v-pred test merge_normalsteps:
This version might work fine at 16+ steps, and CFG around 3 (still relatively low CFG).
these v-pred versions are still test merge, migt not be so stable.
Normal v3.0 might still be the best version to use.
このバージョンは16ぐらいのStepsと、３ぐらいのCFG推奨です。
頭にvp0.5testとついたバージョンはまだテスト段階で、動作に不安定があるかも知れません。
安定性やクオリティ的には普通のv3.0がまだ一番かと思います。
v-pred test merge_lowsteps lowcfg:
This is a personal merge based on noob vpred 0.5.
While it still is a test merge, I'm uploading it as the results seem not bad.
First, there are some points to note. Unlike previous EPS versions that were stable with moderate settings, this version is extremely sensitive to Sampler and CFG settings.
Works nice at low CFG and low steps.
Recommended Sampler: Euler a SGM Uniform (Not recommending just Euler)
Recommended CFG: Between 1.0~1.5
Recommended Steps: 8+ Steps are recommended, but it also operates at a usable level with 4~6 Steps. Try it out for fun.
Also, this is not necessary but it seems to work nice at 1024x1344 resolution.
Noob vpred 0.5ベースのテストマージです。
テストマージではありますが、ある程度満足できる結果にはなりましたので、共有します。
まず注意点として、既存のPersonal mergeとは違い、SamplerやCFGにかなり敏感です。
低CFG推奨で、低Stepsでもいい感じの結果を出してくれます。
推奨Sampler：Euler a SGM Uniform (ただのEulerはおススメしません)
推奨CFG：1.0~1.5
推奨Steps：8 Steps以上推奨ですが、4~6 Stepsでも割と使えるぐらいにはなりますので、試してみて下さい。
また、必須ではありませんが、1024x1344での結果がよさげでした。
※ When using the vpred model, the following environments are required:
Latest dev branch of a1111 webui
Latest version of forge / reforge
Latest version of comfyUI
※ vpredモデルを使用するにあたって、以下の環境が必要です：
a1111 webuiの最新のdevブランチ
forge / reforgeの最新版
comfyUIの最新版
v3.0:
Seems to be quite an improvement for hands and human body, thanks to NoobaiCyberFix fix recipe.(hands and human body from CyberRealistic XL and text encoder CLIP from NoobAI-XL v1.0)
Of course, thanks to noob 1.0, it also works better for e621-related dirty NSFW stuffs than previous version of personal merge.
NoobaiCyberFixのレシピを導入したお陰で、手と人体に結構な向上が見られると思います。（CyberRealistic XLから手と人体のデータを取り、TE CLIPはNoobAI 1.0のものを使っています）
もちろん、noob 1.0のお陰で、e621由来の汚いエロ絵も従来よりうまくこなせます。
Full recipe:
Implemented NoobaiCyberFix's fix recipe into 8:4:2:1:1 mix of
NoobAI-XL v1.0,
comradeshipXL-v14T13B,
Obsession v2.1,
PaSanctuary v2.0,
The I WonderMix v1
v2.1:
Merged 30% of noob-0.75 to v2.0.
This version is currently my favorite personal merge. While sustaining stable and flat anime style, it's also very capable of e621-related dirty NSFW stuffs such as sc〇t or best〇〇〇ity. (these themes might not be suitable for civitai, so I'm not uploading examples)
v2.0にnoob-0.75を30％混ぜたものです。
今んとこ自分用にマージした物の中で一番のお気に入りです。安定的かつフラットなアニメスタイルは維持しながらも、スカ〇ロや獣〇などe621由来の汚いエロ絵もバッチリです。（Civitaiの規約的に怪しいので画像は無しですが）
v2.0:
A simple merge of three noob-0.5 based models.
2:1:1 mix of Comradeship v14T13A, Obsession v2.0 and PaSanctuary V2
All credit goes to the original creators of each model.
This is a simple merge model created for personal use by merging three Illustrious XL-based models.
The Silvermoon, Comradeship, and Copycat models were merged in a ratio of approximately 1:2:2.
It was originally intended for my personal use, but I'm sharing it because I'm quite pleased with the results.
All credit goes to the original creators of each model.
I'm using these quality prompts and negative prompts as below, but it seems to work well even if you remove them or reduce them. Feel free to use them according to your own preferences.
