Yup this model is amazing

需要开v预测

怎么开启啊，有点没看懂说明

我也是醉了，网上找不到任何关于开sd V预测相关的东西。

作者就不能好好做个说明？

中文说明就轻描淡写的一句需要额外插件。

具体什么插件也没说。

搜dev 和 v-prediction 什么都没有。

难道现在已经进入全民comfyui时代，就我在玩webUI

@shoree 方法三就是给webui写的，定位到仓库目录然后git切换到dev分支

Hi, 很抱歉V预测不易上手带来麻烦。这是用户手册，其中包含了V预测的详细使用方式：https://d0xb9r3fg5h.feishu.cn/docx/WWOHdr6RMoQZxQxCZRGc5KlEnUi?from=from_copylink

如果您仍然遇到了任何困难，请继续联系我！

 我自己的电脑上的dev已经切好了。

云端的还没搞定，autodl要如何打开终端切换dev?

Same here. I'm using SwarmUI and nothing good is coming out of this model. I don't know what's the hype about, this is not even close to Pony.

@k1logr4m it knows far more knowledge than Pony, it is probably something to do with your guys settings. Try copying someone's output that displays on civitai and work from there. 

Hi, could you please share your generation parameters? Maybe I can help you take a look. Just copy the content below the image generated by forge.

@Euge_ Steps: 35, Sampler: Euler, Schedule type: Automatic, CFG scale: 4, Seed: 4280384847, Size: 1152x896, Model hash: dfdae6eac2, Model: noobaiXLNAIXL_epsilonPred075, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer version: 24.6.0, Version: f2.0.1v1.10.1-previous-626-ga332f7cc, Module 1: sdxl_vae

It seems to really struggle without artist and character specified in prompt? Like reaaally struggle.

@jaimuh731354 Since you're using epsilon prediction version, I recommend to use Euler A with 5~7 cfg scale, instead of Euler.
In addition, since you didn't share the prompt information. If the aesthetic quality of the generated image is low, you may need to add the quality tag like "masterpiece" and the aesthetic tag "very awa" into your prompt (you can think it as pony's score_9, etc.)

@AkioAI I did notice it understands a ton of concepts from booru tags, like Illustrious does. But in the quality department it falls short. I will keep playing around with it, maybe I just don't know how to use it.

@k1logr4m Yeah, at first that was me for Illustrious, it took me some time to figure out how to use the model. All you need to do is use other people's settings and change it to feel right for you.

If you want a specific style, you can find a lot of amazing creators who have made styles based on NoobAI (you can probably use artist tags before using a style lora, it is likely the model already knows the artist). NoobAI knows better concepts and composition than Illustrious, at least when I was comparing the two models. You just need to find the right prompts and negative prompts to use the full potential of the model. Good luck 🙏

@AkioAI Alright, thanks for the advice.

V-Pred offers higher contrast (brighter/darker) and impressive lighting, but it requires limited set of samplers/schedulers and different parameters.

Highly depends on sampler. You are stuck with euler for vpred and it gives mushy texture and messy lines, while eps produces clear and clean image. So, depending on artstyle you want to get eps may be way better.

try adding "dark" to positives

@SakanakoChan Periodic success

@anifibous try "chiaroscuro"

dim lighting, night time, dark,

you can also put tags in brackets for emphasis:
(dim lighting, night time:1.2),

For outdoor scenes you can mention tags like:
stars, moon, dark sky,

Would recommend experimenting.

isn't that the pony experience?

skill issue

cough model? what model?

You know that both models work different, right? Illustrious has a much better sentence understanding and is not locked to tags. Not to mention we toy with trained version of 0.1, so as soon as the model really comes out, it will blast the current PONY models away.

There's no native clown reaction, so you'll have to put up with the inline version, here you go - 🤡

That's what he's saying... OP is trying to say "I did a bunch of work to try to get good results. This model is so good that I didn't need to do that extra work, I just didn't know when I started."

3/10 rage bait

you mean pony that was too scared to have artist tags?

