@L_A_X Thank you very much in advance! I love this model and am looking foward to experiment it

和 nai3 差不多，建议看这个300画风法典：单画师类

细节上有些差别，总体差不多

确实是，0.75废图概率要高一点，不过色彩要好一些

来晚了。。Civitai是下不到旧版模型吗？

最前面的例图里，黑发甘雨和弹琴的芙莉莲也是我跑的，大家可以参考一下

之后再测试测试

for a certain character i like, i honestly cant tell if its undertrained or overbaked with how awful the quality is in every attempt at a gen, compared to even older models for sdxl and pony it falls behind pretty harshly.

I agree. Unfortunate.

They hated him because he told the truth.

0.5 had e6 in it too lmao, are you sure your prompting isn't just bad? Aesthetic finetuning can always be done once the model's training has finished and all the knowledge is there.

*Shit on the table.
"Why are you complaining bro. You can always clean your table later."

@Konan Literally just learn to prompt. Complaining about it giving you shit just because you can't use it is quite moronic. Every model is shit if you don't use it right (which you've clearly shown you don't do), this is no different. Furry haters are so delusional.

@Konan Skill issue

@upscaleanon537 Checking your profil I think you can teach me how to prompt plz.
@Ligmanese I know :(

@Konan Lol, going with the typical "ur profile" because you're too much of a brainlet.

Did it occur to you to check the upload date of the stuff on my profile? TWO YEARS AGO LOL.

You have brain damage and it shows.

Have to question this. Every X/Y plot I've done compared with 0.5 has been worse.

Judging by the pictures on your profile, you're not exactly gifted with a sense for aesthetics and should maybe sit this one out.

@scythesaint99 He is right though :)

@scythesaint99 Imagine judging someone for gens they made HALF A YEAR AGO LMAO

For all you know, they could be genning 10x better stuff by now.

Also you have 0 pictures on your profile, hypocrite.

@scythesaint99 Do you really think that is a gotcha, those are outputs from 1.5 models that were begrudgingly put up because the models needed preview images. AND dating back more than a year ago at this point. The 0.75 "degradation" is entirely a end user issue with how they are prompting the model. The model simply is anticipating longer more fleshed out prompts as the e621 side of the dataset generally has more verbose tagging compared to Danbooru. Thus the model's median prompt length has increased. Users should be building descriptive prompts to begin with. (keyword: control) With sufficiently detailed prompts, outputs between 0.5 and 0.75 are nearly identical. It's only low-effort short prompts that have degraded by any quantifiable degree. And that crowd of prompters aren't happy about it. They want to prompt less than a handful of tags and have the model autocomplete the rest of it without any further thought or input.

@zatochu people are extremely lazy and are incapable of reading a few manuals to see how to properly prompt. Most of these issues is dumb people who prompt without even using any artists. Of course it's shit when the prompts are basically "1girl, standing". It's gonna be shit regardless of model. It's just that this version is less finetuned for garbage prompts.

@zatochu Curious how all defenders of the current epoch have copious amounts of furry artwork on their profiles. Almost as if they are the only ones profiting from the inclusion of e621 data in the TE training.

@scythesaint99 And? The team behind NoobAI decided on their own to include the entirety of e621 in the model's training dataset. Why do you think that is?

@zatochu The inclusion of e621 obviously worked as long as the TE was frozen. For 0.75 the TE was trained and it resulted in degradation. Users are now rightfully voicing their concerns and you try to undermine that with blanket "skill issue" statements. Kinda reminds me of Lykon when SD3 launched.

If the new model spits out worse gens using the same prompts that were used on an older version of it, that's definitely not a good sign no matter how you look at it.

@scythesaint99 Source on them unfreezing the TE? Or it being frozen to begin with? I'd love a link. I fail to see any benefit of training the entirety of e621 with a frozen TE, what would it even be learning? What would be "working"? That's a lot of compute to be expending with no clear purpose.

@zatochu It was mentioned by euge on the artiwaifu discord. Either ask him directly or ask for a server invite and check old messages.

i think the model is an improvement, i havent seen any regressions beyond what i would consider normal seed to seed variance that you just correct for by tweaking here and there. my gigachad lora had a significant boost effect to contrast in many of my gens, which leads me to believe aesthetic issues if any are ripe for specific lora tuning with much less pain than in say pony

It is entirely that. The only degradation to outputs occur with shorter low-effort prompts. When prompting with sufficiently long prompts, outputs between 0.5 and 0.75 are identical.

@zatochu If the new model spits out worse gens using the same prompts that were used on an older version of it, that's definitely not a good sign no matter how you look at it.

@poefgwjorh Can you explain why it's happening? Or is it just "they added data and content I don't care for and it must be at fault."

