RA-Mix: Ri-Mix Style Merge
RA-mix was created for Phinjo’s 26k follower celebration bounty. I was honored to take second place in the event, and with permission, I’m now sharing the merge with the community. @phinjo
Models used:
A-mix:
https://civarchive.com/models/1692152/a-mix-illustrious
Ri-Mix Pi v2:
https://civarchive.com/models/996495/ri-mix-pony-illustrious
What is this merge?
Ri-Mix Pi v2 has great painterly detail, strong backgrounds, good lighting, and more interesting composition. It also has a dynamic western influence that gives images a different feel from a standard clean anime checkpoint. The tradeoff is that the facial area can sometimes come out softer compared to the rest of the image.
With RA-mix, I tried to influence Ri-Mix Pi v2 with the cleaner anime style of A-mix while keeping the parts that made Ri-Mix interesting in the first place. The goal was not to erase Ri-Mix or turn it into a standard anime model, but to clean up the facial structure while keeping the painterly detail, background strength, lighting, and more dynamic composition.
RA-mix is basically my attempt to balance those two sides: the cleaner face and anime polish from A-mix, with the western-influenced composition, lighting, and painterly detail from Ri-Mix Pi v2.
There was a little bit of tuning for the final version I'm posting but I think it turned out nice. I'm very happy with it's img2img abilites, and it's color range is very nice when using it with this application.
I wrote a step by step compariosn of the checkpoint here:
I also made an article that breaksdown the process I used to make the bounty image for Phinjo's compatition here: https://civarchive.com/articles/30721/under-painting-in-ai-part-2
🔧 Recommended Settings:
A well-rounded setup for general SDXL prompting.
Steps: 30
Sampler: euler_ancestral
Scheduler: karras
CFG (Guidance): 5
Description
RA-Mix V2.0
V2.0 is more of a refinement of RA-Mix than a complete overhaul.
One of the main changes was bringing some of the more realistic and 2.5D influence back into the merge. V1 leaned a little heavier toward the anime side than I originally intended, so I wanted V2 to have a wider stylistic range while still keeping the overall look and character of RA-Mix.
I also spent some time adjusting the merge layers to improve stability. With V1, I would occasionally run into structural issues during generation, so part of this update was focused on cleaning that up and making the model a little more consistent.
Lighting and depth also seem to have improved with the changes, particularly when pushing the model toward more detailed or less strictly anime-style images.
Overall, V2 isn't meant to be a massive change from V1. The goal was mostly to give RA-Mix a little more stylistic flexibility, bring back some of its realistic and 2.5D range, and improve the stability of the merge without losing what I liked about the original.
FAQ
Comments (2)
V2 is batting 4 for 4 so far for complex styles, that alone puts it in the pantheon for me. Thanks very much.
Glad you're liking it. Been trying to figure out how to give a checkpoint merge a nice built-in style without sacrificing is flexibility. With my img2img method, the more a model can pull from, the more creative it can be. Part of the reason I make sure my models can do males and furry designs.