It takes 5 tags, 1 artist tag, 2 negatives. skill issue lil bro :)

There are three significant upsides with NAIXL over PDXL.

Firstly, you don't have to bloat your prompt with that lengthy score_9, score_8_up, score_over9000, score_score_over_score bullshit in two prompts. These mandatory tokens are a HUGE issue for PDXL because it eats away a lot of tokens and dissects model guidance way too much, making any attempt to prompt in natural language futile. Even if you only use tags, the shorter your prompt is the better it's going to be followed. NAIXL on the other hand is prompted precisely like AnimagineXL, so it doesn't require a lot of slop and it has decent natural language understanding. Not FLUX.1D level of natural language understanding ofc, but it's among the best for SDXL anime models.

Secondly, NAIXL, especially the V-pred series, handles drawings significantly better than Pony. It's not like Pony doesn't understand styles, it does, but it takes more to explain that, and the foundational base of the new NAIXL model clearly is more capable for that purpose. Illustrious just diffuses illustrations much better.

And finally, Pony is biased to ponies. Who would have thought! I mean, I don't judge, but that's really not my thing, and it takes a lot of tinkering with negative prompt to make sure it doesn't summon some banging furries at the background when you want a mere SFW illustration. Thankfully it obeys to the prompt well enough, but it only aggravates the issue with mandatory prompt even further.

@Erilaz Idk, last point is applicable to this model too. Not pony, but it clearly loves generating animalistic stuff on certain prompts. Especially with no humans tag and no artist tags.

Also you forgot to mention black screens for pony, because it's latent is more like a minefield

@Kellenok not even there lil bro. 99% of people are just borderline 1/10 at prompting.

This is sdxl based, it struggles with more than two. Easier to get ok composition then inpaint.

@Volnovik Thanks for the reply.

If you know the composition, please let me know.

Would it be better to use something like Latent Couple?

@kansann11 Composition is up to you :) Just roll until you get something satisfactory. Also you can use controlnet for that.

Check bunch of extension that I typically use, they generarly increase prompt adherence. They are explained in my article. Also be sure to use my detailer lora trained specifically on noob, I'll load eps1.1 version today.

Hi, could you please share your generation parameter? Such as prompt, sampler, scheduler type, steps, etc. So that I can help you.

@Euge_  Thank you. I was running this through the civitai generator - Rating_explicit,(masterpiece, best quality, highly detailed), very detailed face, close up, (UHD quality details), 1girl, elf, depth of field, 8k uhd, absurdres, volumetric lighting, tack sharp focus, beautiful composition, intricate detail, professional illustration

Negative prompt: (extra fingers), (bad hand), (worst quality:1.2), (low quality:1.2), simple background, censored, lowres, low detail, blurry, watercolor, monochrome, bad anatomy, bad proportions, deformed, deformed anatomy, deformed fingers, text, caption, 3d, cgi

Steps: 40, baseModel: SDXL, quantity: 3, engine: undefined, width: 832, height: 1216, Seed: 956470018, draft: false, nsfw: true, workflow: txt2img, Clip skip: 2, CFG scale: 4, Sampler: Euler, fluxMode: undefined, fluxUltraRaw: false

Having experimented a bit more, it does seem to give better results with a more complex prompt. Still get a weird smeary visual effect though. Am i overcooking the steps? 40 for euler has generally worked in the past...

@SimianHose Hi, which version are you using as the base model? Also, could you try different prompts? For example, different positive/negative prompts.

Also, you can try adding "very awa" to the quality prompts and adding "worst aesthetic" to the negative prompts.

@Euge_  I'm using v-pred-0.65S - it's the only version available on-site. Thanks for the advice, i'll try playing around with the prompt a bit more, as it sounds like my settings are basically ok.

@SimianHose Sorry I overlooked that only 0.65 works in on-site generation. With that in mind, you can try the Euler A sampler with 4~6 cfg.

@Euge_ Ok, thank you! I'll give that a try. Thanks again for your assistance with this, much appreciated.