@zatochu It's been long known that you are not supposed to finetune an already aesthetically-aligned model because that just dilutes the effect of the said aesthetic alignment. In this case, most of e621 art looks like utter trash with cartoonish taste to it (read: pony) and it's definitely less appealing from the aesthetic perspective than most of the danbooru dataset, it definitely brings the aesthetic quality down (which is average of both datasets). The good-looking furry art exists, aruurara for example, but the overwhelming majority of art on e621 is pure SLOP, and this is a huge problem, don't pretend it doesn't exist.

@poefgwjorh Maybe you should get in contact the model authors directly and tell them how badly they messed up then. Maybe offer yourself up to help them work on it. Though being less snide, or maybe more snide. That still sounds like a prompting issue where users don't know how to meaningfully drive a model aesthetically in the right direction. This isn't Pony where you don't have artist tags.And there's plenty of boilerplate tags that are REAL tags that can be placed in the negative prompt. Opposed to the nonsense that is "text,watermark,bad anatomy, bad proportions, extra limbs, extra digit, extra legs, extra legs and arms, disfigured, missing arms, too many fingers, fused fingers, missing fingers, unclear eyes,watermark,username" where a large number of these tags don't even register because the NLP has been trained out of the TE and tags like "extra digit, extra legs, extra X" were never Booru or E621 tags to begin with and they will have less and less effect the further the TE goes away from interpreting natural language.

@zatochu You've been told multiple times that the same prompts should not looks worse on the updated version of a model. You are either deliberately trolling or just conveniently trying to bend the narrative. I assume the model authors aren't kids and can figure out what to do on their own.

@poefgwjorh If they aren't kids and they can figure things out for themselves, why are you spouting rhetoric that more or less expresses they made a huge mistake that should have been "common knowledge."

@zatochu There's no simple way to finetune an aligned model, or at least I'm not aware of it. I'm sure they realize what they are doing if they decided to base off of illustrious instead of something like cosxl.

I tried the model personally elsewhere, and to be fair 0.5 is still better than 0.75, I assume that because of e621 tags that poisoned the model, I hope it will be reworked or fixed because i'd rather use 0.5 than the final version.

e621 :)

Come on bros, let's keep it civil! Instead of staying at each other's throat we should encourage the devs to improve the model!

Please post your example images

@Korewaai Okay, I'll do it a little later

yeah, idk what the other comments are talking about. my only guess is that they just happened to find styles that they liked the less accurate look of and are disappointed that they don't look the same any more. 0.75 is far better in my experience. as far as illustrious v0.1 goes, it was not trained on an aesthetic dataset.

which vae you are using?

@PM786 i either use the baked in vae or the one from madebyolin that fixed fp16 depending on which ui i'm running at the time since i have the madebyolin one set as default for SDXL in any non-comfy based ui's that i use

@nubby thank you & what do you recommend for A111?

Earlier I wrote that I didn't notice any degradation in the quality of the NoobAI-Epsilon075, but after longer testing I have to admit that this model has some anatomy issues in some cases. I decided to do a simple merge that fixes this problem. And from what I can tell, this merge came out pretty good. I created this merge for myself, but I decided to let you play with it, maybe you will like it too NoobaiCyberFix

P.S. I just noticed that version 1.0 is out. This version seems to have the same problems, so probably I'll make a second version of this merge soon

Yes, I know; I suppose it's like v prediction. Even if the TE is unstable at this moment, the UNET is stable, so using a 0.5 TE with 0.75 UNET works better than using 0.5 directly.

This format is working really well for me with the 0.75 TE:

Positive: (masterpiece, best quality, good quality:1.1), absurdres, 1girl...

Negative: (worst quality, normal quality), censored, bad anatomy, jpeg artifacts, lowres, (artist name, speech bubble, logo, watermark, signature, patreon, web address:1.2), twitter username, patreon username, english text, speech bubble, bad hands, bad feet, bad proportions, bad perspective, bad leg, bad arm, bad neck, extra, fewer digits, extra digits, missing finger, scan artifacts, adversarial noise, artistic error, wardrobe error

Steps: 28, Sampler: Euler a, CFG scale: 7

I just want to know how to use different models' clip and unet. Or is this just a conceptual statement? I don't quite understand what the underlying technology is like.

@2022211688660 That depends on which UI you're using. It should be somewhat similar to using another VAE, but since the default clip is rarely changed, the option might be a bit more hidden.

Look at the image gallery posted below and copy the metadata to start with.

@LazyTrainer I have the same problem too.But it didn't work.Though I copy the data for those images that did not use any lora, it still couldn't draw the same.

