if i want to use the artist prefix, do i also need the _ if the artist has 2 names? or should i write just with the space like "Short hair"

i do not use artist and the styles work for me and i keep underscores of the username

I was antagonistic towards illustrious too, thinking why use it when pony exists but i genuinely love to use illustrious now. Not only can it recognize thosuands of characters without the use of lora, it even recognizes obscure anime characters. It's more stable than pony too and has better prompt adherence, when trying to get two unique characters in a single image they don't mixup unlike pony. There was a artist style i really loved and since no one made a LoRa for illustrious, i just used the artist name in the prompt, and guess what i got perfect results without the use of a LoRa! That's the power of illustrious, try it and see but do make sure to use proper prompts. It doesn't use the score system like pony does rather uses prompts like "best quality, Masterpiece, ultra detailed, etc", i hope you can get it to work

Soul's about right with the advantages of Noob/Illustrious: Artists are existent along with varying styles, obscure characters of both sides (Anime, Furs) may get noticed much more easily, and Noob seems pretty easy to steer.
Only downsides most can spot is indifferences in styles for some artists or at random (Sketch-looking art for example, which happened to me a few times), and probably some eye/hand artifacts that goes along with the sketch look. Most of my tests/generations were done without any hires magic on both to keep it fair, but that also might be why if not my settings.

Overall with bias aside, it comes down to personal preference: Pony is a tad more "noob" friendly and requires less prompt fighting, and it is tuned to more natural wording to get a decent photo despite lacking artists. Noob is a perfect all-rounder that gets you what you ask for with style and knowledge in mind, and that's before needing like hundreds of LoRas.

What I love though is, that both are still growing. Pony's planning a giant 7 series release if you've followed it, and Noob is nearing a 1.0 of Vpred and maybe further work on the Epsilon series according to their huggingface (some model called "V1.1", might be wrong but they're for sure kicking up something.)

@Veryhelpfulfur yeah i was gonna state some disadvantages but my reply was already too long, i hated illustrious in the beginning because of just the sheer amount of low quality output of flat coloured images, it almost seemed like it was deliberately outputing amateurish images. But tbf i was using illustrious 0.1 checkpoint instead of the more advanced ones. As soon as i figured out the prompts and how to use artist tags plus other style LoRa's, my opinion on it changed. Crazy how fast the image generation world moves, tho i can't wait for pony v7! Hope it lives up to the hype

I use both since they’re both fantastic. I use Noob and others when I want to generate a specific character, or wish to use that character as a baseline. I use Pony to generate everything else.

Not having to fumble around with LoRA for specific characters is wonderful.

It can draw a person holding a sword right out of the box.

Honestly, I don’t see why anyone would choose NoobAI-XL over models like Pony Diffusion XL or Illustrious. Both of those models are more permissive and respect user ownership of outputs.

With NoobAI-XL, the license changes give them control over your generated images, which feels like a massive overreach compared to the lenient and user-friendly terms of the models it’s based on. Why use a model that locks down your work when you can use alternatives that don’t impose such restrictions?

@karnas damn I didn't know "holding sword" was a massive hurdle for image models, flux struggles too?

Answering your question, it can recognizes far more characters than Pony it does not have artist styles hidden/ofuscated (nor hidden characters either) and the prompt adherence is better. It gets hands better than Pony as well (I'm comparing it with V6)

@soulchilli Flux is just fine at holding swords.

in my experience this model is so much better at creating unique styles and concepts i have never needed a single lora for anything, the prompting is much more enjoyable and simplified, i really disliked the score up stuff, i didnt use pony ever since i started using this

I had same issue if I skipped basic data processing.

1. Remove non-informative and rare tags
2. Remove non-informative and rare images

Also you should consider training for model version 0.5 if your loras are overfit

Civitai online image generation and online training don't support v-pred, to properly train a v-pred lora you have to use --v_parameterization --zero_terminal_snr --scale_v_pred_loss_like_noise_pred --debiased_estimation_loss args, i already asked for them to support it but no response yet. An alternative would be training on noobai eps 1.0, the loras kinda work.

@Darkwen thank you！

I did the same thing and confirm it's true

I think the commercial restrictions are good. Might scare off at least some talentless ai grifters.

