very epic for the win

Chan Sankaku Complex has a lot of artists too, maybe even more than Gelbooru

@Zeivon The problem with Sankaku Complex is a lot of their art is hidden behind that stupid membership.

@madaraxuchiha88 No it's not hidden, you just need a free account and you can see everything, I have a free acount and I see everything, it's only hidden if you open a lot of tabs, but the dataset can be obtained in a different way, no?

Rule34 has a lot of garbage on it though. And I'm not talking about just the deluge of AI slop, there's a whole load of horrible quality human-generated art on there too. Lots of quality filtering would be needed.

@Zeivon Sankaku is filled with low quality, untagged furry and doujinshi pages. I'm pretty sure that'd be a detriment to the model and literally hell to sort through and fix.

@Boatr0t that's why there are filters of quality and I would say that Danbooru has a lot more images with low quality

@Zeivon That's the thing. If were tagged, there would be no issue. However, plenty of them aren't. Sankaku's quality control is much much lower and became worse after they added that damn furry bot that uploads all the furry images.

@Boatr0t I see, yeah maybe only the top quality ones are good but I guess it's not enough to train a model

R-18G isn't locked and the loli lock doesn't apply to AI training because the datasets models like Noob use have that content in there. Also the payment option doesn't exist anymore but you can get access to censored content by uploading images to the site or winning one of the raffles they do pretty often. I do think that free users should have at least 3 tags to search, however.

I don't know the official reasoning but eps has better compatibility with sd ecosystem while vpred can in theory achieve better training results.

not til NovelAI v4 release thou

Yes, but not release version. You need git clone it.

@special_offer_ubik 方法1reForg 404了

@506291394340 他現在改成叫 dev2 枝

Ask Civitai not the uploader, they don't decide this.

Pretty helpful, btw, do you use natural language or you stay on tags?

@P_Universe tags

From your examples, I definitely prefer 0.65S. Though we might be using different settings overall, so who knows, how they'll perform on my setup. But thank you for the comparison!

saturation is easily fixed through various cfg techniques but detail is not fixable so is a good job for me

Not just artists, it has more useful tags and better tagged images overall.

I always thought its the other way around :\
Gelbooru was my first imageboard, I do notice that some artists have a lot more artwork on it compared to Danbooru

More images does not mean it's better. I don't know why this has to be explained at all, but Gel in general is where the bad images that can't get uploaded to Dan go to be dumped. They're not good to train on. In general the only reason you're seeing more images is because those images were deleted on Dan. I'd rather not see the dataset filled up with garbage.

80% of Gelbooru is a direct scrape of Danbooru which doesn't get updated tags from Danbooru posts. The remaining 20% is content that would be rejected from Danbooru because of its quality or off-topic nature. Danbooru is a far more well curated dataset.

@lopi9999 But, some things have been deleted from dan, like artists with 1000+ pics who asked to have all their data removed, those still exist on gel, so gel is better as archive & where those artists pics can still be uploaded/etc.

@fizalpher So like... 1% of Gel is actually good to train on? What's even the point then? Train a LoRA if you're interested in those artists. Unless if you can compile a list of actually good artists who got banned from Dan but are still on Gel and only train on Gel with those artists, I don't think it'll help the model in any way to just train everything on Gel. Let's not even talk about how bad Gel tags are, they are some of the worst tags I've seen, the only site with worse tags is Sankaku. So many minimal tag posts and posts with just bad tags in general.

@lopi9999 sakimichan? 1000 posts but removed from danbooru for "DMCA"... yea no. Gelbooru keeps that shit up.

@fizalpher I think in case of deleted artists it would be better to add gelbooru images only for them. Not just blindly grab all gelbooru.

ah yes so Dan is much more organized after all haha 😅 but I do agree that some artist have their work deleted on Dan which is understandable, but then again if you are looking for those artist on Gel that aren't on Dan then lora is the only way to go about it I suppose..?

"NoobAI-XL provides both noise prediction (or say epsilon prediction) and V-prediction versions. In short, the noise prediction version generates more diverse and creative images, while the V-prediction version follows the prompts more and generates images with a wider color gamut and stronger lighting."

v-pred doesnt run in A1111 while e-pred does, which is the main one.

