Natural Sin Final and last of epiCRealism
Since SDXL is right around the corner, let's say it is the final version for now since I put a lot effort into it and probably cannot do much more.
I tried to refine the understanding of the Prompts, Hands and of course the Realism.
Let's see what you guys can do with it.
Thanks to @drawaline for the in-depth review, so i'd like to give some advices to use this model.
[expand to see Advices]
Advices
use simple prompts
no need to use keywords like "masterpiece, best quality, 8k, intricate, high detail" or "(extremely detailed face), (extremely detailed hands), (extremely detailed hair)" since it doesn't produce appreciable change
use simple negatives or small negative embeddings. gives most realistic look (check samples to get an idea of negatives i used)
add "asian, chinese" to negative if you're looking for ethnicities other than Asian
Light, shadows, and details are excellent without extra keywords
If you're looking for a natural effect, avoid "cinematic"
avoid using "1girl" since it pushes things to render/anime style
to much description of the face will turn out bad mostly
for a more fantasy like output use 2M Karras Sampler
no extra noise-offset needed, but u can if you like to đ
How to use?
Prompt: simple explanation of the image (try first without extra keywords)
Negative: "cartoon, painting, illustration, (worst quality, low quality, normal quality:2)"
Steps: >20 (if image has errors or artefacts use higher Steps)
CFG Scale: 5 (higher config scale can lose realism, depends on prompt, sampler and Steps)
Sampler: Any Sampler (SDE, DPM-Sampler will result in more realism)
Size: 512x768 or 768x512
Hires upscaler: 4x_NMKD-Superscale-SP_178000_G (Denoising: 0.35, Upscale: 2x)
Useful Extensions
!After Detailer | ControlNet | Agent Scheduler | Ultimate SD Upscale
â No VAE needed but it is better to use one for more vibrant colors
â Feel free to leave Reviews and Samples - and always have fun creating â¤
Description
Since i'm low on time i still wanted to release the RC1 of an new aproach.
ordinary faces, natural look, more face diversity
hand fix
! USE VAE
Recommended simple Negative for average looking output, avoid "optimizer" Embeddings
cartoon, cgi, render, illustration, painting, drawingseems to be a bit nsfw heavy and very lose on prompt guidance for now
FAQ
Comments (112)
Man v5 really doesnt wanna make black people, ive tried every prompt i can think of using terms like: black, African, African American, Ethiopian, Congolese, Ugandan, Afro Caribbean, and so on.
there has to be a lora for it right? for darker skin tones?
Try for example (ebony woman), (dark skin) in positive. And: white, caucasian, asian, chinese, japanese, korean in negatives.Also check your negatives for tokens which may overwrite certain content.
@vitokeorlini225Â there is Ethnic helper (embeding) that can do a pretty good job : https://civitai.com/models/69734?modelVersionId=74394
wake up hun epinikion uploaded a new model...
wake me up in 2 days đ¤Ł
@SyntheticSunsets knock knock....he switched it to free mode now...
@SyntheticSunsets wake up babe epinikion switched it to free mode
@vitokeorlini225Â thanks dad (and epinikion). This was pleasant to wakeup to!
Early Access!? You've blue balled my entire weekend! D:
I switched it to free mode, didnât want to ruin your weekend đ
@epinikion Oh my god this is why I keep a dedicated folder for your models and LoRAs, out of pure respect. Absolutely well earned one of the top creators here Epi!!
No more blue balls!!
@Gairm I don't get the hype, but well, thank you đ
Awesome work as usual, would really appreciate a inpainting version please!
Since this is only a Release Candidate i wont put out an inpainting model yet. U alwys can merge one by yourself
How to make your own inpainting model:
1 Go to Checkpoint Merger in AUTOMATIC1111 webui
2 Set model A to "sd-1.5-inpainting" model
3 Set model B to any model you want
4 Set model C to "v1.5-pruned" model
5 Set Multiplier to 1
6 Choose "Add difference" Interpolation method
7 Make sure your model has the "-inpainting" part at the end of its name (Anything3-inpainting, DreamLike-inpainting, etc.)
8 Click Run button and wait
9 Have fun!
(Source: Reddit)
I tried to generate few images use settings for V5 and for RC1 cover picture, the result seems blurred...I uploaded them to gallery, not sure where did I do something wrong or used wrong settings:(
As mentioned in the Version info, use VAE. Iâm uploading a VAE version right now, but well, my internet sucks and civit isnât helping with breakups. 3rd time I try now.