Don't worry, I'm doing my best to experiment and research. The latest version has solved the residual noise problem and most of the oversaturation problems. It will be released after testing.

@Euge_ thank you euge <3 I look forward to your improvements

Looking forward to vpred 1.0

I have trained LORA on more than 15 characters with vpred 0.6, and I can’t go back to epsilon.

我已經變成了vpred的形狀了

It's up to Civitai to decide whether or not it gets its own category. I think Illustrious got its own fairly quickly so I don't see why they wouldn't do it soon

It probably will once the full release is out

https://rentry.org/wtfvpred
they're all good for nsfw and you can use either as a base model for lora, but loras for one may not work well for the other since epsilon and vpred are distinctly different things that require different settings in both generation and lora training. its important they named them this way because you need to know if a model is vpred or not to use it correctly.
epsilon is what almost every other local model you've probably used is, and its why you get "sepia" "yellow/brown tinge" and grey instead of black/white. vpred is what NAI (as in the SaaS model, not noobai) uses. current noob vpred has some issues and may not be as stable, they are being worked on. I wouldn't bother with loras for vpred unless you plan to re-do them once the final version comes out. you should always bake a lora for the model you intend to use the lora with; bake it on the version of epsilon you plan to use if you're using epsilon, the version of vpred you plan to use if you're using vpred. wait til training is finished and the ecosystem settles if this is still too much for you to process

These are technical terms that refer to the way the model produces the image. Epsilon is an "older" method that predicts the noise added to the image during the original forward diffusion process. When the model is trained, it learns how much noise is added to a latent during each timestep, so it can estimate how much it should remove during inference to produce the desired image. This is literally "stable diffusion".

V-pred is velocity prediction. Instead of only predicting noise, it predicts velocity, which is a combination of the original image used during the training process plus the noise added to it. This increases stability during inference, which results in higher quality images and better convergence (less steps needed when using non-ancestral samplers), as well as better noise schedulers and sampling methods. This is why the sample images provided by NAI for these models have improved contrast, richer colors and more coherent image composition compared to Epsilon examples.

@scamm i don't think he is going to give a shi what u said 🤣

@Y_X maybe not, but hopefully others that want to know will read it or Hasoo's comment. I tried lol

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

vpred - better colors and other things
epred - speia pred makes model yellow and grey
by pred they mean prediction of noise in training v does velocity of noise and e(epsilon) does noise

for now vpred is not stable and will be updated each week, their naming is version of model (its bad yeah) basically 1 is final model and 1.1 is aesthetic tuned in, 0.<number> basically means what percent of training is done 1 is fully done.

Might learn to read lil bro.

"pretentious" what is pretentious is believing everyone is as dumb as you

not sure if appropriate to ask here but is there somewhere to sign up for access to the gated models like the "Laxhar/noobai-XL-Vpred-experiments" or other similar ones that have been locked for a week+ before releasing in an open state?

@fizalpher experiment models are early models only testers (people helping euge with vpred) have access to

rest are uploaded on civit already or you can access without any permission (like vpred 6.5 without mix)

@fizalpher Thanks for your enthusiasm! Actually, it's just a repo for sharing very experimental checkpoints with a few people who are related to the recent V prediction experiments. So please just ignore it lol.

Hi, thank you for the document. I was wondering about the aesthetic tag "very awa", was it trained from the very start of the model's training?

I would love to have some help on which nodes I should be used in comfyui for v pred

@ViperF ModelSamplingDiscrete, set it to v-pred and turn on ztsnr

@kgl57944453 Thank you. Got it. I'm assuming after the Checkpoint but before the LORAs and Sampler.

@FondantViper Yep, make sure it connects to everything that needs the model node from it, if you're using facedetailer for example

@kgl57944453 It was introduced from 0.75.

