how do I do this? I'm in forge and not sure how, do I just paste it in?

@stygianwizard42 yes

不需要，直出就好。

@Euge_ 好的，谢谢了

新版本支持更多角色并提高了画师还原度

comfyui出来的图质量有点低，请问你们是怎么调的，我出来的图饱和度太高了同时质量也不行

just search ModelSamplingDiscrete at link it to models,and switching sampling to v-predction

That's how it would work with any token, If you don't want an artist name to influence it, either do not put it or put it in negative prompt.

@L_A_X Sorry, I only saw the reply three days later.invalid invite again.

加上两人互动的TAG

比如什么“看另外一个人”“抓住对方的胸部”之类的

comfyui出来的图为什么怪怪的，怎么改都没用

俺也一样，怎么弄都很怪

请问解决了吗

@wumingzi49357 我也是，用几个几乎没黑线的画师凑起来硬是多了一堆黑线

@wumingzi49357 可以试一下降低去噪强度，好像会好一些

@akiame_ 没有啥用

i found the NAI3 format works but i just use the names like ciloranko, sho \(sho lwlw\), etc

What I did is use the format artist: XXX(artist's name), as long as the artist has a tag on Danbooru it will work if you are not sure the artist you want has a tag or not you can copy the artist's name and try to research on Danbooru's artist page.

This is what happens when things stay open source and non-gate kept, the community comes together and makes amazing models like this.

@Light7799 indeed open source FOREVER.

one should just discard their "muh ethics" to achieve this

I am sorry for noob comment, what can it do over pony? like it has built in characters or something different?

@PM786 not a noob comment. What it does over pony is it can do pretty much everything pony is able to do, but it makes the generations look more human-like, and less-ai. like, it doesn't have that AI feel or look that most models do. (I can't explain it, it's just something you notice if you've used AI for a long time) It knows a lot of artists also, on top of that it's pretty good with knowing obscure characters out of the box, can do some more niche concepts, too. It also doesn't need that score nonsense

@Light7799 Yup, pony also have pretty horrid lighting.

whoa what you can do that?? I always thought it will bleed into each other. how do you prompt it assigning one character trait for that character only and not having it bleed to the other ones??

@Ainokura like this, 3girls, yuri, breast press, (Character name), (Character name), (Character name), on chair, voyeurism, bed, bedroom, netorare. i also added thing things like hololive english or nier automata/pokemon etc, it is highly dependable on how many images you can find on danbooru of that character tho anything under 1k is not guaranteed be a trained tag. the logic of how make describe a scene with only tags instead of natural language is each character has a task, so you know that on chair and voyeurism and netorare is locking them to chair, i think order have deciding factor so. also, if there is a lot of fan art of kronii getting cucked by mumei and fauna so it kind of knows what order to put them in.

@zhanzix93913 I see, I thought so as well. maybe its just me but if I put things directly next to the name of the character's name it most likely will assign it to that one, maybe its just placebo effect lol but you're right about the order of which things you prompt first is important as well.

for most people, it will be better to wait until it's finished training, since this version is an extremely limited proof of concept.

@hasoo Hmm, well when it's finished, is v-pred going to be the "offical" version? Does it have that much of an advantage over the legacy samplers?

I know it sounds lazy, but I just have Fooocus set up now and am wondering if I should take the time to set up a diffusers-based script or UI like reforged.

@brucethemoose2 in terms of official by the creators, they said there will be both versions available fully completed. vpred will have a full range of color and should make much better shaded/night scenes and contrast especially, if those things are important to you.

here are some prompts I stole from the gallery and made lazy comparison gens of between the current unfinished vpred vs eps 0.5 models, to try to highlight some of the difference: https://files.catbox.moe/mfjaoj.jpg

https://files.catbox.moe/2am859.jpg

https://files.catbox.moe/w8629d.jpg

(please note the judgement in these images should be largely focused on color/contrast since we don't have a 1:1 comparison with training)

you can see that the eps version has a tendency to "wash out" colors with grey and cause night/dark scenes to have a look a bit like there is a spotlight in the image, or lean towards grey instead of pure white. it can still look very nice on its own, but vpred/zsnr is used to fix this coloring fault. According to the NAI paper that was released, vpred zsnr may also positively impact image composition, but I think it's too early to tell yet if that will be the case with the noob models.

I'm biased towards personally preferring vpred, but if you're already comfortable as-is, I'd wait til they're both released then look at both posted galleries. If you notice the color difference of vpred and prefer it, might as well switch your setup if fooocus doesn't get support for it. In terms of official as in loras, my guess is most creators will bake eps loras or loras for other eps models that still have some forward compatibility, which I think should work on either version? (I might be wrong there and vpred needs vpred loras, sorry for misinfo if I'm mistaken..)

V-pred is heavily advantaged over non-vpred versions of equivalent training time. The primary advantage is color accuracy, composition and diversity. You can get dark blacks and brighter brights with high amounts of sharpening with v-pred models. But the main rule of thumb at the end of the day is more training time = better model, so just use whichever one has the most epochs.

@hasoo Do I need a config file to go with it? I read that that's the case for v-pred models.

I figured out that I have to add a "ModelSamplingDiscreet" node to the mix but I'm not sure if I need to do anything else.

Also is it worth using that node for the epsilon model too? It seems to make things look better with that too but it might be confirmation bias.

@patchworkpants Generally you don't need the config yaml anymore. Comfy has a ModelSamplingDiscrete node that can handle v-pred without issue. reForge inherits the comfy node as a built in extension in the upstream_dev branch. You need to look for the advanced sampling extension. If you're using Forge or the dev branch of Automatic1111 webui @nubby baked the keys into the model so they'll autodetect and run:
https://huggingface.co/nubby/noobaiXLNAIXL_vPredTestVersion_with_keys

You can't use v-prediction sampling with non-vpred models. It just makes a bunch of jumbled noise.

@patchworkpants what @Jemnite said is accurate. if anyone has any issues getting things set up further or related questions, I recommend considering joining the noobxl discord group linked in the model description - lots of helpful people there that usually respond pretty fast ❤️ (sorry for my late reply 😅)

@hasoo awesome! I will. Thanks

@Jemnite I'm finding I'm getting lots of banding or weird glitchy looking backgrounds, especially with img2img or unsampling. I'm at work now so I can't really check the discord but I will later on

@Jemnite what's the deal with the epsilon setting on that node? I tried it with the epsilon version of this model and the results were different but I couldn't tell if they were necessarily better

Epsilon noise is just the regular sort of noise that SDXL is trained on. I'm not sure why it's different, possibly because it's detecting ZSNR settings? I use reForge so I may not be of much help if you're using ComfyUI. You can try asking their matrix space here:

https://app.element.io/#/room/#comfyui_space:matrix.org

To my understanding Illustrious is trained on only a danboory dataset while noob is trained on both danbooru and e621 making for a much bigger dataset although I could be wrong

@Idkanymore50 Oh. Thanks

There is this list for novelai https://raideninfinity.pythonanywhere.com/naidv3_tag_search

im pretty sure they share the same tags

大佬太强了

太强惹

目瞪口呆了，这个连贯性太强了