A merge of all three models (Chimera, Center and Kroma), with some Flux.1 Krea attention layers sprinkled in for good measure.
Description
FAQ
Comments (17)
Ok, you are going full mad scientist now...
yes :) I got one more which is a fixed Krea2 turbo with some Raw and Qwen attention layers but did not test that yet
@TiwazM I need to get another drive labeled TiwazM Models at this rate - good thing I can only run the 12gb ones
@xFennec777 using a NAS for storage ;)
Have to agree, this is some mad-scientist stuff, and it's AWESOME! :D
Incredible did you mix all 3 kroma models in one , that is big project, how about NWSF does it need denial filter ,
no should be fine without
@TiwazM Great i hope we see this model on seaart.ai soon , tnx
Which kroma version did you use for the merge?
version 0.3
thanks i try it, does krea 2 loras still work fine ?
I did try a character and style one no issues
Amaizng output! I have run some tests with the turbor LoRA at 0.25/0.3 strength, and the photographic outputs are GREAT at 8 steps! I have to test with my illustrative LoRA (but I have liked the slight softness of the 8step with it).
Basically doing this as 8step + Turbo = 40s for a 1600x1200 image where as 12step = almost a minute at 1600x1200.
@xFennec777 🤔 I did not try a turbo lora on top. Also yeas I might not notice I use deconvolution sharpen like 90% of the time 🤭 From the KJ nodes if your curious, its in his sharpen node.
@TiwazM Ah nice! I noticed you run 12 steps (which I CAN do, but am too impatient lol)
@xFennec777 yes 12 a lot sometimes 8 or 10. And I get it after SDXL with DMD2 most feel slow 😆
Somehow merging those three models plus the secret Flux.1 sauce created something next tier - style, robustness and prompt adherence are truely great!