@LazyTrainer i did it and the results are trash, it generates something that looks like it came from crayon lite

@lorenzofee04557 post an image that you've had a problem with including it's metadata then link it here. it's hard for us to guess what might be going wrong for you if we can't see what you're doing, or even see the problem

@nubby thanks for replying, i think that the problem was on Automatic 1111, i moved it to comfy ui and worked perfectly using the same generation data

@lorenzofee04557 I also encountered the same problem, do I have to switch from webui to comfyui or forge

@dou_ou i recommend you to do so, it was instant improvement without changing the parameters or the metadata between the UIs

ye, i'm waiting for super duper merge of noobai vrep trained on illustrious 2.0 and artiwaifu 3.0. this will be the best model

@Kellenok yeah a merge of both would be best of both worlds because the stylization customization of artiwaifu is so good.

yes but they will not work

Thank you for your work! This is aways appreciated, I'm glad you used something closer to a character, it helps to notice the difference better. I also noticed that Hatsune Miku is a great subject for this kind of examples. Either way, thanks you for sharing!

this is perfect. UR A LEGEND

I've spend so much time on Pony checkpoints that I forgot that normal SDXL ones could remember artists without LORA remining.

这些对比非常棒，让我注意到更多好的画师和画风

In what regard? Some artists that didn't work before work now and some concepts seem to be better understood as well.

Things that weren't known are known now, and concepts that were built in are stronger. Adjust weights accordingly, your same stacks won't work as model progresses, that's why this isn't a "complete" model yet. once v1.0 is out you can start making tons of stacks you can consistently use.

They just made an announcement on the discord admitting that this version is degraded because of the TE training. Stick to 0.5 and hope that the issues get fixed for the 1.0 release.

@scythesaint99  Please provide the source.

@zeseren there is no such announcement, they are trolling. devs have not mentioned text encoder at all on discord.

@fizalpher ??? https://files.catbox.moe/hsgg5f.png

@scythesaint99 this is not the official discord, therefore could be referring to anything. Unless you have a discord link for this as well?

@fizalpher Euge is one of the creators and the announcement comes from his discord. you shouldn't claim something is trolling just because you don't know it..

@hasoo then that discord should be linked. the official nai-xl server is Laxhar, if the info isn't there there'd be no reason to believe the claims being made about the Laxhar product on another server unless verifiable.

@fizalpher You're talking BS with a very serious face. It's beautiful.

Can you try this format? it's working for me.

(masterpiece, best quality, good quality:1.1), absurdres, YOUR PROMPT HERE

Negative:
worst quality,bad quality,bad hands,very displeasing,extra digit,fewer digits,jpeg artifacts,signature,username,reference,mutated,lineup,manga,comic,disembodied,futanari,yaoi,dickgirl,turnaround,2koma,4koma,monster,cropped,amputee,text,bad foreshortening,what,guro,logo,bad anatomy,bad perspective,bad proportions,artistic error,anatomical nonsense,amateur,out of frame

@bionagato 我打算明天试试你这种方法

Where was this ever confirmed? They never said it in Laxhar server... Someone posted a screenshot but not from any of the common anime AI discords I know about, so kinda doubt authenticity.

@fizalpher They're training the text encoder (I think it's still in progress). They want it to recognize characters with fewer than around 100 tags in Danbooru. If you look for a comment I made about 20 messages back, you'll find that he replied implying that the text encoder is training. They also discussed it with a Waifu Diffusion mod and others on Discord. I personally find that 0.75 works better than 0.5, but it's a bit more unstable for now.

The openpose doesn't work

@leegridmaster Depending on the checkpoint, you may need to change the ControlNet model for it to apply effectively.

@changpingjitian641 Thank you! The new controlNet Models are released, looks like the general model is deffective.

Hi, thank you <3 I think the link isn't working for me, idk why

illustrious was not an option when this came out

@pumpkindolphin illustrious tag already came out as v 0.75 pred version came out. It's just wrongly tagged. OP can fix that easily.

I requested the same thing in the discord but I don't think they will do that

ill已死，noob当立

@Madafada1991 illustrious category is currently bugged, so i wouldnt change it for now.

clip skip doesn't work on XL architecture.

@Enigmata thank u!

you have a lot of drawing styles, i can tell you one that i like (artistningen_mame0.9),(artistsho_(sho_lwlw)0.9),(artistrhasta0.9),(artistwlop0.7),(artistke-ta0.6),(fkey0.5),(tianliang duohe fangdongye0.5),(hiten (hitenkei)0.6),

@carlosalextorres15238 Is there a page or section where all the styles are listed?

@Santiagosfx If you mean "drawing style" as in artist style, my way of doing this is going to danbooru.donmai.us , search for the artist i like (from twitter or pixiv), or click one of the picture listed, a new tab will open then i'd copy what's in the artist tag or the search box on the top left corner (ex: muku_(muku-coffee)) then paste it in the prompt box.
I do the same to any character, too.
Correct me if i'm wrong tho and if you learn any other methods later please share it too.