@epinikion haha silly me, I'm ready to try it as soon as it comes out:)
Wow, with Vae its actually incredible.
Amazing job epinkion
What's the difference between Pure Evolution v5 and Natural Sin RC1?
i think that he added into merge model called "next photo v3" its almost copy....
@vendo19324Â Nope, next Photo is crap imho, due to prompting and outcome
Please check version info, itâs all there
@epinikion I am left with the same question regardless. It is not clear why I should pick one over the other.
So the VAE Version is also uploaded now, i'm off now for the weekend. Don't break things.
I tried to generate images with other VAE b4 you uploaded and it is an incredible excellent work, ty for your work!
I'm gonna break them, muahahah!
I cannot reproduce your examples, is there an ENSD ?
Thanks a lot for your great work(s) !
Great job! I am very interested in training models, I have already tried many different techniques, and I wanted to try something bigger, 2-7 thousand images to fine-tune. I am pretty sure that the first results will be crap, but I had great results with small datasets with 300 images. I would be grateful if you could share some top/tricks or basic workflow about dataset preparation, etc. I don't ask you to share the full details because it's your own work, but at least something to begin with a larger dataset. Thanks.
sideline but someone tell me where can i find and download <lora:ER4ZQV5:1>
What does the Natural Sin version fix/add to the experience?
@epinikion thank you for this new model.
I like the skin, but may there is a little conflict with the age prompting.
If there is a nsfw token inside the positives, it starts hardly to ignore older ages. I added a post with 3 generations. Inside the single image there are the generation datas. May you have a idea.
Thanx for sharing. Have a nice weekend.
Hey man first post.
I'm loving your models, tried many at this point but I am seeing myself always coming back to yours, especially when you drop a new one which happens quite frequently :).
Just wondering if you could tell me how you got the "restart" sampler? First time seeing it and i didnt find any tutorial how to install it.
thanks man and have good one,
Hey thank you. The restart sampler is included in the the dev branch from Automatic1111
we really need this as SDXL :)))
Do you have a discord account? How can I contact you?
man the inpainting version is great.i got about 6 inpaint checkpoints,this one of the 6 is my most used,it generates almost everythign i wanted,looking forward for updating,tho its already graet
difference between natural sin vs v5?
See version info or just do A/B Test. Since it is a release candidate, I will remove it soon for the new 1.5 model on a different Name, so we wonât see any new epiCRealism checkpoint again đĽ˛
@epinikion RIP epiCRealism? Wh00t? ;/
Oh man, this is MIND BLOWING! I tested with some Loras and.....WOW! I did like: TXT2IMG 512x680, then I 'refined' it by using this same epic checkpoint, IMG2IMG, with ULTIMATE SD UPSCALE, like 1130x1500.... with even fewer prompt tags. Wow! Just imagine bigger pics! IT worked with single prompts! And......strangely (or not), distortions and other problems almost didnt exist anymore! I just use 'easynegative' and a couple of other negatives, but with even fewer words there!
Guys, what did you do to this checkpoint (epiCPhotoGasm naturalSin)? Is it just me or it's just TOO damn PERFECT? And I did like a newbee, by using TXT2IMG 512x680, then IMG2IMG by using ULTIMATE SD UPSCALE, res 1130x1500.... DENOISE 0.40! Wow!
Where do I download the VAE for Version 4?
Sorry for the spam, had to unpublish the whole model and republish the versions to unpublish the naturalSin Versions đ
Nvm head over to the new model epiCPhotoGasm
Hey I'm looking to download the VAE for epiCrealism, where has it gone?
@LuckeyG_420Â what VAE do you mean? If u mean naturalSin it is back up.
Lol, no problem, downloading epiCPhotogasm now...
what was wrong with naturalSin, why unpublish it?
@SAC020Â was pre-state of epiCPhotoGasm, so i thought i'll remove it, but hey one model more or less, who cares :D
I would be really helpful to have a table differentiating all of the models variations.