Quality prompts: masterpiece, best quality, amazing quality, very aesthetic, absurdres
Negative prompts: lowres, worst quality, bad quality, bad anatomy, jpeg artifacts, signature, watermark
個人的な使用用途で作ったIllustriou XLのマージです。
Silvermoon、Comradeship、Copycatの３モデルを大体1:2:2ぐらいの比率で混ぜたものです。
個人使用用途でしたが、なんか結果がかなり気に入ったので共有しておきます。
元々のモデルの製作者さんたちに感謝いたします。
自分は以下のようなQuality promptsとNegative promptsを使っていますが、別に適当かいっそ消しちゃっても結果が悪くない気がします。色々とやってみてください。
Quality prompts: masterpiece, best quality, amazing quality, very aesthetic, absurdres
Negative prompts: lowres, worst quality, bad quality, bad anatomy, jpeg artifacts, signature, watermark
Description
This is a personal merge based on noob vpred0.5.
While it still has a strong test merge characteristic, I'm uploading it as the results seem not bad.
First, there are some points to note. Unlike previous EPS versions that were stable with moderate settings, this version is extremely sensitive to Sampler and CFG settings.
Recommended Sampler: Euler a SGM Uniform (Not recommending just Euler)
Recommended CFG: Between 1.0~1.5
Recommended Steps: 8+ Steps are recommended, but it also operates at a usable level with 4~6 Steps. Try it out for fun.
FAQ
Comments (137)
Does the v-pred version have hyper mixed in? would it be possible to have a version without hyper? I like to use it as well but I have a preference of tempering it's weight on different steps with lora block weight.
最高クラスのcheckpoint。有名なほとんどの作品においてLoraを必要としない。
nice
When I use ADetailer-face_yolov8n.pt, the mask area is darkened. Do you know the cause of this problem? changing vae and Sampling method had no effect. pony base does not have this problem.
try setting it manually to the same vae your checkpoint is using
this is too much power for one man
Hi there, wondering, how did you do the merge? I mean, with which tools? Is there instructions to do it? For the 3.0 version.
Thanks!
this merge is AWESOME! will you do one for noobai vpred 0.65s? it seems to be a lot better!
This is insane, I don't even know what to say
Using ntrmix as base and this as 85% refiner creates amazing images!
What the hell this model is insanely good
hi :3 would you consider making a version with epsilon 1.1? its really good
Not sure I'll be releasing one because no satisfying result yet, but still testing.
@dims2 ahh i still hope you do release it! your work is incredible :)
This is the best version of illustrious/noob ai that ive tested. for realistic images i use Cyberrealistic as a refiner
Best anime-style model I’ve tried, BTW, if I wanna train LoRAs with Onetrainer, which model should I use to make sure it works with yours? >>v3.0<<
Hi, I've never used Onetrainer, but I'm using noob eps v1.0 for Lora training.
Its ridiculous how good this is for anime pictures compared to EVERY oder Checkpoint on the entire website. everyone should take an example
Better than other similarly merged illustrious and noobai model like ikastrious, fetteredmix and morgymix?
The checkpoint is good, but the fact that it's using Danbooru instead of Gelbooru for a dataset kindof limits what you can do artiststyle-wise. Danbooru is very strict with quality and images get deleted if they aren't approved after three days, and some artists are missing hundreds of images on the site.
Having a high-quality dataset is crucial for the performance of both LLMs and text-to-image models. While expanding to include Gelbooru could increase the dataset size, it would likely require significant work to recaption, curate, and rescore the images due to the inconsistent tagging quality compared to Danbooru. Some of this could potentially be automated using AI captioning models, but human oversight might still be necessary to ensure the data meets a high standard.
@QH96 Well, in a sense, an AI model can't really be expected to flourish unless a lot of effort is put in to ensure quality. It would probably be wise to have some filters in place to screen things out on the user's side when training AI's (for example, blacklisting "lowres" to weed out low-quality jpegs/screenshots and "ai_generated" to prevent AI inbreeding)
So, which model is better use as base to train loras for 3.0? I'm currently using 3.0 itself with mixed results. Should I use Illustrious/Smooth or NOOB as base for my loras?
I ended up trying v3.0, base Illustrious, Illustrious Smooth and NoobAI 1.0. To answer my own question, the best results were from loras trained on NoobAI (EPS 1.0).
I gotta add tho that for datasets with realistic images (photos) the best results were attained by using v3.0 itself
Any plans for future models/updates?
Ok, this one is simply ridiculous and I don't really understand why more people don't use it already.
This is clearly a underestimated model for Anime Style and maybe more.
Fewer steps and fewer CFG Scale, and quality well above that of all Civitai models for less Steps. Using ByteDance Hyper-SD XL 8 Steps and 1024x1344, you can have stunning images by default without HighRes.
With this I use 18 Steps and 3 CFG and also added "high contrast" in Negative Prompt.
