Why is this license different from illustrious license?

v1 is live!! POG!!!

This is problematic when one does not want to copy an artist style. One advantage of the Animagine-based models (although they lack a lot of knowledge compared to Illustrious-based ones) is that you can tweak a default style or use an artist style.

Since here you have to mix and match, sometimes getting "rare styles" (flat color, anime-coloring style is one of them) is problematic, and takes a lot of trial and error. Note that this is not a criticism of this model per se, but a feature/problem of all Illustrious (and Kohaku) based models. Not too noob-friendly, unfortunately.

Looking foward to test it when it's available for Civitai Generator.

Any idea when it will be available for on-site use?

If I want to train my own artstyle should I use this model or illustrious model for model base training?

well, time to go back to 0.5. here's hoping they'll eventually update illustrious past 0.1

最新的1.0版本测试下来感觉在各种概念上有了强化，但有个比较明显的问题，人物的大小腿比例很不正常，有的时候还容易扭曲，脚的崩坏率也比0.75版本更高。。。希望作者可以想办法修复下~

"Not only legs, but also arms turn out worse compared to model 0.75. And in general, version 1.0 has many more minor issues that weren't present in previous models. I don't want to upset the creators, but all tests show that the previous version is much better! I'll stick with 0.75 for now, thank you so much for it!"

Sorry if I made any mistakes, my English is not very good

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

strong picture,

Adding this word to the prompt will cause the screen to crash. You can try it. This problem does exist, especially in the case of long prompts. Put this word in the front or increase the weight.

YYDS

So far I'm experiencing poor backgrounds and too much creativity in character poses. I wish I had more control over those aspects...

Thank you for sharing amazing model for free!!!!!

For v1.0 stability is not good enough. Probability of anatomy distortion rises high. Anyhow thank you for your hard work. Hope a fix for it in the next version.

Good One

It's good

Awesome! :)

in 0.75, it applies lora styles perfectly but in 1.0 the style loras that I have trained are not being applied consistently to the images. Not sure if it is because the style loras needs to be retrained on this specific model for the style to be applied or not.

✨👍Unspeakable Masterpiece👍✨

Thanks for upgrading to v1.0!

1.0 overall improves the color, line and light effect. However, it's a downgrade to certain art styles (e.g. overfit, the hair always covers the pupils no matter how emphasized the relative negative prompt was) and the anatomy is not so perfect compared to v0.75. And I don't know why but the face skin color always became darker every time I tried face restoration using detailer nodes. Hope there'll be a fix on these issues soon!

From what I can tell, the Text Encoder wasn't fully frozen after 0.75, unfortunately - Whether they let it train partly after getting the feedback from 0.75 or let it go all the way through is unknown, but the Unet never had time to catch up like it should've regardless.

If anything, this is made some of the issues more apparent, depending on the style you're using.

I'll go back to waiting for the vpred release, but if you plan to use EPS 1.0, downweighting artists and strong tags to ~0.5 or lower seems to help significantly if they're trying to overpower the prompt.

Is there a lists of characters that the ckpt knows. or maybe at least it can be determined in the data set?

Is it finished? Or there will be new version?

For versions 0.75 and 1.0, there is indeed a decreasing trend in Lora's control, and pairing with character tags may require reducing character weights to achieve Lora's art style or character. However, the effect is still great without using Lora. Please try more!

i think the base model should be illustrious-xl instead of sdxl

These models are really great, but I still believe 0.5 is the best model. I would suggest a re-evaluation of the dataset both aesthetically and technically. There seems to be a lot of low quality data holding the training back. Considering you're training on illustrious, I would suggest following the illustrious training strategy outlined by its author or maybe contact them as they have been sharing their thoughts on Noob and what could be improved.

My suggestions:

-Rescrape the dataset instead of using a compressed version. According to the huggingface page you are using a e621 dataset that is both scaled down and compressed @ 90% quality, which introduces artifacts.

-Remove excessively unaesthetic images. Does the model really need, no exaggeration, 40000+ feces/diaper images? Because that's how many are in the furry portion of the dataset. With 12 million images, this means 1 out of every 300 images involves some sort of furry scatplay.