@scythesaint99 good or not, the contradiction remains.

@scythesaint99 Those types will use it for commercial purposes regardless of what the license states... it’s nearly impossible to enforce. What’s more troubling is the idea that the developers of this model retain control over your outputs, even when significant human input is involved.

Ironically, the data this model was trained on likely came from open-source communities, and its foundation, Illustrious and SDXL, embraced open-source principles, including freedom over outputs. Restricting ownership to use outputs is a rejection of those principles. It’s a bad look for a model claiming to inherit from such open and collaborative origins. I find it to be a spit in the face to those who support AI and those who are against AI.

The licensing terms may vary depending on the content within the dataset. Maybe the dataset has copyrighted materials? Allowing AI-generated outputs to be used commercially requires the exclusion of copyrighted materials, which can impact the model's overall quality. Retaining these materials allows the model to remain more versatile and accessible, aligning with the goal of democratizing AI. This approach may strike a fair balance, ensuring the model isn't overly restricted while still respecting licensing considerations

Saying "outputs are not governed by the license" is not a legal statement permitting a legal use of outputs under any specific jurisdiction. The copyright of the training materials is fully irrelevant if a jurisdiction considers a work transformative. In reality this license is leaving it open for the model author to claim ownership over the outputs, disregarding the legal jurisdiction initially and then factored by any restrictions by the jurisdiction afterward. They are creating another legal barrier to ownership of outputs prior to factoring the jurisdictional laws. Since all materials trained are copyrighted to some extent and few will have any specific declarations giving the trainers to use it for this specific effort, there isn't really a balanced approach that can be taken with these licensing statements. Any and all materials trained could be found to bar any claims of ownership to anyone, including the model author or the prompter, depending exclusively on the ruling of the jurisdiction a case is found to fall under.

I am not giving any legal advice for the author of this model. But I do think that better legal advice could have probably been obtained if this license was made under these vague notions. In my personal opinion, there is not any realistic legal protection granted to the author of this model by not including these statements, regardless of the material that was used to train this model. Their primary defense of ownership will have to lie under special permissions transformative use type legal arguments according the jurisdiction. They could attempt to designate a jurisdiction themselves as a default and make reference to it in the license in any corresponding legally weak areas, but the lack of these statements could likely only be used in the case of trying to seize exclusive rights of outputs from the author. But even that claim may be overruled by a jurisdiction if they find the transformative work to fall more under the prompters influence than the model author.

I would rather not assume nefarious intention or disrespect in the license to prompters, but I absolutely agree with the poster that the ideal scenario would be that these specific statements found in prior sdxl models be amended to the license.

It's normal that this version cannot be used for commercial purposes. I don't believe any model that can generate specific Japanese anime characters would openly claim it can be used for commercial purposes. If caught making a profit, it would surely lose in court.

As for those who say commercial use is inevitable, please differentiate between commercial business operations and fan creations.

I tend to believe this is a protective clause. Allow me to say directly that any anime model capable of generating copyrighted characters cannot claim to be for commercial use.

Actually, most Japanese anime companies have policies that do not allow any fan creations/derivative works, regardless of whether they are for profit. They often choose not to sue in most cases, but if they see a profit-making tendency (especially with R18 derivative works), there are many examples of lawsuits. Most anime models operate in a gray area.

Of course, I believe there are people who do not draw copyrighted characters, but I also believe they are in the minority.However, if there are clearer provisions from the output owner, the prohibition on commercial use can remain, but they should not retain the copyright ownership over the outputs.

@NTR_BLACK I completely agree. Given that the Danbooru training set contains a large number of copyrighted works, it's wise to avoid commercial use

just a guess, but r34's tagging isn't as good, it's flooded with shitty uncanny 3D renders and poorly generated AI works, and its rating system only supports upvotes and is practically unused.

@Floofers when what problem creating furry edition?

@Floofers Yeah, r34 is a pure p0rn site, e621 uhm... "considers" itself an art gallery. I wouldnt really agree with that per se, but they at least have some form of quality control, tagging is taken somewhat seriously (at least by some like me), and its just a less amount of absolute unusable garbage you really wouldnt want in an SD model. It is strictly furry tho, while r34 is just... literally anything as far as i can tell.