@burnera679889 A1111 is the main one? Ahaha funny joke. A1111 is relatively dead.

@burnera679889 vpred models have worked on a1111 for over a year

For your reference: https://d0xb9r3fg5h.feishu.cn/docx/YpOQdtHTDoetcZxIO9fc33onnee?from=from_copylink

@scringle Not in main branch, no. The model loads, but it gives bizzare results that dont produce anything remotely resembling a coherent image.  

@burnera679889 you have to create a .yaml file from the safetensors model configuration, which includes the prediction type parameter, a1111 doesn't support the new standard but you can still load vpred models using the old format

You are actually using guidance equal to 5 and don't even know it. You see, since you are using PAG set to 3 - it is being added together with CFG to total guidance of 5 (that's just how it works), which is something that preview images here use too.

is PAG even working? I got the extension but turning it on and off made no difference on reforge...

There's some more settings you may refer to: https://d0xb9r3fg5h.feishu.cn/docx/YpOQdtHTDoetcZxIO9fc33onnee?from=from_copylink

@Euge_ Thx!

I finally got it working well with DPM ++ 2S ancestral, the PAG doesn't really work with it but I put the CFG at 1.7 and it works like a charm. It's very setting sensitive and it needs very specific prompting especially with artist styles. Small changes lead to huge differences - I think that's what a lot of people aren't used to.

@zendashi528 @munchkin I'm using ComfyUI PAG, could be why. Also quite sure PAG doesn't work like increasing CFG, it makes outputs more coherent but it never causes any saturation artifacting unlike CFG above 2 with these models

@atdc You need to read how PAG works, then. or you just didn't understood me. PAG doesn't increase CFG, it replaces it, and it can also work together with it like in addition.

It's specifically recommended to have lower CFG for PAG because its value and value of CFG are combined in one total guidance, which isn't the same as CFG. You can set CFG to 0 and then just set PAG to 5 or something, that would work too (not recommended to do so).

Since you're using ComfyUI, you can see it for yourself in the code even. Open custom node for perturbed attention, pag_nodes file, it has post_cfg_function with the next comments:
"""CFG+PAG"""
And after some code there is # Replace Self-attention with PAG comment and code for how it works.

I was going to comment somehow PAG works for the DPM++2s Ancestral CFG++ but yeah never mind still doesn't work for reforge

well, the actual 0.65 (non-s) that you could find on huggingface was just a direct upgrade of 0.6 more or less. but this one is like much more improved for sure.

@fizalpher I have been using the non-S versions, as I prefer to tweak the model myself. Base 0.65 also suffered from severe noise issues which made me skip out on it. I believe this was due to the addition of a good amount of data which the model did not have enough time to generalize.

you can use it in forge like any other model

@ShencilLizard I have tried forge and reforge, both didn't give me good outputs compared to the preview images.

@AkioAI share your prompt, I'll upload one of my images now

Hi, you may refer to this: https://d0xb9r3fg5h.feishu.cn/docx/YpOQdtHTDoetcZxIO9fc33onnee?from=from_copylink

@Euge_ I think I have figured out the reason why it was giving me really bad images. Thanks for trying to help but I solved the issue. The reason was because I did a fresh install of reforge (the branch dev2) so the settings weren't set correctly to give me better results.

@AkioAI What settings did you change though?

@Xeno443 I recommend using forge, the outputs look way better. The two settings I changed was "ENSD: 0" to "31337" and enable "Always discard next-to-last sigma"

i agree with you

You might have to specify your problem. Overfit, underfit, NaN?

I'm currently training some LoRAs for Illustrious using the recommended settings on Civitai. At the moment they work even better on Epsilon-pred v1.0 and v1.1 than on Illustrious itself. My training params:

@klikkeri1 Hello, I tried a lot of different configurations. First few were giving NaN, but even the one that completed didn't work properly and gave very noisy outputs when used in a v-pred workflow in NoobAI works. It's hard for me to pinpoint the problem, that's why I straight-up asked for a sample config.

@klikkeri1 and @TonyVan I'm sorry, I missed mentioning that I'm talking about the v-pred versions. I've edited my original comment to reflect that now.