-Implement NovelAI's loss-weighting to bring even more focus to character/style tags

-Re-evaluate how you prune artist tags as they seem to be losing effect and aren't as strong as they are in NovelAI v3

-Begin training on vpred

-Pull additional data from other sources, including pixiv and gelbooru which contain more images not available on Danbooru

-Pause the TE training fully for at least 0.5-versions worth of training.

Thank you for these models and I hope to see them continue improving.

I really like how its better in colors and style composition. love it (●'◡'●)

OMFG why so BRIGHT, CONTRAST and OVERSATURATED?

please make ver 0.75 or 0.5 available again for on site generation ... 🙏🥺

insane, bro, really super good! But i cant understand how to increase quality, maybe some tags?

南無三，太感謝各位大佬了！！！

Nice model ^_^

The colors are perfect, don't listen to people. I think a lot of people just like desaturated images.

No matter what I do I can't get the same quality from v1.0 as I am getting from v0.75 or my Stable NoobAI model. I guess further testing is needed...

If you want to learn more about artist styles, you can check this:

NOOB 1.0 Version - 1000 Single Artist Test

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

If you find that using certain artists results in oversaturated images or too strong artistic styles, remember to lower the artist's weight to 0.1-0.5. The intensity of specific artwork styles can also be reduced by lowering their weights.

If you find the art style, character, and costume are over-fitted, or the image is overexposed, consider lowering the CFG to 4.5-5. Of course, this depends on your preference - some art styles and characters perform better with CFG above 5.5.

If you find character stretching or distortion, you can change the resolution. Recommended:

- Vertical images: 1024*1536

- Square images: 1024*1024

If you still want to use special resolution outputs, consider adding more camera language like "cowboy shot", "from side", "upper body", etc.

If you don't want to draw furry images, you can put furry prompt words in the negative prompt box, which can significantly improve the output quality.

In summary, version 1.0 is an excellent version, and you need to learn to understand and use it. Have fun!

score_9 model

After using version 1.0 for a while, I find it really works well. For those who think there might be issues, I suggest adjusting your prompts and settings carefully to get better results.

請問大佬，該模型的藝術家名稱大概有多少個，有標籤嗎

Yet to try it out but hearing it's potentially THE Pony Killer model. Very interesting.

大伙有没有去除文字的负面tag？不知道为什么我出图一直有文字,用了这个anatomical nonsense也没用

requesting a vpred version of the 0.5 version of noob

这个模型的出图分辨率会直接影响色调，对比度，画风和细节。1024*1024和1024*1536(宽*高)的推荐分辨率出图质量很好。我一开始无论如何弄出图质量极差就是分辨率的问题。还有不同的画师提示词也会影响画面的元素。比如写实风格的画师背景细节就会比较多。

Very bad to train LoRAs. 1.0 is ultra over-contrasted.

You can see grid comparison:

https://civitai.com/posts/8737466

LoRAs trained on 1.0, 0.75 noobai and Illustrious (All used on NoobAI 1.0).

LoRA which is trained on 1.0 is over-contrasted all time (even with artist style, which is not oversaturated).

Illustrious LoRA forgets some details of character because between Illustrious and NoobAI a gap of 12 million art - It is not the same like PonyV6 and Autismmix where PonyV6 had shit over-contrasted style but autismmix (or style loras from pony page) fixed everything with small adding of style loras (obviously not with 12 million arts, lol).

So Idk which is the best option to train loras now...

I think I try 0.75. Maybe 1.0 loras will be good on future mixes and other finetunes of this models.

P.S. Actually model is not for "noobs" because it doesn't support natural language (pony partially supports it), so danbooru tags only. And you must to know artists styles and good weights for them or style will be shit all time.

1.0 is really good! Thanks for this amazing model.

Absolutely amazing!

I find 0.75 more stable and better. ver1 tends to oversaturate and using artist style produce style burn quite often

On this model I trained lora, one of my favorite characters, but in many cases my cfg needs to be set to 4 or even lower

你不应该做的：
1，在新版本的模型上使用老版本的画风并期待他能保持原样或变得更好

2，因为自己调试好的画风在新版本变样了就来评论说新版本模型劣化变差

你应该做的：

1，在新版本的模型上尝试构建全新的画风

2，如果想证明某个概念真如你所说“过拟合”了，请拉一张xyz表格来证明你的观点

Whatever character or concept knowledge that the model learned from training the TE after 0.5 is not worth it, the model is increasingly unstable in various regards (particularly eyes are very bad in 1.0 compared to 0.5), some artists are fried compared to before, anatomy has taken a hit in certain situations, etc.

