I wanted something I could use out of the box for the style I enjoy, mobile or otherwise, without mashing a bunch of lora prompts and minimal prompt wildcards to remember them or whatever. This is kind of experimental. It's not designed to win any awards, or get on the front page.
This is a merge of Duchaiten Retro and NijiDiffusedMix4. It's weighted towards Retro, since Niji is pretty insane (in a good way) It has some loras merged into it, EnvyBeauty16, Detail leco at low strength, and EpiOffset at moderate/high strength.
For people it's good with clipskip 2, the rest its fine with 1. But YMWV.
Don't worry about the controlnet in the metadata, I have a bug where controlnet shows up enabled after every restart, even with blank data.
Description
FAQ
Comments (10)
Weird, using same generation settings & prompt and getting low quality images. Nothing near the realism in your pics. using ComfyUI
Edit: Can confirm it works fine for me in A111. I tried the custom node that is supposed to allow for a111 weight interpretation but it is throwing errors. alas.
I actually want to know if anyone else has this problem, as I just redownloaded the model and it seems fine. It could be how comfy handles embedded loras?, or just comfyisms. I am in A1111 so eager to hear others
@Sophorium ill test now
Comfy uses different weights, is generally the issue. There's a way to get it to use A1111 weights out there, but someone who knows more can give you the details!
@Aishavingfun I just ran a whole series on Vixons Puffy Dress with it, seems ok. Also I used your Natalie in there =)
https://www.reddit.com/r/StableDiffusion/comments/18vds0p/what_makes_a1111_and_comfyui_produce_different/ maybe that helps? I dont know, I only run Comfy when I want to be a SDXL masochist unfortunately.
I get good quality, even better. Using 2 on clip skip and upscaling with 4xUltrasharp_4xUltrasharpV10. Using A1111 though.
I just tried Comfy TODAY and was getting really shitty gens basing it off of A1111 weights, had no idea there were nodes to do that. Guess I'll try again, I'm just much more comfortable with A1111... ironically
Seraphelle slays! It's been great for high-key lighting, low-key lighting, and can achieve wonderful contrast. I used a mix of photographic prompts and concept art prompts with no issues.
Thank you for creating and sharing this checkpoint.
Thank you for your kind words! It definitely is a fun model to use. I use sinistrella for most of my benching loras because it has less bias when demoing, but this is the one I use when I want to let loose or get more moody. Seraphelle has a edge to it that I'm glad you enjoy!