I also was struggling to find a Illustrious/NoobAI model who doesn't demand exaggerated or overkill prompt, and this one is perfect.
An insanely good optimized model, really good job bro.
so, try using "eyelashes, lips, nose," with pretty much every prompt to change it up. "realistic, photorealistic" isn't exactly cyberrealistic level of realism but it's pretty neat.
"western comics \(style\), cel shading", optionally with "high contrast, flat color" and/or "outline" is great.
"toon \(style\)", doesn't work as well as western comics but it can be pretty good.
a lot of artist tags work, a couple i like are reiq and dandonfuga, often there will be a lora for an artist style and it isn't actually needed since the base model knows it already. some artist styles actually work better with a very low emphasis, like 0.3 or 0.5
It also performs exceptionally well with a CFG of 1 and 8 steps when paired with the 8-step Hyper LoRA, delivering excellent results along with a significant speed boost.
@spunkymcgoo Bro, I just tested with "western comics \(style\), cel shading, high contrast, flat color, outline", the style that comes out of it is just crazy. For more relief, simply remove “Flat color” and, depending on the style you want, the image will take on a whole new dimension, sometimes even crazier.
Thanks for the tips.
@Wolfenris "anime screenshot, anime coloring," also helps for a more pure anime style
it seems obvious but it's not like pony actually makes intuitive sense like this lmao
Just try this prompt:
1girl, eyelashes, nose, lips, half-closed eyes, (anime coloring:1.2) (intricate details) newest, masterpiece, best quality, absurdres, highres, (cel shading:1.2) <lora:hyperSDXL8stepsLora.XP3k:0.5>
Guys I'm testing the stuff you've suggested here and it works so well! I've tried Hyper with Illu merges before but it always somehow affected the images. Here it works surprisingly good!
EDIT 2: putting "3d" in the negatives (except when you actually want 3d, and in that case also try adding "source filmmaker \(medium\)" to the prompt) improves things a lot of the time
EDIT: getting rid of "amazing quality, very aesthetic" has a surprisingly big effect, try it out, i'm not sure what to think of it yet
after some experimentation, i've settled on DPM++2m Karras at 20 steps with 7 CFG (and 1.5x highres at 0.6 denoising with 10 steps)
I put "masterpiece, best quality, amazing quality, very aesthetic, absurdres, highres," at the start of the prompt and "lowres, bad quality, worst quality, jpeg artifacts, sketch, 4koma, multiple views, signature, logo, patreon username" in the negative prompt. (if you actually want a sketchy style remember to remove sketch from the negative lol, and i don't like to use "watermark" with any model since SD1.5 since they tend to see the "water" token and just remove water from the image)
pretty simple.
"eyelashes, lips, nose" is good to add detail to faces with pretty much every style, but if you want the character facing away from you you'll want to remove that. Many artist tags just work, but some work better with a very low emphasis on them, like 0.3 or 0.5. a lot of artist loras are just superfluous, try the tag without the lora first if you're trying out a lora.
Try "western comics \(style\), cel shading" optionally with "high contrast, flat color" and/or "outline." it works surprisingly well, add emphasis if you want. "toon \(style\)" works but not quite as well. "realistic, photorealistic" isn't exactly photorealistic but it's pretty good
it follows danbooru tag conventions way better than pony, and it actually knows a lot of tags that pony just doesn't.
A lot of pony loras actually work pretty well with this too
What other resolutions can be used besides 1024x1344?
How can I make the negative prompts really work? I tried increasing the weight to 1.2, but it didn't work. (Current CFG is 3, as recommended). For example (antenna hair:1.2) is in negative prompt, but antenna hair still appearing!
The lower the cfg, the more creative you want it to be. Higher cfg makes it stick to your prompt.
vpred 1.0 version? pretty please?
This model (v3.0_noob-1.0based) is easily the best Illustrious variant so far. Thanks for the merge ❤️
por descargar debería estar abilitado
Any way to prevent the same glazed donut anus that's been identical in every single furry image since the late 1950s?
Also, a simple "Godzilla" prompt will produce the titular monster with, erm, tits. So one has to be aware of the proclivity for sudden onset breasts.
Wow, this is just an absurdly good model. Image quality is outstanding without need for any "masterpiece, best quality," type prompting. Very good prompt adherence and character knowledge. Very surprised this isn't in the top rated/downloaded models for illustrious
It's a shame that everyone can use this model except me.
You can use it on sea art
This is great at sex scenes, positioning of hands, limbs etc
I am struggling with getting a realism/2.5D image style, any tips?
use danbooru artists who draw in this style
@seawolf338 any recommendations?