Please consider returning to 0.5 and converting it to v-pred, don't train the TE.

What am I doing wrong? NoobAI doesn't work for me, it generates simple images with thick lines, almost like sketches. What I've tried: different styles (it gets much better, but the backgrounds are clearly broken), different VAEs, no VAEs, changed clip skip from 2 to 1, different samplers with different schedulers, different resolutions, different CFGs - nothing changed.

Dreams and hopes for a real anime model have flown away like farts, carried by an autumn breeze.

使用建议（转发）-Suggestions (from others):
· 如果你使用某些画师后觉得画面饱和度过高，画师风格太强注意给画师降权，权重降低到0.1-0.5，某些特定作品的风格也可以通过降权来减弱强度。

If you feel that using certain artists results in images with excessive saturation or an overly strong style, try lowering the artist's influence by reducing their weight to between 0.1 and 0.5. For specific works with a very distinct style, lowering the weight can also help soften its intensity.
· 如果你觉得画风，角色和服装过拟合，画面过曝，可以考虑降低CFG到4.5-5，当然这个得根据你的喜好来，有的画风角色在CFG 5.5以上表现更佳。

If you find the art style, character design, or clothing overly fitted or if the image appears overexposed, consider lowering the CFG scale to around 4.5 to 5. Adjust this based on your preferences, as some styles and characters actually look better at a CFG scale above 5.5.
· 如果你发现人物拉伸扭曲，你可以更改分辨率，推荐竖图1024*1536，方图1024*1024。 如果你仍然想要特殊分辨率出图，你可以考虑加入更多的镜头语言，cowboy shot,from side,upper body等等。

If you notice issues with characters appearing stretched or distorted, you can adjust the resolution. For portraits, try 1024x1536; for square images, 1024x1024 is recommended. If you still want a specific custom resolution, try adding more camera-angle prompts, like "cowboy shot," "from side," "upper body," etc.
· 如果你不想画furry图，你可以把furry提示词放到负面提示词框，可以显著提高出图质量。

If you want to avoid furry images, add the term “furry” to the negative prompt box to significantly improve the quality.
感谢提出这些建议的大佬们，但还是希望模型做的更好，假如不需要这些额外措施就好了~

Thanks to everyone who shared these tips! It would be great if the model could handle these details without needing extra adjustments, but in the meantime, these tweaks can help.

about cfg & steps

Image posted by Wenaka_

just choose the one that suits you

Your current model is the best. The prompts are understood very well. I look forward to the next version. And if possible, please consider SD3.5 as well.

1.0太强了，感觉画风的复现远强于0.5！！皮肤质感、衣服细节更棒！！！！

1.0 has improvement in understanding prompts and better scenery. The only thing I have problems with it is when I prompt a certain character, my style lora that is trained directly on NoobAI 1.0 doesn't work but works wonderful on different characters or OCs. My lora that I have trained on illustrious works fine on 0.5 and 0.75 but the one that I trained on NoobAI 1.0 doesn't work at all when I prompt for certain characters.

If someone knows how to fix this problem then please let me know, but if it is a problem with the model, please fix this, creators.

After playing around with the epsilonPred10 version for a couple of days now, I can say that while yes, it is soo much better than the previous versions. but despite being called NoobXL, it is really not easy to prompt this model, in my case I tried with my always go to combo for artist style which is Hero neisan and Ishikei, I notice that it favour the artist that have a lot more posts on danbooru, in my case it was Ishikei.

from what I observe I have to be extra careful as to not put in too much value on one tag furthermore due to it having soo much knowledge of different artist and style it became a lot more sensitive to weights

CFG is also something I notice while generating with this model, basically speaking more CFG create more saturated images with bold outlines. the style became a bit more rigid and less "moe" or rounded. while less CFG create much more curvy, rounded and balance coloured images, things like sweat, skin surfaces and...the hair down there become a lot more natural with low CFG.

