Follow me on X/Twitter for exclusive content: https://x.com/Citron_Legacy 😁
Wield a rainbow's radiance,
Create Colorful Anime Treasure with CAT!
This Checkpoint (attempts to) bring Illustrious knowledge/power to Pony. The goal is to be compatible with Pony but also have some Illustrious magic.
CAT PDXL v5 - (Pony Diffusion V6) x (Zankuro_Lora & Afrobull_Lora)
CAT PDXL v4 - (CAT PDXL v2) x (Collection of my best Pony Loras)
CAT PDXL v3 - (CAT IL v4:0.7) x (CAT PDXL v4:0.3)
CAT PDXL v2 - (CAT_Illustrious_v10:0.2) x (WAI-ANI-PONYXL:0.8) - It also has the VAE from CAT_Illustrious_v10 baked into it.
Description
FAQ
Comments (6)
Pony?! wuzzat?
(tbh, my first thought was "yep checks out that now that citron is the undisputed King Of Zit loras he now goes looking for another frontier". And, Pony is, like, a dead wasteland these days. So yeah it's a frontier.)
Pony is like beating a dead horse with another dead horse. ROFL Just kidding!
I know this is not very popular but I had an idea I wanted to try. Most of my Pony loras aren't getting any attention so I'm guessing this wont get used.
HOWEVER the top 4 creators for Pony are always displayed on CivitAI's homepage so its apparently something worth investing in 🤣
As for ZIT, I'm cooking something on CivitAI's trainer that has been running for almost 24 hours and I have two more things that probably take at least that long too.
I want to see if I can overcome the "doesn't like anime" problem by training a 1000 image dataset for 6000 steps. My 3000 step lora is taking 24 hours to train so I guess the 6000 step one will take several days lol.
@CitronLegacy lol, yeah good hack to win the leaderboard. (but you're on there already for landscapes, and last I checked for smth else too! is there a leaderboard for "most on leaderboard"? hmmmmm).
Cool about the toomany-steps ZIT loras. (any training that has it's own Danbooru tag for length is cool ^_^). Good luck.
Is there any place where you trainers at the leading edge of ZIT share ideas/results? just curious. Because I follow a few people here that're trying ZIT training, and you're all making comments about still testing to figure out its parameters.
@ravemry9 Lol thanks for noticing! Yes I'm #2 on Architecture and ZIT and #1 on Vehicles, but that doesn't get my name on the homepage.😅
As for sharing ideas, I'm notoriously bad about not asking for help lol. I've basically learned everything on my own for my entire AI career. I always learn things the hard way or randomly when I see someone's technique. For example, I want to learn how Navimixu makes world-morphs but I never ask 🙃.
I'm happy to share what I know but its always ends up in random comments/DMs like this.
I even have a nice excel document which lists my training parameters for many of my loras so I know what settings produced good results but I have no where to share my data lol. (For example, if training an anime with tags instead of captions 3000 steps produces pretty good results but less than 1000 tends to be janky. If training something realistic with tags, like Seraphina, then 1000 steps for 39 images produces great results)
I was thinking about publishing an article with all my research notes but my ideas are still very "rough draft"
@CitronLegacy oh that's hilarious, I'm the same way in so many areas, I just never get around to asking for help and then spend loooong amounts of time getting there the hard way, trial and lots of error.
I've actually gotten better when I started AI gen, smth about this site encourages my asking friendly users. I sort-of feel like we share a space in pioneering a new frontier, it gives me fuzzy feelings ^_^
Your excel doc sounds precious! take good care of it! thinking back on your career on this site, there's way too much hard-won experience in there for it to get lost. (and maybe one day we'll convince u to post it... gentle pikachu poke)
I've been thinking to try an Illustrious 1st sampler for composition followed by bong sampler res_3s bong_tangent at 0.3 denoising at 3 steps using Z-Image-Turbo to speed up the zit generation i think it's an underrated upgrade to SDXL era upscaling.