@elfsmasher wlop :)
I'm very impressed. It's really easy to use. It's compatible with just about every XL lora and control net ive thrown at it, and generates cleaner images in fewer steps than average.
any chance for an update?
how do you use this model?
What do you mean?
This is truly magical model. Great outputs from get go, can generate interesting poses, accepts pony/illustrious/noob loras all very well.
This model is very unique and works very well. When will version 4.0 be available?
By far the best model I've used. It's a step above vanilla Illustrious and NoobAI. I really hope we can get this in the Civitai web-based generator soon.
Fantastic model, knows a ton of artists and represents their style well. Combined with LoRAs I'm making the best sh*t I've seen yet.
This is insanely good. Like actually. I'm trying to follow the CDF, but finding at 3.0 its already killing some style art Lora and okayish at 1.5. There anyway to fix that? anyone know
CDF?
@Nedok CFG typo lol.
@YomuSonsue Ah right haha. I use CFG 5 on my gens, even with loras on top.
Absolutely amazing model! I'm having a small issue with it, however, whenever I inpaint at high resolutions (1536x1536 or higher) my characters seem to start sweating profusely (for lack of a better word lol). I'm not prompting for "water" or "sweat" or anything related to those concepts, and yet it'll start making everyone look like they're currently taking a shower. Does anyone have any idea what could be causing this issue?
Did you ever figure out what was causing the sweating? When I do a hires fix equivalent to a higher resolution, I get the same thing, characters profusely sweating.
@ScoobyDoo25 I ended up reinstalling auto1111 completely from scratch & using a different VAE and the issue seems fixed now. It was likely an issue with my installation (personally I noticed that my Python version was the wrong one).
Hi, I was wondering if somebody else has this problem. No matter the CFG scale or sample steps, I always get washed out colors with this model.
I tried every setting and all my loras, even just the base and I get very weak colors. I tried multiple embeddings and other things, but nothing helps.
The overall images it creates are awesome, so I was wondering if I'm just stupid or if someone could help me out. Thanks in advance!
Are you using a VAE? This model doesn't have a baked VAE.
@Nedok Thanks for the response it worked. I just selected regular sdxl vae and now I get vibrant colors again. (@ykuni2003421 To change VAE go into Settings tab and VAE section)
@Saturnalien Awesome, glad to be of help
Sorry from Japan.
How does the erotic expression of this version compare to other models?
It's the same as other illustrious + NAI merges. Somewhat better than base illustrious, much better than pony.
please upload to Tensor Art!
im noob , im using stability matrix interface and use stable diffusion webui FORGE .
how can i use those merge model plz ?
This is on the top end of "better things you could be doing" if you ever pick a checkpoint. If you like ntr mix Xiii you will love this checkpoint
exellent checkpoint ! i really hope for a update
you only had one job, just one
I have seen many discussions below claiming how good it is and I have tried many different il checkpoints. For me this merge is by far the best of them all. It's really that good. Lots of the others are not even at the same level.
Not even wai illustrious?
@Annonff I've tried that one. I see at least this one suits my taste more. Wai and ntr are great ones too though.
@Robert_Zimmerman mmm, which of all the versions of this model do you like the most? there are many in this post
@lovelygeez1 The lastest one. Have you tried it?
@Robert_Zimmerman mmm I will try it when I have time, lately I have been looking for illustrious models that give results with contrast, I come from the oblivious model:
https://civitai.com/models/1052485/oblivious-mix-illustrious
but this model breaks down a bit with the use of style loras and character loras.
@lovelygeez1 I don't know I haven't tried that one. Good luck anyways.
And now it was removed from civitai generator, huh... Nice one
It's fine I guess because wais shuffle noob is replacing it, and that one is also good as this one ngl
Yeah. Three of the models I like just vanished a few hours ago, super lame
@Annonff are there any other alternatives you'd recommend?.
@TeamHeresy Actually coco-Illustrious-XL in 5.5 can achieves the same result and sometimes even better.
EDIT : NVM I just see that also this one was removed from civitai generator, sorry :(
Now that It is no longer available, what other options do you recommend for achieving similar results?
coco-Illustrious-XL in 5.5 achieves the same result and sometimes even better.
EDIT : NVM I just see that also this one was removed from civitai generator, sorry :(
I just wanted to say that this is my favorite merge; I was impressed by Illustrious but it was somewhat inconsistent at times. Whenever I use this merge, consistency is rarely an issue and if I ever struggle to get what I am looking for then Regional Prompter helps sort it out. Absolutely superb.
I do have some questions; first I wanted to ask if you intend on doing an update? Secondly, I was curious if you recommend Euler a SGM Uniform for the current version. I have had luck with DPM++ 2M SDE on Karras scheduling, but I didn't know how strongly it actually affected results? My steps are high (30) so I might have misunderstood some of your explanations in description.