I also notice that with my own combination of Hero Neisan and Ishikei, that it tend to draw better in mosaic (especially for vajayjay) but maybe because on danbooru it have alot more censored than uncensored images. (I am still sadge that it tend to ignore a much niche feature of an image like dark groin mark for example, a more dirty looking images if you will)

I agree with what some other people say that it need a personal fine tune in order to make it easier to get the style that you want reliably (which I have to learn to do sooner or later, anyone able to be my mentor on this matter? dm me, I really needed that Hero Neisan and Ishikei fusion, its too good of an art style lol)

Anyway, I wouldn't say its bad, its really good at listening to prompts but reaaaaaaly hard to get the style you want unless, the style that you like comes from a well known artist on danbooru or...the other site E something something. I definitely looking forward to the V-pred one because THAT makes the style that I like much more frequently.

I tried a little more with loras for characters.

My grid: https://civitai.com/posts/8737466

1) LoRAs which are trained on 1.0 are sucks with almost every settings. They are over-contrasted.

2) LoRAs which are trained on 0.75 are OK with 4-5 cfg. Also sometimes "red theme, colorful" in negative can help.

3) Illustrious loras work but worse than 0.75. You can see that these loras forget details of characters. So you don't need to train it purposely on Illustrious.

About styles in noobai and loras - they work well together if style or loras are not overfitted. But I still can't find balance for art style tags. Some of them are poor, some good, some are overfitted.

Sad thing personally for me - I don't know which 3D styles model can recognize except gacha / JRPG Tales of / Ys like games. And I can't to find out it because danbooru is greedy - only 2 tags in search.

Impressive !

Anyone else got issues with watermarks showing up on 1.0? Even generating with (text, watermark, patreon logo, signature:1.2) in negative doesn't help.

I think 1.0 looks better than the earlier versions but I didn't get watermarks that I couldn't prompt away with 0.5 and 0.75.

made an updated artists sheet for version 1

PART 1 | WILDCARD

Upd: PART 2 | WILDCARD

If there is ever a 2D image generator history book, your name should be on at least page 1.

POST I ARTIST NAME LIST

Based on my personal preferences, I specifically tested the styles of artists who have drawn pantyhose, to further study how to enhance the overall feel/aesthetic of pantyhose, click the link above for details

基于我的个人喜好，专门测试了一些画过连裤袜的画师们的画风，以进一步研究如何提升连裤袜的质感,点击上面的"POST"以查看具体内容。

I have currently tested how these artists' styles perform in complex scenarios. Given that some artists' work is mostly based on a simple white background, my next step will be to test their styles on a simple white background.

目前我已经测试了在复杂场景下这些画师们的画风表现，鉴于有些画师所画的内容大部分都是基于简单的白色背景，下一步我将测试简单的白色背景下这些画师们的画风表现。

1.0 seems to have more issues with anatomy than previous versions i. e. necks turning too far and elbows bending the wrong way. I'm also having issues with results being oversaturated at similar settings to 0.5, but that may be an issue with my settings that I need to figure out.

v1.0. I wonder is this by design or what? every image generated by this model has watermark & random blurry effect on certain area. Other than that.. all is okay for me, with combination of "masterpiece, best quality, highly detailed" the image looks better.

new v-pred model when?🤔

长文警告!!!

Long text warning!!!

自己玩了几天了，分享一点经验吧（，sampler方面试下来euler A cgf++ 和 cgf scale=2（schedule type 选automatic，自动就是normal map）感觉这次发挥的很好，要细节有细节要构图有构图，对我来说SAG插件还能进一步提升细节，值得一试，hires.fix可以切换到经典的DPM++2M karras，可以修复一些扭曲（很少有就是了）。也用了好久的pony系列模型了，对比起来最值得称道的优点就是——知识储备充足！没有奇奇怪怪的审查和删减，如果有tag记录，基本上就可以生成，随心所欲畅通无阻的太爽了，让我找回了当初刚开始玩模型时候的热情。当然缺点就是对3D和现实的支持不是很好，但也正常就是了毕竟底模摆在那，这方面还是未来可期的。总体来说非常好，革命性的模型，加油！