@Floofers what problem ignoring tag for rule34

because relu34 is a dumpster

This model focuses on anime I think

People put useless stuff like "what has science done" and "jesus fucking christ" on unconventional pics they don't like

i dont think you realize just how much effort, time and money it takes to "create a realism version" of a model like this. i suggest you use flux or sd3.5 or juggernaut or whatever else is already available.

@Hugs288 I never said it would be easy, nor am I demanding that it be done. I'm not a fan of 3DPD, but I do like having variety available.

PonyRealism managed to pull it off, somewhat (still has a slightly anime-ish look). It's not like it's learning new concepts but rather just a "style" so it's not outside the realm of possibility, and if we manage to get to the level where it makes a sorta realistic result, then you can just img2img it with a standard model to get it the rest of the way there.

It looks like someone on Civitai here is actively working on a "noobReal" model. Let's see how it plays out.

If you're using comfyui, remove the ClipSkip or keep it -2. That can give this error.

I'm getting the exact same thing, and idk what's causing it

@nymical I'm using a1111 and I'm getting the same thing op is saying 

@nymical I was using webui, and switched to comfyui, and it started working properly.
But without "ponyv4" lora it takes 10 mins to generate 1 image for some reason...

I use forge instead of automatic, its faster and better (still same ui with every feature)
also use Euler scheduler with v-pred versions, I still recommend espilon-pred-0.75

@ShinobiSpartan I don't think A1111 properly supports this model. May be a dev branch does, but I'm not sure. You should switch to reforge instead. It's an A1111 fork, and the support and speeds are better, in my experience.

@Ka0su That's weird. If you can share your workflow I can look into it.

Hi, the problem is caused by prediction mode mismatching. You may refer to the model introduction for more information.

The main branch of the automatic1111 webui doesn't work with v-pred models.

@nymical do you know of any Google collabs that have reforge? As unfortunately I don't have a good pc to be able to generate.

@ShinobiSpartan I'm sorry, I don't. Are you using A1111 via a colab?
You can use Forge or ComfyUI colab as well, that'll work fine. I haven't used any colabs, but they both support v-pred models.

@nymical Hey, I just found one luckily, and the checkpoint works! One little question, do you happen to know what the option on the top left means, it says "UI," and under it shows "sd, xl, flux, and all." Does this just mean it shows the UI of the selected one, or is it something different

@ShinobiSpartan Yes, it just hides the irrelevant options. You can choose the one according to your model, or just choose 'all'.

@nymical k thanks again

Okay, update on the comfyUI thing.
So i've AGAIN tried fully copying someone's image, because (as i assumed) noobAI started working.
But when the result (after 10 or so minutes) came in, it gave me a... texture like drawing.
The image can be described as a light-gray png that can be used as a paper texture in game.
Suprsingly this AI can be very good for generating textures for games, since i got a lot of results like that, but i also got a lot of normal-underwhelming results, that were comparibly worse than it should have been..
Im still sticking to AnythingXL because atleast it gives me a result that somewhat works, but i will atleast try looking into the noobAIXL problem

@Ka0su You should turn on live preview in the settings, to tell if the gen is going to be a mess after first 4-5 steps. I don't know where it is in the new UI though.

@nymical I was trying to find the setting somewhere, but can't seem to find it...

@Ka0su Your KSampler might also have preview method options. I'm using 'Efficiency Nodes', but that haven't been updated in a while, so I can't really recommend it.
PS: If you want to talk further, I think you should DM me instead, others in this thread must be getting unnecessary notifications.

@nymical I don't really wanna bother someone with an issue that only i have, and where the core of the problem is not really known, so ill just try to resolve it myself or try using a different AI instead..

@Ka0su I understand, but I'm okay with it. As I said before, if you'd share your workflow, I'll look into it.

This is most likely because the ":3" tag exists in 18k captions on e621. Illustrious is not trained on e6 tags.

There are 115k captions for :3 on danbooru. This problem does not come from the e621 dataset.

你用的是A1111的WebUI吧？那个不支持V预测类别的模型，需要用Forge/reForge/ComfyUI才行

是的已经解决，感谢回复