If you ran into any troubles with my mix, please read description below. There is plenty of useful information. Also, you can contact me through message system on CivitAI.
Если вы столкнулись с каким-либо проблемами при использовании моего микса, пожалуйста, прочитайте описание ниже. Там есть много полезной информации. Также вы можете связаться со мной через чат на CivitAI.
Also try: DivineAnimeMix!
Introduction:
All models i've used before first release of DIvineElegance did not fit my requirements. I've decided to do something on my own to fix this issue. There was a lot of different mixes based on Anything v3 or v4 and i've taken it as a base as well.
At the beginning, it was a mix for my personal use, but it was warmly received by Unstable Diffusion discord community and some time later i've decided to share it here, on CivitAI.
Right now Divine series is my favorite models and quite unique thing for me. The main goal of DivineElegance - anime characters with as many details as it could have. Hair, clothes, backgrounds, all of this things should be beautiful, slightly more realistic and detailed.
I hope you also would like DivineElegance, so, stay tuned for newer versions of mix!
Recommendations:
Use danbooru tagging for prompts
Do ->not<- use "masterpiece" or "best quality" tags (v9+)
Sampler: "Euler a" at / 30 steps
Cfg scale: 8
Clip Skip: 2 <---- VERY IMPORTANT
99% of anime models only works with clip skip set to 2
ENSD (Eta noise seed delta) - leave blank
Negatives:
(EasyNegative:0.8), (worst quality, low quality:1.2), blurry, bokeh, dof, fog
Recommendations (Hires Fix):
Upscaler: Remacri
Hires steps: 10-15
Denoising strength: 0.3
Upscale by: any
Troubleshoot:
My colors are bleached out. What should i do?
You're missing a VAE file. For this mix i would recommend kl-f8-anime2 VAE. Just put it into SD folder -> models -> VAE folder. Then go to your WebUI, Settings -> Stable Diffusion on the left list -> SD VAE, choose your downloaded VAE.
If you want to get mostly the same results, you definitely will need negative embedding:
EasyNegative, it's better to use it at 0.8 weight.
Don't use EasyNegative (or other embeddings) with Style LoRAs, it may ruin the style!
Until you like the result, of course.
Don't use a lot negative embeddings at the same time. It definitely will ruin any possible output on all models! If you want to use negative embeddings - use one for quality, and one for hands (if you want a placebo for hands).
Technical support:
If you're ran into any problem with my mix, you can DM me on discord:
troubledarkness - discord nickname.
I'll be glad to help | Буду рад помочь.
Credits:
DivineEleganceMix Recipes (pre-v6)
Really BIG thanks to:
Nigurumi for BasilMix
WarriorMama777 for AbyssOrangeMixes and recipes
andite for Anything4.5/Cocoa/PastelMix
Ikena for Grapefruit
Cyria for ZeroTolerances
devdevice545734 for FaceBombMix
You, for reading, downloading and reviewing this mix
Description
I've started to working on an update immediately after v8 update to investigate layers deeper. But force majeure circumstances happened and i needed to put this update aside for a while. Anyway.
New version has came out, so, lets find out what changed:
Mix:
- Changed block merge recipe to improve visuals and restore some broken concepts
- Updated LoRA merged in, now it has less art with brush-like strokes to keep more clean artstyle
Visuals:
- Adjusted style to be less anime (less Anything5-like)
- Adjusted textures, should be slightly more detailed than before (v7 - v8)
- More digital style of art like on some previous versions of model (v5 - v6)
Characters:
- Fixed some really weird underwear being generated, it still can be a little funky, but regular panties is working okay now
- Fixed some simple NSFW concepts like regular cleft of venus, anus, pubic hair and nipples, was broken on v8 on wide range on generations
- Better hands and fingers in some situations like "arms up" or while lying on back
Lighting:
- Fixed too bright lighting being applied to characters
- More warm lighting, absolute zero begone
Thanks for keeping an eye on Divine series models and thanks for reading this changes, i hope you'll like new version of DivineElengaceMix as well.
FAQ
Comments (41)
v9 Patch Notes:
I've started to working on an update immediately after v8 update to investigate layers deeper. But force majeure circumstances happened and i needed to put this update aside for a while. Anyway.