@nymical  Got it, hope you figure it out.

https://civitai.com/models/833294?dialog=commentThread&commentId=612391

all other settings same as normal models

ive trained some models with these params and got pretty good results

apparently you should also disable all noise related arguments (noise offset, min snr gamma, multires_noise_discount)

@Hugs288 Thank you! I'll look into it.

Please oppen no half vae.

With Euler I try to use low CFG, but it depends on the results, it has worked well with 10, but with some artist it's better to use 5, it's a trial or error.

Hi, you may use Euler Ancestor with CFG=3~5

@Euge_ I use CFG at 4 though :\
it could be a bad install, I think I might want to reinstall Reforge

@Ainokura I'm on comfyui and Euler A 3.5-6 seems to work better for me too

Lower the CFG or try using Euler A or a CFG++ model.

I think it's mainly Civit-related SEO optimization, they're all based on Illustrious v0.1

@Machi Im sorry I didnt get the first part but its written SDXL as base model

还是说我应该更新下显卡驱动？

同样问题，至今无解

你好，我认为问题出在VAE上，与硬盘空间和显卡本身无关，在启动参数中添加 --no-half 即可解决。

如果问题仍然发生，请与我保持联系。

But you didn't mention what you actually did to fix it, tohelp others suffering from the same problem

I recommend using forge, the outputs look way better. The two settings I changed was "ENSD: 0" to "31337" and enable "Always discard next-to-last sigma" (copy & pasted from a different reply). For some reason for me at least, reforge wasn't producing not so great images compared to forge.

@Xeno443 Hi, for your reference: https://d0xb9r3fg5h.feishu.cn/docx/YpOQdtHTDoetcZxIO9fc33onnee?from=from_copylink

@Euge_  Thanks! Just getting info from some others that it seems to not work properly with reForge right now, while its fine with original Forge

Why not? These models work for LoRA training, You can use Epsilon-pred v1.0 or v1.1

Here's a simple example with only native nodes: https://imgur.com/a/r55otKR. However, using custom nodes with lora stackers is much easier, such as Efficiency Nodes or Comfyroll Studio.

@Lost_Meteor shouldnt u use a node that turns v-prediction or something?

@Haiigdso3 He should

@Haiigdso3 The current version of comfyui handles this automatically iirc. And I've already generated a lot of images on top of this set up with absolutely zero problems.

@Lost_Meteor thank u, good to know!

Hi, for the usage of artist tags, you may refer to this: https://d0xb9r3fg5h.feishu.cn/docx/YpOQdtHTDoetcZxIO9fc33onnee?from=from_copylink

@Euge_ 有中文版链接吗（

@akiame_ 有的：https://d0xb9r3fg5h.feishu.cn/docx/WWOHdr6RMoQZxQxCZRGc5KlEnUi?from=from_copylink

@Euge_ 谢谢！

You need to keep in mind a lot of artists have different art styles, and their art has evolved during time. And the model is trained in all their art available in Danbooru. So, keep that in mind, so unless the artist has had a really stable style for tens of pictures, you will get a good stable style, if not, expect some variation.

Also try to play with CFG, some artist work well with low CFG others with higher, same with their multiplier, try to loser it to 0.8 or so, testing is your best choice.

im having this issue too, did you find a solution?

@karina5057 Nope! I messed around with 0.9R and my results suck. I'm getting very "incomplete" looking eyes, bad hands, etc.Tried low CFG, high CFG, Allign Your Steps, no Allign steps, Clip Skip 2, no Clip Skip...

Then I try Illustrious models and my results are ultra beautiful again.

yeah you can, some preform even better on noob ai, I recommend using eps prediction versions.

Stick to eps, quality seem to degrade on v-pred. Imo loras trained specifically on this checkpoint as base produce better results. But dont forget that you can use artist tags here.

yeah, AYS and beta were the only schedulers that produced great results in my setup. The others output watery or noisy images, which seems to be a convergence issue.

@alter2611 Oh really? That’s interesting, personally I didn’t seem to notice noisy images, but I did notice more distorted subjects and worse prompt adherence for sure.

