DETAILS
Epîc Diffusion is a general purpose model based on Stable Diffusion 1.x intended to replace the official SD releases as your default model. It is focused on providing high quality output in a wide range of different styles, with support for NFSW content.
Epîc Diffusion is a heavily calibrated merge of SD 1.4, SD 1.5, Analog Diffusion, Wavy Diffusion, Redshift Diffusion, Openjourney Diffusion, Samdoesarts Ultramerge, Elldreth's Dream, postapocalypse, Inkpunk Diffusion, Ghibli Diffusion, Mo Di Diffusion, Archer Diffusion, Classic Animation Diffusion, Arcane Diffusion, Van Gogh Diffusion, 3DKX, HASDX, Flexible Diffusion, Cinematic Diffusion, Shady Art, dvMJv4, dvAuto & mj-v4-look + some dreambooth trained models of my own, blended and reblended multiple times until I got the quality & consistency I was looking for...
MORE INFO
More example prompts & the diffusers version of Epic Diffusion can be found at the Huggingface repos (v1.0 - v1.1)
If you like my work, feel free to buy me a coffee.
Description
FAQ
Comments (23)
I really wish more people making big mixes like this paid more attention to what mixes they are actually using as ingredients. Some of them, such as Protogen (which is not exactly their fault, but Seek.art's and Dreamlike's) include restrictive licenses that poison all further derivatives. Which is a shame, because the mix is rather good, but I'd never use it as an ingredient in a mix of my own that I'd intend to share.
sounds like Seekart is actually removing their license. So that’s cool.
@shadow97 :
For 1.1, Protogen has been removed as a dependency and several others have been added. AFAIK none of the remaining dependencies depend on seek_art or Dreamlike.
Question here is how anyone will know you base on this or that models when all drinks from Midjourney... Simple as that. You can always train MJ images into a base model you want, make another with same base as previous and mix them. Train on top of them, simple as that,
What is the difference, in downloading, between pickle and safetensor? I have been downloading the safetensor. Perhaps pickle download downloads a .pkl file?
The pickle is a *.ckpt file. It is the exact same model as the safetensor version, but many people prefer a satetensof version over a *.ckpt file because the latter allegedly can contain a virus or something like it...
Rilevant differences between the 1.0 and 1.1 versions ? Really great model anyway.
Relevant differences?
A few models were removed and various others were added in my merge.
That, and dozens of additional hours invested in testing and merging my models.
Overall, though, the concept remains the same. I want this to be the most versatile and highest quality general purpose model out there. And I feel all the extra hours spent on tweaking 1.1 gives it a minor edge over 1.0 in this regard... but just a minor one.
Not sure if there's going to be any new versions after 1.1. There's only so much you can optimize & tweak. And it's taking ever more time to get ever more minor improvements. For the time being I'd rather experiment with creating some more style-specific models based on my own Dreambooth or LoRa training.
Thanks !
There is no triggerwords for this?
You don't need any trigger words, no.
Are there negative embeddings recommended for Epic Diffusion to solve distortions like bad hands and extra body parts, etc.?
try easynegative
I would love if there were a inpainting version.
No idea how to make an inpainting version. Can you recommend any sources on that?
@johnslegers I found this post on reedit with instructions how to make it in AUTOMATIC1111 https://www.reddit.com/r/StableDiffusion/comments/zyi24j/how_to_turn_any_model_into_an_inpainting_model/
@grtmate I'll check it out. Thanks for the info!
Very impressive realism, lighting, poses, etc.
One down side, though, is that a lot of the "female" faces in the examples posted, have something particularly man-ish about them though.
Maybe it depends on the ethic group. Dunno.
I made a new post with some Norwegian & French girls / women.
Do those still look "man-ish" to you?
@johnslegers 1st one... ehhh. 3rd one yes. 4th one, dolph lungren vibes, so yes. 5th yes. 8th yes. 9th yes. Probably a big factor is the hollowed cheeks. Secondary is overly prominent man-chin
This was one of my favorites for a long long time. I wish you did an XL version. It had a very unique way of rendering photos.
Thanks a lot for the kind words.
I actually recently started working on my first SDXL model around the time you posted your comment, but I got a bit sidetracked digging into FLUX, ComfyUI & custom Python coding.
Maybe I'll continue my work on the new SDXL model in the near future.
Your words are inspiring.
@johnslegers hope you do if you feel like it. i don't know what you used for a dataset but even in 1.5 it works better than most XL models when it comes to prompting "analog gloomy photo". good luck sir
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.