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 š
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
ā No VAE needed but it is better to use one for more vibrant colors
ā Feel free to leave Reviews and Samples - and always have fun creating ā¤
Description
As V3 is pretty perfect, i tried to add less Photoshop/retouch Effect, even more skin enhancement, tried on better eyes again and overall prompt variety.
Let's test it out
FAQ
Comments (25)
Hey! Great model! One question though. I want to create images of women but when specifying breast size, it tends to make them topless, even more than in V3. Is there any negative prompt you could recommend for the purpose of not making them naked?
Hmm didnāt test that but there are some breast size Loras on civitai, maybe try them out.
Try clothing adjuster LoRA by Leosam, Maybe at (-0.5) weight.
(Please refer to my XYZ plot Images for reference)
Epinikion you know I love you but for my use, V3 >>> v4.
Worse TI, Loras, Lycoris, famous people compatibility: This is such a deal breaker for me. It has a slight bias toward a certain face aesthetic; a long slim face, with longer chin, and unnatural (even creepy) eyes. Check XYZ plot #5 and #6 to see long face Kim Jong Un and Donald Trump. I'm not sure what is going on with epicrealism evo v4 and realistic vision v5, but they both have this same problem. They are changing peoples' faces. This means you can't produce the look of any celebrity using TI/LoRAS/Lycoris/prompt with the level of accuracy of V3.
Skin texture: I'm not seeing an improvement with V4 tbh.
Lighting: Flat and bright (though not always).
Landscape and product photography: It may perform better in landscape and product photography, although I'm not an expert.
Summary: V4 is a great checkpoint. But it has a certain bias in portrait generation. Fortunately, V3 is already the peak of sd1.5 so really, it's difficult to expect anything more. Thanks for your continual work on sd1.5 models Epinikion, you keep winning my dear. Can't wait for SDXL epiCRealism!
Thank you! As I said in the Version info I also thought v3 is perfect. There is a bit more variety in this one and less āeffectsā also the skin details depending on prompting and should be more rigid than v3 overall. Didnāt test Loras that much with the model. For SDXL, first I have to test it out in depth, hadnāt much time for that, prompting seems fundamentally different from 1.5. So it will not be soon I guess, since Iām on a heavy workload privately. But Iām sure there will be an awesome improvement by someone else anyways š
@epinikionĀ V3 is godlike I agree. And no worries epinikion, life comes first always. SD1.5 took months to get to this point. We will patiently let the master work.
Tutorial for this model epiCRealism easy and Fast
Is there an inpaint version of v4?
Or did you just do the normal merge with the 1.5 inpaint model for the other versions?
Bcs I know some models have a specially trained inpainting one
No, I just merge them with the 1.5 inpainting model. So everyone could do it themself, maybe I will upload it the next days.
@epinikionĀ Please do. Most of us don't do anything for ourselves. š¤£
@epinikionĀ ok thx. that I can do myself then :)
This model is just amazing, may I ask one thing, how you curate the images you use? I been training models for a while but I can't get the precision I see in models like yours, you finetune several times, any tip you can give?
It certainly produces great results.
I just have some issues with skin textures and skin color matching while inpainting tried multiples samplers and settings but not giving up yet
@firedelta5Ā have you tried upscaling instead of inpainting?
Use high-quality images; that's why there is no need for any 'masterpiece' or '4k' in this model.
Precision is variation of images.
I'm a complete noob but while playing with SDXL I got a bit annoyed by the terrible smooth faces and overall look (might be me). But if I run them afterwards using img2img with PureEvolution V3 it changes to a way more natural appearance.
@epinikion I have some issues with the 4.0 Edition.
3.0 and 3.5 colorfix are absolute insane. I uploaded two Batch runs - first with same seed nr. , second with -1 keep. On the 4.0 the skin texture is not so perfect as the last two editions. Also the tendency for portrait style seems to be back / higher.
Btw: The mystery edition is absolute insane. It has some hand / arm issues and a little "boring" (sry) for situation backgrounds. But the skin texture, reflections, people bias .. absolute gold. Is there a update planed?
As always.. thanks for your great work.
Hey thank you! Since the 4.0 bias is a bit different it probably takes another seed to shine against v3. Iām not really planning any updates on mystery, but maybe there will be if I have a checkpoint that doesnāt fit anywhere in epiCRealism š
V3 inpainting is my go to model - fantastic! Any chance you will a V4 inpainting version? Thanks.
Itās possible to do it yourself. Iāve done one based on V4 and it works really well. Instructions on Reddit. https://www.reddit.com/r/StableDiffusion/comments/zyi24j/how_to_turn_any_model_into_an_inpainting_model/?utm_source=share&utm_medium=ios_app&utm_name=ioscss&utm_content=2&utm_term=1
@vinterAIĀ Interesting! Fancy uploading it please?
Maybe for the eyes a good lora would be more beneficial. There are a couple decent ones, but I haven't seen a good eye lora yet, that really delivers with comprehensive prompts.
I have many problems with blurried faces in duo, trio or groups of people... What I can do, is the model or it's my prompt?