btw:一些进阶技巧emmmmm 1.prompt施工工程也挺重要的，简单说就是多尝试，根据出图结果，观察图里的元素，加减正面反面提示词或者调整权重，然后就是一些经验之谈（比如说我在负面提示里加shoes，角色是没穿鞋子没错了啦，但“脚部”这个概念也跟着减少了，生成的图里更少出现角色的脚了，单词与单词之间会有一些概念上的连锁反应）。你的经验越多，掌握最后出图结果往往会更得心应手，多练多试就好了。（我现在正面和负面提示词加起来有1000多个token，hhh，但是还是要注意越精简越好，比如负面提示里，没必要的细节或模棱两可的就不要加了，对于tag库里搜不到的就更加要谨慎了。 2.哦还有sd-dynamic-prompts插件也挺好用的，每次生成像开盲盒，又能精简token，提高细节质量，我全都要.jpg. 3.我设置里选了pad prompt选项，感觉能好一点 4.生成多个角色时，为了防止概念的溢出，就是prompt bleeding，（角色特征的杂糅很恼人），我一般是一个角色用框框起来，加特征用逗号隔开，像这样——（character A，white hair，etc），能减少特征溢出的风险 5. 福瑞不福瑞的，你把monster girl放负面提示词里，权重拉高点就行了

It takes too much time to write it myself, so I just use AI translation，sorry for the inconvenience

I've been playing around for a few days and would like to share some experiences:

In terms of samplers, I tried euler A, cgf++, and cgf scale=2 (schedule type set to automatic, which is normal map). The performance was great this time, with both detail and composition. For me, the SAG plugin can further enhance details and is worth trying. The hires.fix can switch to the classic DPM++2M karras, which can correct some distortions (though they're rare).

I've been using the pony series models for a long time. The most commendable advantage is the extensive knowledge base! There's no strange censorship or editing, and if there's a tag record, it can basically generate whatever you want. It's incredibly liberating, bringing back the enthusiasm I had when I first started playing with models. Of course, the downside is that support for 3D and reality isn't great, but that's normal given the base model. This area still has a lot of potential for the future.

Overall, the model is revolutionary and very good. Keep it up!

btw some Additional advanced tips emmmm:

1.Prompt engineering is also quite important. Simply put, it's about trying more. Based on the output image, observe the elements, add or remove positive and negative prompt words, or adjust weights. For example, I added "shoes" to the negative prompt; the characters no longer wear shoes, but the concept of "feet" is also reduced, so fewer feet appear in the generated images. There are conceptual chain reactions between words. The more experience you have, the better you can control the final output. Just practice and try more. Currently, I have over 1,000 tokens combining positive and negative prompts, but still, the more concise, the better. Avoid unnecessary details or ambiguous words in the negative prompt, and be cautious with tags not found in the tag database.

2.The sd-dynamic-prompts plugin is also quite handy. Each generation feels like opening a blind box, and it can streamline tokens and improve detail quality. I want them all .jpg

3.I selected the pad prompt option in the settings, and it feels a bit better.

4.When generating multiple characters, to prevent concept overflow (prompt bleeding), which is annoying when character features mix, I usually enclose one character's features in brackets and separate traits with commas, like this—(character A, white hair, etc.), to reduce the risk of feature overflow.

5.Furry or not, just put "monster girl" in the negative prompt words and give it a high weight.

个人认为属于是上限很高、下限很低的模型，确实不太符合“菜鸟”的名字。不知道是不是因为提示词权重的问题，即使在同一串画师tag下，更换不同的描述出来的画风也会不同，不太会存在同一串画师tag可以保持画风任意出图的情况，不过这可能是“幸存者偏差”，因为数据还不够多。出图大小最好保证最小为1024*1024，特写类的图细节会比较好，也比较清晰，全身图基本上都会很糊，细节不是丢失就是一团乱线，当然这可能是我自己的问题。已经上传了几张图供大家参考，都是在同样的画师tag下的图。

I think this model has remarkably high potential but can also produce underwhelming results, making the name a bit misleading. It seems that even with the same artist tags, variations in the description lead to different art styles. Consistent style across outputs with a single set of artist tags doesn't appear reliable, though this might be a case of "survivorship bias" due to limited testing. Ideally, the minimum output resolution should be 1024x1024. Close-ups are generally detailed and crisp, whereas full-body images tend to be blurry, with details either lost or appearing as a jumble of lines. Of course, this could be an issue with my prompting. I've included some examples generated with the same artist tags for reference.