@Primaveri not gonna happen đ
@epinikion when I had this browser page up a couple of nights ago, there were two NaturalSin tabs. One is still there (RC1) and the other is gone. What happened to the gallery images of the one that is now gone? I really want to try and access the images posted by other users under the non VAE NaturalSin but can no longer get to them
@sdshifted published it, have fun
Thank you to republish 'Natural Sin' version. I am really impressed with the results and I am testing more tags. I used a prompt like: "high detail, raw photo of 20years old woman, smile, full body, standing, nsfw, brown hair, single long braid, (narrow waist), wide hips, (thick thighs:1.1), hairy, boots, backpack, (massive scale stunning environment), ruins, cave, sharp focus, daylight"
Negative: "easynegative, (low quality:2), lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, blurry"
Perhaps the way I used 'high detail + raw photo' in positive, with 'easynegative' in negative made this checkpoint 'wake up' with more quality. Well, it's all about tests in order to find the sweet spot.
whats the difference between this one and the new model and 5V?
I'm reasonably new to this. Where do we find the differences between the different versions?
Right column - under the "reviews" there's an About This Version tab you can open where the creator can put notes on the specific version you're currently looking at.
awesome thank you :) so version 5 is the best one? having all features of the other ones like New Era?
@CaptainKokomo The last one should be the best one. Also V3 is stable and for ânot that beauty facesâ u can use naturalSin
I am a fan of epicrealism, but hard for me to admire the new model's pictures by users based on the type of pictures being uploaded LATELY. If I have to put up with seeing THOSE KINDS of pictures, I would rather there be a separate place to upload "sex pictures". Nudity I guess is fine.
I agree. But the public is immense, some has his/her kinks, and the community is growing even faster (now with Nvidia stocks getting like +5% due to AI research getting hotter). For some it's okay, for others it's gross. IIRC, there is a way to filter certain 'nude pictures' in the settings. Perhaps in the future, we will get a subtab, say: REVIEWS-> (Backgrounds, SFW , NSFW->female nudity, NSFW->male nudity, NSFW->trans, NSFW->gore....whatever). Or a way for each OP to disallow certain tags/pics.
If you go to your account settings, there is a beta Content Moderation feature where you can turn off diff kinds of specific explicit imagery. I think that will help you a lot.
@schsch Yep. I agree with that. I don't necessarily need to see the sex stuff. I find women beautiful and that's pretty much art to me anyway. If I see a naked dude, I'll pretty ignore it because not my thing but that's not what I'm seeing. These dudes are sick perverts. I get why women don't like to be objectified.
I understand. The only way to get comforted in our 'bubbles' (safe zones) is by limiting exposition in the settings. Other than that, everyone is doing wonderful art, and this 'AI era' is just in the beginning!
@schsch My opinion of art is not being grossed out by people with issues, but admiring beauty, being inspired, see things from a BETTER perspective and seeing worlds from alternate universes.
Or maybe you should use the built in features the site provides to filter out content you don't like...
I see your point, but I prefer it the way it is now. Look at it this way; like it or not, nsfw content is what has advanced modelling techniques, AI-aided image generation, digital media itself (many are unaware of this curious trivia) so, in a way, even though I don't generate nsfw content, I appreciate those who do because they advance technology for everyone, so I don't judge them. Also, if you block all nsfw content you're going to deprive yourself of seeing really good gems (I've copied parts of GENIUS promps from some pretty racy pieces (that I wouldn't have gotten to see if I chose not to see that part of the community) Separating the content and hiding the most controversial, in MY opinion, will bring us more harm than good. FREEDOM FOR ART!
While I love all your models, Natural Sin RC1 has become my favorite. It definitely outputs the "realest" images for me without that silky smooth 'touch up' other models tend to have on skin textures. I can't wait to see that one developed more if that's even possible!
Is there anything you can do to improve bare feet?
not working so well with personal trained lora
I don't know how you do it but since I took the inpainting V5 I no longer use controlnet. THANKS
I have tried training a Lora on this checkpoint and got some very strange (not good) results. Is there an alternate version of this that I can use for training? Thanks
Did u use Natural Sin? V5 or V3 are an alternate option which may work out better. I'm not that Lora guy, so i dont know really
@epinikion I have tried v5. I think the problem may be that it is pruned? I do not know a lot about various "flavors" of a model, but I believe a "full" version might be needed for training upon?
@SAC020 For me, it works best to train Lora on regular SD 1.5 (pruned emaonly, the base model should be around 4GB). Then when training is done you can use the lora with epicrealism. When I try to train Lora on checkpoints such as epicrealism or realisticvision etc. the results are very weird, so I stick with training on base SD 1.5 model
@vorkje700943Â Thank you
Is the VAE included in this model and other realistic models, the original VAE ft mse from stability?
I think it is generating same nose. :D. That's really odd.