I kinda agree. Danbooru has been removing images/tags due to complaints. Its filter is just more restrictive in general (though that can be a good thing for example to filter out double postings). In terms of tagging, gelbooru has somewhat more diverse vocab (though maybe less accurate?)
Though I think to use gelbooru dataset you can't use danbooru dataset as there is significant overlap.

Agreed, danbooru is a little outdated.

Sankaku Channel and Gelbooru are definitely more up-to-date and standard.

@sonnenkind Gelbooru is basically Danbooru but gives you a worse training dataset. Sure it's less restrictive, but it was literally always seen as the second site to dump images on if your stuff doesn't get approved on Dan to begin with, or can even be posted on Dan. Yes sure, you're able to have access to more artists, but unless if they are a banned artist that actually has a really good artstyle, expect to have a lot of garbage mixed in. I'd recommend just training a LoRA on banned artists if you're interested in their styles instead of needing to pull in all the crap with Gel.

@softtaco1712 Honestly man... I sincerely hope you're joking when you suggest Sankaku Channel. There is literally no image moderation at all on Sankaku, so anyone can dump anything onto there. They literally changed the post number system so that people wouldn't be able to see how many posts go unapproved (not that it matters because posts are never autodeleted when not approved unlike Dan) so that people wouldn't see how much goes unapproved, not to mention the tagless posts that may as well not exist. It's worse than Gelbooru to train stuff with. If you want a garbage dataset, go right ahead and train a finetune with Sankaku and see how well that goes, especially because AI people dump a lot of pure shite onto there.

Sankaku does have a pretty decent scoring system though, but the problem I see is the severe rate limiting on their API which would likely hamper any attempts to pull images/tags on this sort of scale and I'm still not sure if it's the best option for an additional data source either way.

That said, the issue I see with Danbooru is that it generally does not allow content that isn't broadly defined as 'anime', and while Noob mixes e621 to give furry data to the model, we're still missing 3D and other more western content so it'd be nice to have some other data source to cover that in the same way Pony did.

I never heard of Sankaku before so I went to check it out. I have 2 adblocks and I still see NSFW ads, while I can't view ANY artwork, innocent pictures are rated R15+. Who built it like that?

@Stinkek I haven’t used it actually in awhile. Wow. They are money hungry to say the least.

If you log in with a free account, you can still view most work.

@lopi9999 you’re right tbh, haven’t used the site in awhile but wow they’re really hungry for ai content now huh? It looks like that’s what they’re trying to get into the business of, at least.

@DrA_263 That is exactly the reason I originally suggested Sankaku. They have a pretty decent ranking system (by quality) and a more wide and diverse pool of artstyles than danbooru. Scraping data off of it would likely be a nightmare though as it has lots of paid requirements. The site seems much worse off than a couple years ago, tbh.

NO NEED ,BAKED VAE IS OK.

Hy

both are SDXL. V-pred & epsilon pred are just different ways of dealing with noise. In theory when V-Prediction is more stable/better that it should output better results, but for now Epsilon Pred 1.1 is better... And to emphasize, all of them use Illustrious as a base, finetuned off of that to start.

No they all based illustrious

@zeseren and illustrious is on SDXL.

What characters/series/styles data are missing? I've done some pretty intensive testing across the NoobAI models and they all have been more versatile than most Pony or SDXL models I have used. This most recent iteration seems even more so. Seems to handle natural language prompting much better than previous versions, so I am not sure that the danbooru tag training is empirical evidence of lack.

@mewtsy Don't get me wrong, the amount of data this model has is up there with the best if not the best atm. But for what I read, the Databased used it's like 6 months old aprox. It has a lot of characters I wasn't expecting like Wuthering Wave characters, but it's missing some characters from ZZZ like Piper, which was a launch character, for example.

It's just a comment, that it would be cool if the characters/series/styles database could keep updating, making this model the model go for and also that would require less Lora production, since based on some peoples mentions, it seems it doesn't play well with SDXL, Pony or even some Illustrious Loras.