Added another 4,000 artists. Perhaps someone will find it useful, idk

PART 2 | WILDCARD

Single Artist Document Update [Tencent Document] NOOB Version 1.0 8888 Single Artist Test

https://docs.qq.com/sheet/DZGxRSXhvcmNmeHFv?tab=q5lnh5

Time to cancel Opus.

I have also tested the Epsilon Pred 1.0 checkpoint and find the images very saturated. I will continue to use this one.

ok

玩了些天了，1.0可玩性其实很高，就是不太好上手，稍微重口点的涩涩特别好

Meh. I tested it with a simple Miku and personally prefer the output from PrefectPonyXL. It feels like the model isn't ready for the big stage yet. Maybe a future version will convince me otherwise

Is there any place that has documentation for this model? I want to know if it is necessary to include "artist:" when referencing artists.

我在webui上用这个模型，prompt过了75出图的质量就会大幅下降，而其他的模型没有这个问题，请问一下是对webui的兼容性比较差吗

偉大的模型!

嘎嘎好用

Why does the 'base model' notation keep going back and forth between 'sdxl' and 'ilustrious'??

Amazing model

Хотел бы сказать что модель вообще нормис 10/10

Is there an entire list of characters/artists present in illustrious?

maybe Tip? : if you're bored with that weird faceless guy in your nsfw gen, type "male/female" in negative with high weight, and boom, he's gone, i'm sorry buddy, you're not invited this time lol

我给此模型（noob1.0）做了个定制的danbooru.csv（用于webui的TagAutoComplete或者ComfyUI的Custom-Scripts）

记得之前的简介中，数据集是2024-4，并且只训练的count大于80的tag，但是我现在找不到这些声明了、可能我脑子out了

现在的简介说是当日最新，我按10月算的话，也有一些七八九月，甚至更早的count大于80的tag不能复现……

总而言之，我抓取了今天的最新danbooru全部tag，并移除了count低于80的所有tag

另一份是在上面的基础上，移除了所有创建于2024-3月之后的新tag，因为如果我记得没错是4月数据集的话肯定不包含这些

理论上说，当使用这个csv时，如果自动补完功能能够识别，那它应该在此模型中经过训练了

如果没有，那也不是我的错（逃

I made a customized danbooru.csv for this model (for webui's TagAutoComplete or ComfyUI's Custom-Scripts)

Maybe I'm going a bit crazy, but I remember from the previous profile that the dataset was 2024-4 and that only tags with a count greater than 80 were trained, but I can't find those statements now.

I grabbed all the tags from today's latest danbooru and removed all the tags with a count below 80

The other one builds on the above and removes all new tags created after 2024-3

Theoretically, when using this csv, if the auto-completion feature recognizes it, then it should have been trained in this model.

danbooru_2024_4

danbooru_2024_11

另外，经常有人问noob包含哪些画师，csv里B列的1代表画师名，4代表角色，你可以用筛选抽出来所有画师词

Also, people often ask which artists are included in the noob, and in the csv column B has 1 for the artist name and 4 for the character

Are you supposed to use underscores in tags or not? Only on artists? Not at all?

There's no official guidance I can find.
OPs images are a mix of having underscores in some parts of the prompts and mostly not. Artist wildcard sets are thrown around with underscores in them (not by OP to be fair).
Base Illustrious doesn't want them afaik and you'd think the common consensus is to strip underscores... which only adds to my confusion.

I'm sure an answer to this will be "barely matters", but what was trained?

which one is better for use characters and series tag?
1. keqing, genshin impact
2. keqing genshin impact
3. keqing (genshin impact)
4. keqing \(genshin impact\)

What are the quality tags for it?

Your model looks very good but tags system is kind of hard to understand.

Guys, what character tag format are you using?

I noticed that devs said in the post, the prompt format is:

<1girl/1boy/1other/...>, <character>, <series>, <artists>, <special tags>, <general tags>

Does this mean the character tag from danbooru, for example, "xxxxxx (xxx xxx xxx)", should be "xxxxxx, xxx xxx xxx" instead?

However even the post cover images are also using danbooru format "xxxxxx (xxx xxx xxx)".

I'm a little confused.