>>64451033my model is up on Civit, LS Vividus v3
HLL version is hll6-lyco-c5b
I'm using 2 LoRAs for the style:
the miwol/Wool lora that a anon here trained and posted some threads ago at 0.4
mamaloni lora at 0.2
the colours are coming from Dynamic Thresholding along with its interpolate_phi combined with the effect of Recenter
the final touch is from LCM, which has a tendency to clean things up
Normalisation is added where it makes sense to smooth things out
my current workflow is not fully optimised, so the initial gen tends to overcook the face (especially on tags like lips, smile) which I downweight to reduce the issue
Even if the inital gen is a bit jank, LCM will clean it up for you, so don't bother too much to get it perfect
tidbits:
Recenter makes the colours darker, but also seems to imporive composition
my entire workflow is built around 3M SDE and sgm_uniform. you may not get the expected results if you use a non-SDE sampler
I am using custom noise, which also has an effect on composition, and 3M SDE works well with this noise(3M is a bit weird on normal noise)
I use separate prompts for the LCM step, because the composition is coming from the Img2Img, and quality tags and negative embeddings can be added to refine the result
>>64451209kek