Regarding Natural Language use, I haven't had that much luck getting exactly what I want many of the times. I feel I get better results just using Danbooru tags instead mostly. But it could be just the way I wrote it, even so, I feel some concepts are missing in general, which is a shame. I gave the example of an 1950s Diner, I can replicate it with Danbooru to some extent, but I need to name a lot of things, like the checkered floor, stools, etc etc.

mochi \(circle rin\) <- Try this format

don't use underscores. as stated above the correct format is mochi \circle rin\), euge's user manual would be good to read through: https://d0xb9r3fg5h.feishu.cn/docx/YpOQdtHTDoetcZxIO9fc33onnee

"Prompts should not contain any underscores "_". Influenced by websites such as Danbooru, the use of underscores "_" instead of spaces between words as tags has been circulated, which is actually a misuse and will cause the generated results to be different from using spaces. Most models, including NoobAI-XL, do not recommend including any underscores in prompts. This misuse can affect the quality of generation at best, and even make the trigger words completely invalid at worst."

@Hannibal_Lecter very thankful

@fizalpher understand. It seems that I am not careful enough. Thank you for your patient answer.

@Hannibal_Lecter @fizalpher Just tried, according to the rules in the manual after adjustment is still not recognized.

@Black130516 the artist needs to have at least 80 high quality artworks typically, also unsure if the "loli" content was trained, most of mochi's artworks aren't visible to check quality on danbooru without gold so can't confirm.

@fizalpher What about hyphens? I usually use those instead of underscores to keep unrelated words that form a phrase from being individually processed, or are spaces still better for this?

@bigredguy2 Use spaces unless specifically the danbooru tag uses a hyphen. Words shouldn't combine, use commas to separate and go full tagging. NLP begins with a capital letter and ends with a period if you add that in.

muk_(monsieur) will not work either, this tag will force the muk of the pokemon to be incorporated.

Have you tried to see if your concept exists without LORA? NoobAI has full danbooru training for 34 epochs and you shouldn't really need to use a LORA. If it's the case you still need to, then the LORA needs to be trained for NoobAI as it's gone through a lot more training and has a bit different formatting than Illustrious.

@fizalpher This was just testing a LoRA I made of a synthetic, consistent character(s). Mostly for me to learn how IL works at all. Formatting I don't know at all yet, alas, hence me trying these things

Best practice, is to read the model guide, it includes recommended starting prompt, schedulers, dimensions, etc. Also avoid using underscores _ . Search Danbooru and e261 for tags similar to the concept you are trying to build. In my experience, avoiding natural language (sentences) works best. Other than that, I guess just look at other people's prompts, loras, etc. . . . Kind of difficult to gen without experimentation, considering that's the nature of these models (near limitless combinations of possible images).

@mewtsy The "natural language captioning" seems to be a scam from the word diarrhea tag salad prompts I've found so far on decent gens... there's also esoteric quality tags like "very awa" or "year 2024" strewn everywhere. The generation guides I've found so far are by and large useless and uninformative on actual prompts. Basically, this model seems to have been trained by someone who was stroking out, good fucking luck prompting it without a ton of experimentation.

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

Parenthesis in WebUI is basically a multiplier, makes the prompt take more attention, so (text) = text x 1.1
Adding / helps keep it as just how we use parenthesis in normal text

@darionk You're supposed to use \(backslashes\) to escape parentheses, not forward slashes. But otherwise you are correct. One thing to note though is that when training a lora, image captions with parentheses do not need backslashes, they are only used for escaping in prompts.

@CosmicElement thanks! That makes a lot sense. But why they didn't use other marks since Parenthesis is a multiplier

@XDDDAI Because they are using Danbooru database and tags, the site has existed for a long time, changing all () will take way too long, so using \ is faster. Also, this is more of a software thing, they decided to use () as a multiplier, they didn't thought people would use Danbooru tags to create prompts at the time, well that is my assumption.
And yeah @CosmicElement I made a typo, but since you corrected me, I don't see the need to edit my comment, the conversation wouldn't make sense lol

@darionk Thanks for your patient reply! Yeah that makes sense, however I feel it's not a hard thing to change that in tagging automatically since they are all text files...I had this complaint is because \ is also a special mark in coding...yeah but nvm, they already did a great job training this checkpoint

You can use the Keyword " BREAK " follow by the animal you prefer.

