Its nothing too special, its based on Kroma (Krea-2) - OPD | Krea 2 Checkpoint | Civitai by Lodestones. Kroma suffers a little from fine detail formation, my goal was to fix this while preserving Kroma's diversity. It should also have better LoRA compatibility with much less artifacts. I am not gonna write with AI and pretended that its some state of the art complex thingy, but I did manage to fix some issues so I thought I should share it. Also huge shoutout to Lodestones for his remarkable progress! Do check out his model even if you decide mine is waste :P (For Lodestones: If you want this model removed please DM me as I didn't ask for permission and I will remove it.)
Sampler settings (Recommended): Euler/Beta - 10 steps (Can go up to 12 steps, can make much detailed in some instances)
Pro tip: Use SNOFS 1.5 for better sexual positions with shorter prompts. I did create a version with SNOFS merged in, but SNOFS decreases facial diversity by a lot! I am trying to figure out the blocks of SNOFS that effect only the poses and only merge them, if I figure it out I will share. Adding SNOFS does make it super prompt adherent for NSFW especially for anal and vaginal differences in complex poses.
On a sidenote: With all due respect fuck everyone who is just merging a bunch of free - publicly available LoRAs and checkpoints and asking Buzz for it and calling it CHECKPOINT TRAINED. I am sick of this utter bullshit that has started to ruin this site. This rant is NOT meant for creators who really put in effort to train their own models and LoRAs!
Description
FAQ
Comments (20)
Look good! Waiting for bf16
Snap. Waiting for bf16. Looks incredible!
@EuroCluddy I will try to upload tomorrow
added!
@mqblbhai thanks a lot! It's a great model
Kroma is so nice already, and this is on the radar for sure. It looks like it'll let you keep fast euler gens but always keep that casual look. The only real question going forward is that NSFW capability. All I can hope is this is like Chroma-30 was, still twenty or thirty updates away from being "finished". can only hope really
Yes! Fingers crossed hope he keeps making it even better. I heard the Krea team gave him the true RAW model for him to train. The NSFW is already very good, it just requires a good prompt but if NSFW is what you are after try using this + SNOFS, it gets you close to ChatGPT level prompt adherence for NSFW already.
@mqblbhai After messing with the teacher for an hour or so, it's just not worth it without the built in turbo yet anyway. He might be able to get it wrangled but at this point its like a magic spice lol. Anyway thanks for putting this together, I've simply given up on using the teacher alone and am going to try this out lol. having the int8 speed boost should be worth it alone
tell me a little more about chroma-30. I Use the final version but never tested all the other versions. is it the hidden gem?
@WhiteRider no that was just about the version (maybe 27) that it became stable enough to become pretty popular. People always point back to those versions as they were likely the ones they trained their early chroma loras on, or worked better with the Flux loras they already had
@makiaeveli gotcha gotcha. i have a load of flux loras. i think ill test just for the fun of it. thanks!
Can you do a FP8 version?
That's fantastic, the perfect checkpoint. Is there a compatible workflow?
it's just any Krea2 workflow without a turbo lora. However It does need more noise reduction than Krea2, so you might need to tweak a few settings
I actually use Forge Neo, its simpler and this checkpoint works great at 1 pass, so no need to do a complex workflow. You can use the default Krea2 Turbo workflow with Euler/Beta 10 steps and thats about it.
Very good checkpoint. The prompt adherence and lora compatibility are fantastic. My go-to. Converted to nvfp4 fine and retained excellent output quality. This is a winner.
Thanks!
This checkpoint really stabilizes the OPD version of Kroma. Can you share your recipe - for educational purposes?