i like to play around turbo loras, this one turned decent, tho it feels like having a wet cardboard lora sometimes.
there's quite the instability but i can't touch it much because other loras add too much style as a fix, its either glorious or just bad, just the kind of variety i wanted (wild) in the name because of that reason, every other turbo lora has the same bland generic same pose style so yeah, if you can remove the cardboard texture feeling do tell me
most images are made at 10 step er_sde, kl optimal
i'm using cosmos predict distilled lora with some blocks removed
1.0,0,1.0,1.0,0,1.0,0,1.0,1.0,0,0,0,0,0,1.0,0,0,0,0,0,0,0,0,1.0,0,1.0,1.0,1.0
and a mix of not-so-styled rdbt's
rdbt 0.2 at 0.2 strength with 1.0,1.0,1.0,1.0,0,0,1.0,1.0,1.0,1.0,0,0,0,0,1.0,1.0,1.0,1.0,1.0,1.0,0.8,0.8,0.8,0.8,0.8,0.8,1.0,1.0
rdbt 0.39b at full power, no blocks removed, its a nice lora but incomplete, the other loras balance it
merged and lora extracted with kj node at default bf16 64 and 32 rank
the way i mix-match is in the workflow of the images with the smart loaders too
i add some styles before my prompts to guide it to be more of the anime styles i like
alright just posting here to share in some servers xD
Description
FAQ
Comments (4)
"if you can remove the cardboard texture feeling do tell me"
sure thing bub, let me post a few pictures with a workflow
here's the auxillary Stuff while they get approved:
https://huggingface.co/sorryhyun/anima-turbo-2step/tree/main
https://github.com/pamparamm/ComfyUI-ppm
https://github.com/sorryhyun/ComfyUI-Distilled-ResShift
It just so happened to be that this silly buzzsaw-squiggles-turbo LoRA does what this other silly turbo model was lacking, so if you smash them together - you get surprisingly crisp and wildly varied outputs in a few steps (4/5 +1 +1xCFG [-> 7-8 sampling steps])
whether a particular setup responds better to 4 or 5 initial steps kind of varies
i'm glad that the lora helps your case at least, quite impressive for 5 steps
tho my problem remains, i tried that lora but the problem persist, but the ppm/reshshift seems to fix it somehow (?, i was thinking the lora was the solution but the reshift was, will do some more testing but thanks for the tech, sadly i can't block weight merge that fix/shift into the lora xD
i like the starting composition of 1024x1024 images so i'll just find a way to adapt it into my workflow fun next time, thanks VeerGeer xD
also saving as animated webp seems to slightly heal that cardboard feeling XDDD maybe because of the quality settings somehow xD back to square 1 for me but will keep the techs in mind
Thank you for this! Of all the turbo loras I tested, this one works best :)
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.