New version has came out, so, lets find out what changed:
Mix:
- Changed block merge recipe to improve visuals and restore some broken concepts
- Updated LoRA merged in, now it has less art with brush-like strokes to keep more clean artstyle
Visuals:
- Adjusted style to be less anime (less Anything5-like)
- Adjusted textures, should be slightly more detailed than before (v7 - v8)
- More digital style of art like on some previous versions of model (v5 - v6)
Characters:
- Fixed some really weird underwear being generated, it still can be a little funky, but regular panties is working okay now
- Fixed some simple NSFW concepts like regular cleft of venus, anus, pubic hair and nipples, was broken on v8 on wide range on generations
- Better hands and fingers in some situations like "arms up" or while lying on back
Lighting:
- Fixed too bright lighting being applied to characters
- More warm lighting, absolute zero begone
Thanks for keeping an eye on Divine series models and thanks for reading this changes, i hope you'll like new version of DivineElengaceMix as well.
Отлично, щас будем тестить
Надеюсь результаты генерации тебе понравятся, за время тестирования этой версии микса в моей папке прибавилось почти 5000 изображений. Надеюсь, что никто не столкнётся с какими-либо проблемами, которые я мог упустить при тестировании.
Будем-будем. По превью всё выглядит мегасочно.
@TroubleDarkness как я заметил нейронка намного лучше начала работать с пальцами и их количеством. Как удалось добиться такого?
@artereket Привет, при мерже я не трогал блоки модели, отвечающие за анатомию и/или имеющие большое влияние на неё. Результаты при генерации пальцев получаются неплохие, да. Лучше чем на предыдущих версиях модели.
Каво? Еще же недавно 7 версия была. Мегахорош
Седьмая версия была выложена уже два месяца назад. Ну и торопиться было куда. Проблемы было необходимо исправлять как после релиза седьмой, так и восьмой версии. В первом случае это было необходимо из-за того, что это был первый эксперимент с block merge'м, во втором - из-за некоторых проблем, образовавшихся при миксе, которые я каким-то образом пропустил или не заметил. А именно - чрезмерная анимешность и пересвеченные персонажи.
По небольшому количеству фидбека, что я получил, можно сказать, что девятая версия определённо лучше двух предыдущих вместе взятых. И это хорошо.
Спасибо за комментарий.
абсолютли инкрэдбл! очень нравятся детализация и освещение, хоть я и люблю простоту, но это действительно прекрасная модель для прекрасных вещей... good job
Привет, спасибо за комментарий, рад что людям до сих пор нравится эта модель, несмотря на то что при сокращении количества миксов при её создании у неё был нестабильный визуал на протяжении целых трёх версий. Надеюсь, смогу и дальше продолжать поддерживать и дорабатывать модель.
I suggest use 3D slider with this model. Stunning results are obtained. Also style arts.
Undoubtedly, it is a complete and beautiful model
Hi, thanks for commentary! I hope you'll get great results using this model in a future. I'm still learning about block merging techniques to improve results i have right now. I hope someday i'll do even better update.
Thank you very much. We will wait for the new version.🌹❤
@saman_a464 Where can I find/install the 3D slider?
@Tsukilicious @PM786 https://civitai.com/models/154017/3d-slider
value is between 0.8 and 1.5 is better
@saman_a464 Thank you! This helps me a lot!
i've generated hundreds of pictures just to test the myth about multiple negative embeddings being bad. And the answer is, as expected, it depends. BUT multiple negative embeddings, multiple GOOD negative embeddings gives better results %99 of the time. If you do not use bad embeddings and you use them at proper weights you should always use multiple negative embeddings.
1. What multiple negative embeddings?
2. Proper weights? What weights?
3. Always? No. Embeddings interfere with style LoRA functionality, because it changes style as well.
4. Some embeddings may even ruin concepts, can make the same face for every generation and more.
5. > GIves better results 99% of the time. Where?
ditto
So which negative embeddings do you consider GOOD negative embeddings? Anything to recommend?
@Tutankan0n aurora and easynegative are my go to lately. usually add some variation of bad hand negatives too. But yeah depends on what model you use. I mostly tried these on any/yesmix/divinelegance/other 2/2.5D models. It might be completely different on 3D
and as TroubleDarkness said, do not combine multiple embeddings if you are using style LORAs. I personally hate style LORAs. my trials were with concept/clotthes/character LORAs
very cool model
Good to know you've liked it, thanks for commentary!
@TroubleDarkness welcome bro, if you got time take a look at my models too
Thank you for the good decryption! It may be dumb, but I learned some important things that I've overlooked before (not only using your beautiful model but also others)
Thanks for reading, good that this info can help someone. Also, thanks for commentary.
I still don't know why people recommends that awful VAE that always gives me a weird pinkish tint on the images, and blurry lines.