Example:

masterpiece, best quality, newest, very aesthetic, source_furry, Equine, horse, anthro, park background, BREAK furry female, horse, big breast, wearing a dress, in the city

if your prompt is long, you may prefer to use the example below:

masterpiece, best quality, newest, very aesthetic, (source_furry, Equine, horse, anthro), beautiful background, BREAK (furry female, horse), big breast, wearing a dress, lost in the city

Avoid all danbooru tags like 1girl which signal to the model "this is anime." Instead use e6 tags like solo, female. Adding a furry artist lora of any kind even on low strength will also be enough.

I can't help you :D I have the furry tag banned globally on this website lol

@Technoyote This. Putting typical danbooru tags in negatives can also help. Whether or not furry as a tag even works in negatives or positive is hit and miss because danboorus own furry tag has quite a lot. I've put furry in negatives before and actually had humans turn into furries.

@SexyUmbreon "source_x" won't do anything in particular, that's pony specific.  

@Speedhoven Good to know lol I usually run test "source_furry" on different checkpoint that I often forget, that not all checkpoint needs it. Not just that, but I'm also a noobie playing around and learning new things about the different versions too

yes ive realized that too, it was a thing with pony as well. for example "bocchi the rock!" is way more biased towards generating gotoh hitori rather than any of the other characters.

tagging "bocchi the rock!, yamada ryo" sometimes generates ryo with pink hair, unless "blue hair" is added to the prompt, etc, etc

also, less popular artists do required loras to accurately replicate

@Hugs288 That's interesting! I usually use character Loras for Pony because it doesn't seem like Pony has super great adherence to a lot of characters I had tried to prompt for.

For sure, loras will always have a place! With Pony though, I feel like loras are essential especially for art styles. That said, nothing will quite compare to a Lora to really get the best effect. But, if you can simply get a character from a prompt with ~90% accuracy, that's definitely a more desirable method for me personally than needing to download loras.

@softtaco1712 yeah most characters dont need a lora or anything, usually just the character name is enough, if it isnt strong enough you can add copyright/descriptive tags etc. but only about a couple hundred artists can be replicated somewhat accurately by just tagging.

pony was so insanely overcooked even with lora it struggled sometimes lol

from my experience with these AI models 99% of people creating these models have no idea what they are doing, no one knows actually not even the people from big tech AI companies know, and here 90% of creators will spam the website to be on the recent list adding the same model with minimal changes. the best approach is to look at the images OP added, and those from other users. If you can afford the time and the bandwidth download and test the models you are interested in. The result may be either the model keeps ignoring your prompt and just generates the same images because it was "burned" to death by OP or the model will generate amazing images that neither OP or other users got as result.

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

@Euge_  Thank you, this is just what I needed!

for me it's look gray when i using hires. when i using tiling upscale all is ok

It’s not very serious, the picture looks a little gray, loses some saturation, and has a smeared feel. It only feels like a 10% change.

@Flange yeah, it’s about the same for me

Yeah. Sometimes jacking up CFG a little bit on 1st pass only or 2nd pass only helps.

I also use the "vivid" tag to try and squeeze more colour out sometimes.

Everything is here:

https://docs.qq.com/sheet/DZGxRSXhvcmNmeHFv?tab=8ztba1

I've been flipping through this guy's amazing style LORAs because he has lots of image previews.

https://civitai.com/user/SasakiNsfw/models

I use his LORAs for regular Illustrious but they're not working well for me in NOOB. I just type the artist name in when I'm using NOOB and try it without a LORA.

It actually works for some characters with less than 100 illustrations but only if they've been on danbooru for a solid year or two. I've found that with characters where it only has a vague idea what they look like you can improve the results by putting as many descriptive words of their appearance as possible into the prompt. Also make sure you use spaces (not underscores) in the characters name since that's what the model is actually trained on.