Oh, good - it's not just me then! I can't seem to get much variety out of the faces in this model at all. Very similar noses on every one of them.
After reading this comment, I can't take eyes from noses in gallery, and INDEED every nose looks the same... holly....
Same here. I was afraid I did something wrong.
Honestly i keep getting the same face. I can see it across ages and many nationalities. It changes for gender.
Will you release an inpainting version of Natural Sin RC1? :)
Nice model. Thank you for sharing. Would you make Realistic SDXL model too?
Can somebody explain naturalSINRC1 vs pureevolution?
I am currious to
Great model! Beautiful pictures!
I've tried a lot to adjust the age of people in the images but whatever I tried, I had no success.
Do you have an advice how to adjust people's age using only prompting?
It is in my opinion the best model but i have no idea what is the difference between natural sin, pure evolution v5 and epic photogasm, please tell us
You can read the Version Notes. PureEvo5 is the last refined model from the pureEvo Base. Natural sin has the pureEvo base and fintuned on more natural Photos, not that beauty faces and more analog style photography. epiCPhotoGasm has been trained from ground up with different dataset which is more ânormal photosetsâ and finetuned merge with pureEvo afterwards. Hope this helps u guys. EpiCPhotoGasm is more rigid to photography, pureEvo has more flexibility.
So far, this is the best realism model I've tried, and I've tried a lot. Realistic vision, Cyber âârealistic far behind. Almost never distorts faces, 100% realism, rarely distorts bodies, understands complex queries well, and pictures are extremely difficult to distinguish from real photos
Thank you!
Hi, Thanks for the great work!
I have a problem using epicRealismPix2Pix model with diffusers library while I don't have any problems with ui version. When I use StableDiffusionPipeline or StableDiffusionImg2ImgPipeline, It says that "Trying to set a tensor of shape torch.Size([320, 8, 3, 3]) in "weight" (which has shape torch.Size([320, 4, 3, 3])), this look incorrect."
Can anyone help me? Thanks : )
The best real photo model together with epiCPhotoGasm. No other realistic model comes even close.
weird but I can not create male person. I wrote male,boy etc but I got girls only
This worked fine for me "man on basketball court jumping". No negative prompt was given.
Great model, I am having trouble because there's always the same face in the image, but other than that, works wonders.
Probably u are using ADetailer or face restoration?
@epinikion I'm not and I also got also very minimal face variation.
@epinikion No, I was not using ADetailer, also this issue applies to Natural Sin RC1 VAE, Natural Sin does not have "the same face" issue and I'm getting varying faces. Which means: great job all around. I wish I had the skills...
getting great results but colors are very dull/low, how to fix this? ... I'm new to this
Use the embeddings Epinikion has on his page. There is one that specifically adds color. You can also add 'deep rich colors' to your positive prompt, or 'saturation', and in your negative prompt add 'muted, washed out, dull'
Regular customer here, I just keep coming back to Natural Sin RC1 VAE no matter how many other models I try for realism (even your own other ones). Natural plays so well with custom LoRAs trained on it and the renders really do come out looking the most natural with a real photo feel. It is by far #1 in realism models for me......with a caveat that it does start to fall short and get confused once you start really getting into more hardcore/couple stuff, but it even handles hands and feet so freaking well! Hopefully you have another version planned for Natural with that in mind, but nonetheless thank you for this! Amazing work!
Better than any XL!
A million times better than anything XL.
This model delivers mind blowing results, just like the Lullaby-Version it is extremely responsive and satisfying in terms of clothing, locations and lighting. It's like the AI is connected to my brain, I literally get the results I imagine. The only issue for me is, the model comes up with one face and one face only and I am not using any other gizmos like After Detailer or Upscaling or anything else that could influence the result. Not even when I use a self trained Lora I'm able to overwrite the face without sacrificing the image quality. Other than that, this is easily the most realistic model out there and way... waaaaaay better as any XL model.
the problem is not the model but probably your prompt, tokens like "girl" or "woman" will always create the same face. the AI has been trained with thousands of images with the word girl or woman so its hard to get a different face with just one token. try to use different words and the face will change.
@Atega I wish it was that easy but it doesn't matter if I call my subject a girl, a woman, a blonde, or a brunette or a paleontologist or even "Eminems bastard daughter" that's the same facial structure with the same features, browsing through the posted images, unless someone rooped a different head it appears to be the same nice lady. That being said, face swap and roop is always and option and doesn't make this model less impressive.