Well, i can't see pink tint on any generation made with kl-f8-anime2 VAE. But. Talking about blurry lines - it's mix feature (or issue, for someone, i don't know how to call it, because lines should't be always clear on slightly realistic merges). Anyway, if you know any better VAE - you free to use it without following any recommendations, of course.
eyes come out a little bit jankier with v9 than with v7 from some initial testing, so v9 needs adetailer while v7 can potentially do without it, depending on what one is doing
Yeah, you're right, 7th version of the mix has better eyes. Also, v9 can make too bright eyes sometimes. And talking about different model versions, you can't just keep up some details without sacrificing something else. For example, v7 has a ton of issues, like really a lot, because of AbyssOrangeMix in its recipe (and i can even list them if you're interested), but in general, v7 has better visuals, except a lot of brush stoke effects almost everywhere. I don't like it personally, after this version i was trying to avoid this effect. And i want to achieve at least v5 or v7 level of visuals in the next version of DivineElegance without any moves towards AOM again. Someday, somehow. No idea.
Well, right now i need some sleep time after Rabbit style LoRA update. Was really busy today.
Thanks for commentary. It's always nice to see any.
@TroubleDarkness "AOM"
That can't be helping, that model is awful.
That aside, it is a little easier to eventually get some decent hands with DivineEleganceMix versus some of the other models. A lot easier, thanks for that, and this is one of my favorite models in general since I picked it up starting with v5; this one, DreamShaper v3.31, and Deliberate v1.1/v2 are my favorites to experiment with before trying other models I've downloaded, which are generally less "general purpose" for the kinds of things I prompt. Also, girls via DivineEleganceMix are simply the best of all. 😁
@kintilian45 I don't even said AbyssOrangeMix was helping. It can provide a great visuals after merging, but also, can bring a lot of troubles. For example - blurry faces, excessive blur or foreground, broken hands, "yellow-ish" filter on the images and more. But of course, it has good sides as well.
Should be much better with v10 now.
@TroubleDarkness I'm testing it out now, thanks. Also, I forgot to ask, why do you insist in your descriptions that Clip Skip 2 is mandatory with your models for anime style(s)? They don't need it at all, from what I can tell. That's one of those little signs of your models that show they are pretty great, in my opinion.
@kintilian45 Because NovelAI (NAI) models was trained with Clip Skip values of "2", and if you'll not use it, you may loose a lot of anime concepts on generations. Most Anime models (99% literally, are based on NAI) That's a fact. Most anime models would work as it should only with Clip Skip set at "2"
Also, most LoRAs for anime models are trained with Clip Skip value of "2", and if you'll use those LoRAs with Clip Skip set to "1", you may notice a huge downgrade in understanding of concepts trained in. And maybe visuals as well.
Привет Андрей, в последнии месяца ты ничего особо не выкладывал. И я заметил что ты выложил арты с версией zheng v9, ты готовишь что то маштабное?
Привет.
v9 выкладывал, да. Но датасет требует слишком много ручного редактирования. Его я пока забросил.
А со следующей версией DivineEleganceMix я пока что не уверен.
Так что можно сказать, взял перерыв.
@TroubleDarkness ну надеюсь что нас ждёт что то очень грандеозное)
Any chance of you making a PonyXL merge? This style is still one of my favs. <3
Hi, actually, there is no any models that can provide such visuals on PonyDiffusion, right now i'm looking for different LoRAs that can somehow do similar artstyle, but most of them are poorly made of 20-100 pics without any variety in dataset. And everytime you'll try to do something different from initial dataset, style would be lost. Also, i'm not sure if my GPU is capable of merging two XL models. Need to try it out.
The best style i've been able to get on PonyDiffusionXL: https://aibooru.online/posts/66019
Still not anywhere close to DivineEleganceMix visuals, of course.
@TroubleDarkness Thanks for the reply, I see the issue. The main draw from Pony is just its multi-person capabilities, not sure if that would even be possible without much regional prompting in SD1.5. Appreciate the input tho and if I do single characters, I still love using your model!
Details
Files
divineelegancemix_V9.safetensors
Mirrors
divineelegancemix_V9.safetensors
divineelegancemix_V9.safetensors
divineelegancemix_V9.safetensors
divineelegancemix_V9.safetensors
divineelegancemix_V9.safetensors
diffusion_pytorch_model.safetensors
divineelegancemix_V9.safetensors
divineelegancemix_V9.safetensors
divineelegancemix_V9.safetensors
divineelegancemix_V9.safetensors
divineelegancemix_V9.safetensors
divineelegancemix_V9.safetensors
SDcAG0Rk-DivineEleganceMix-v9.safetensors