Is it possible to create a version for pony xl?
Wondered half a day why my creations looked suddenly oddly colored and way less detailed. Thanks Civitai to remove and replace this ressource right under my nose.
Why do we have to document meta datas of creations again?
I suspect that they want to be able to filter certain keywords and resources from the system side. It makes sense that they'd want to do such a thing, because content they've decided to ban is stuff they don't want on the site. It's also stupid and annoying, and probably going to lead to them filtering and banning more and more content for being "too extreme".
I'm new on this
It's a good merge. Sometimes you will have LoRA compatibility issues that require you to tweak CFG, samplers, and schedulers.
First of all I want to thank you for this dope mix. My question is - will you update/improve the low steps version? it is very useful for testing.
Of course, I can always take the dmd2 lora and use it with any other checkpoint - but in my opinion it is not so effective when you have limited resources, plus your checkpoint is best for picks up the styles of the artists.
Wow
someone knows a vae to use for this model?
Definitely the best illustrious/noobai-based model.
even after everything that has come out since, this is still the best imo
The best one I used so far
Hands down, probably the best model i've ever used. It's incredibly fast and detailed, only complaint I personally have is the increased saturation, but can be easily fixed afterward.
I've lost access to this suddenly, what happened?
CivitAI operates on a bid system so only the top (100 I think) models with the most buzz bid on them become available for on-site generation.
This model lost the bid that time. However it is back.
This does not apply to LORAs however, only checkpoints.
My disappointed is immeasurable, and my weeks is ruined because this checkpoint has been vanished from civitai generator
The bid system is the stupidest thing ever..
Is that why it's gone?
Sus17 It is. This stupid bid system is here for over 2-3 months and everyone Mad ever since. And also there's many people quit from this site because of it too
y no more online generate and download only? :'(
aint no way it's coming back to online generation? :(
Hi everyone, I was away from civitai, and couldn't make a bid for checkpoint. I made "recurring" bid now, so it might be available for on-site generation again after 4 days(since thie week's auction ends).
I'm not sure how recurring bid works, but it might be always available for on-site generation since this week's auction ends, unless my buzz depletes.
Yo you still here in civitai this whole time? But why you don't upload content anymore? And is there's a plan to update this checkpoint? But anyway I love this Checkpoint of yours. And I couldn't find any better than this ❤️
Are you planning another low-step (8-step) merge? Like merging NoobAI 1.0 with Illustrious XL 2.0?
My GTX 1060 6GB laptop with 20GB of RAM is in the minimun requirements to run SDXL with ComfyUI, and your low-step model really helps with generating images in a reasonable amount of time.
Also, if I made my own low-step merge, what models I should use?
I don't know the recipe that author used - but you can merge any checkpoint you like with Noob Hyper DMD lora or WAI Recitified 4 steps lora.
(imo best results with those lora is - Hyphoria, YiffyMix, GrayColor 1.4 checkpoints)
Easiest way is merging in ComfyUI.
Something interesting I noticed is that even the newest one produces quality results at low steps which I have done with just 12 steps. But 8 steps might work. I will get back to you on that.
Looks like it is a bust. The lowest I can go on 3.0 before the quality becomes completely shot is 12 and that is with DPM++ 2M Karras with a CFG of 4.
nice
Not available again, bidding sucks
May I ask what does this model do? Exactly?
What do you mean? It does the same as other Illustrious based models around.
@Nedok "Seeing 'merge' in the title made me wonder—what's it merging with? I'm curious to know what makes this model special.😊
@mickyIammicky If you read further down in the description, it's an 8:4:2:1:1 merge of a few things like Comradeship, NAI, etc.
Please find a way to make this available all the time!
Its the best working checkpoint i used till now!
please win the bid
I need it!
It's still available 5 hours ago since I'm comment this and now it's vanished. We lost the bid again 🥀💔
Does that mean we can not use that anymore?
@Weidluscius if it lost the bid then yes
@Weidluscius The bid sistem works weekly, so if the checkpoint won the bid then it'll available to generate as a checkpoint in this site for 1 weeks
I genuinely hate this bid sistem because it's only 100 winner checkpoint to available over 1000 checkpoint. Worst decision ever
@Annonh I see, thank you for your time to expain. It is indeed outrageous.
I took my time to find replacements which one appears to be working as I desires to see, at least.
@Weidluscius no problem buddy. And finding a replacement is always a hard challenge when this one lost the bid
Why is it not available suddenly? I was having so much fun with the model yesterday.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.