You may add the core tags of the character. You could find them in this csv file: danbooru_character_webui.csv · Laxhar/noob-wiki at main (huggingface.co)
Also, you may refer to this document for more information about the model:
https://d0xb9r3fg5h.feishu.cn/docx/YpOQdtHTDoetcZxIO9fc33onnee?from=from_copylink

@NVGE Yeah, was using the tags straight out of the csv/spaces/etc, or even just testing on one-word names. I'll see if adding in some additional help works out and the tag just refines from that though; thanks!

Illustrious is the base model for noob. Due to extensive training, noob differs significantly from Illustrious. Although LoRAs trained on Illustrious can work on noob, the results are not perfect. Personally, I believe that character LoRAs in Illustrious perform better than art style LoRAs

translated from：Illustrious是noob的基础模型，noob由于大刀阔斧的训练，和Illustrious有许多不同，训练在Illustrious上的lora虽然可以在noob上工作，但是效果不完美，我个人认为Illustrious的lora中角色Lora的效果好于画风Lora

NoobAI-XL is what Illustrious should've been in the beginning.

I agree with you very much! I think it's possible to use Danbooru as basic training, but it's not a good idea to train tag. The current Danbooru community shows a serious aversion to artists or styles that do not conform to the preferences of their moderators. As a result, many artists' images or styles were not allowed on Danbooru's website because they did not meet the preferences of their auditors. This is not good for training AI.

@yukiyanagiba Exactly. Hopefully they have already changed to Gelbooru by now.

@AkioAI They would have to train the entire model from scratch if they wanted to train on Gelbooru instead of Danbooru, which is probably very costly. I, too, wish they'd use Gelbooru, though, due to so many styles on there that aren't represented on Danbooru. I'd also vote to train using the Rule34 website as there are tons of artists on there that aren't even represented on Danbooru or Gelbooru. However, the Rule34 website has and allows ai artists and art on it so you'd get those styles trained in the model as well as the other stuff.

@madaraxuchiha88 Idk how they train the model or the details on finetuning/training on a full model but couldn't they just switch to Gelbooru or a better website in the middle of training? I am pretty sure you can do blacklist of "AI-assisted" and "AI-generated" tags to avoid training on A.I or really bad images that was created by A.I if they used Rule34 as well. Unless they are using an A.I to grab all of the images?

The quality control and tagging on Gelbooru is much worse than Danbooru. 80% of images on Gelbooru are pulled straight from Danbooru, the rest are duplicates or of varying quality.

@HDiffusion I see. What is a better database to grab tags and images then? if you know any

@AkioAI For anime I think danbooru is already the best. Everything else requires a ton of filtering and deduplication. I think instead the focus should be on types of anime data that isn't common on image boards, like screencaps. They've already done this starting with v0.65. I think the noob dataset is already big enough.

Try adding "male focus" after 1boy/male or adding "otoko no ko, femboy, girly, crossdressing" in negative. And it could also be cause of artist styles/loras, cause for me it gives me a solid gigachad/faceless dojin man with just "1boy, solo" without any loras.

Best tip I can provide is to make sure the male character has around 100 images in Danbooru, avoid artist that tend to draw only females and if you select one, lower the intensity of the artist by writing it as (Artist:0.7) or a lower number. I haven't needed to use 1girl, woman, girl, in negatives at all so far, but I can see those helping. Either way, good luck

i alway use
1boy,fat man/old man,short hair......
1boy,Asian man,......

It can be used perfectly with 1girl, and there is no problem that men can’t appear at all.
If you only provide "1boy," there is no reaction. However, if you add any related male prompt words, it works perfectly.

Additionally, please note that if there are any "1girl" prompts with "solo" in front, please remove them.

For the problem where there is no specific male character, I chose to train lora myself to solve it.

thank you all for the tips! I have switched a couple artists and now it makes me male characters flawlessly!