Euler SMEA DY CFG++ has been the best at adherence thus far from my testing. Beta and AYS still seem to have best results. I also recommend a large resolution as I’ve noticed this model can struggle fitting multiple subjects in a “normal” SDXL resolution. For example, 1792 x 1024 yields much better adherence for 7:4 aspect ratio than 1344 x 768. Because it’s v-pred, I was surprised but it seems like there are minimal if any artifacts from using a higher resolution.

ive trained several loras on v-pred snr and they all work perfectly fine, youre probably doing something wrong.

@Hugs288 what parameters did you use? I used the same as on epsilon except v-param and zero term snr. I even lowered my lr and enabled min snr gamma 5 (without this setting lora is totally garbage).

it works on my iphone. When importing the model, select v-prediction. Attention with Higher Precision remains off.

Where the hell did you find it? https://civitai.com/models/license/1116447 It bans sale of checkpoint and merges.

@Volnovik It's more than that. Part of the problem and why I say it's "poorly communicated" is because NoobAI-XL appears to be licensed under two licenses. There's the license you linked, and there's also the license terms provided in the description of the checkpoint. The terms described between them are different.

The license(s) put restrictions on how generated outputs are used outside of the norm:
From https://civitai.com/models/license/1116447
"Image Sales: Do not sell or license images generated by the Model. The following are a few examples of what is prohibited: selling or licensing a product such as a game, book, or other work that incorporates or is based on those generated images."

The license(s) makes prompt sharing a requirement:
From the description on this page [ https://civitai.com/models/833294?modelVersionId=1116447 ]
"III. Open Source Community

To foster a thriving open-source community,users MUST comply with the following requirements:

Open source derivative models, merged models, LoRAs, and products based on the above models.

Share work details such as synthesis formulas, prompts, and workflows.

Follow the fair-ai-public-license to ensure derivative works remain open source"

Again,
"...users MUST comply with the following requirements... ...Share work details such as synthesis formulas prompts, and workflows". User's MUST comply with the requirement: share prompts and workflows. It's unclear what this even means. Are users required to share all prompts they use with NoobAI-XL or just of the images they post online? What does compliance even look like?? It's not clear.

I honestly don't see how it'd be enforceable

The license is kind of invalidated because this is spun off from the Illustrious base model which states that everything allowed in that license must be allowed in any adopted license, thus it's more optimal to just not adopt a new license since it'd basically have to be the same thing. So no need to worry about being in violation of anything. Freedom of Development

Boycott Noobai until they stop with their foolishness.

Also you misinterpret the prompt sharing part. That's only for derivative models/merges/etc, you must share how you made the model and any accompanying details like pictures you provide alongside the model must have those details disclosed. Otherwise, as per BOTH licenses, outputs are pretty much unaffected by both licenses.

"Output

The output of this software is not covered by this license, and no contributor claims any rights to it."

This is the only website where their license was updated its still "fair-ai-public-license-1.0-sd" everywhere else, I noticed it a while ago and brought it to Jomcey's attention.

Everyone thinks its kind of weird considering their whole thing was about this model being an open sourced public service basically, so if I had to assume one of their team members likely went rogue and changed it without telling anybody. But we won't know for a few days.

But if you want to use the model to merge or whatever else just grab it off of hugging face
https://huggingface.co/Laxhar/

Information about it will likely be clarified in a few days.

@fizalpher I'm not sure that I am, and even if that's so, that supports my claim that the license is not clear. If I'm wrong, then the license should be more explicit so that there isn't room for misinterpretation. I think that's a reasonable ask.

The addendum part of the linked license (the OpenRail addendum) does put a limitation on outputs as laid out in my original post, so it's at odds with the Illustrious-XL license part you're quoting.

Eps or v-pred?

https://civitai.com/articles/9740

@Volnovik vpred

@Nomad334 A1111 need test branch, main branch not support v-pred model

No clip skip. Use comfyui... a1111 never updates on time. Set ZTSNR on, euler ancestral beta cfg 4.2 works fine. (or euler ancestral cfg++ cfg 1-1.8 for better colors/etc)

skill issue

@Pxrpose Literally why he's asking for help, dog brain.

@hcommand68990 why so mad

@NTR_BLACK Thank you, I was very lost

@Pxrpose yes thats why I need help, I'm a casual