Colorful

Added photorealism, with no loss of artistry

Mixed with https://huggingface.co/thiros/YuzuLemonTea

Script is here: https://github.com/recoilme/losslessmix/

The anime models are very dark, so I mixed with the colorful_v1.4, start from simple mix with some share (0.5 or 0.7, don't remember): result

And mix result with different share about 10 times, before anime was converted to semi - real.

python3 weightedsim.py colorful_v14.ckpt yuzulemonmilk.ckpt --s .77

python3 weightedsim.py colorful_v14.ckpt merged-0.77.ckpt --s .76

python3 weightedsim.py colorful_v14.ckpt merged-0.76.ckpt --s .75

#After that i mix again with good old, not anime models like unstableinkdream

python3 weightedsim.py colorful_sr.ckpt unstableinkdream.ckpt --out ttt

new merging algorithm:

https://github.com/recoilme/losslessmix

Please stop asking me how to run this and how to work with it

this models was removed:

• portrait.ckpt

• izumi.ckpt

• vintedois.ckpt

• unreal.ckpt

• hunter69.ckpt

• analog.ckpt

• wavyfusion.ckpt

• uberRealisticPornMerge.ckpt

• has3dkx.ckpt

Mixed with brilliant unstableinkdream

Agorithm: losslessmix

Completely redesigned. New unique mixing algorithm. All parts with free license, merged with script:

python3 merge3_mindiff.py a.ckpt c_openj.ckpt d_f222.ckpt --out tst

models:

• a_stable_diffusion.ckpt

• c_openjourney.ckpt

• d_f222.ckpt

• h_eimisanime.ckpt

• i_portrait.ckpt

• m_vintedois.ckpt

• q_unreal.ckpt

• r_hunter69.ckpt

• u_wstanding.ckpt

• y_rpg.ckpt

• t_analog.ckpt

• x_uberRealisticPornMerge.ckpt

first release

I wanted to create something new. First of all, I blended five or so inconsistent models like anime with 3D, etc. using custom blending algorithms (minmax & maxdiff). It doesn't make any mathematical sense. The result is an absolute trash, with unpredictable results, but bright and acidic

After that, I tried to stabilize the model by adding new, trained models (not mixes), already with the classical weighted sum algorithm, trying to keep brightness and variety.

I don't know exactly how well it worked. I do not specify weights, because in half of the cases there were none, the algorithm just loosened the tensors, but the composition involved these models:

colorful = abehcmiqruwx.ckpt

• a_stable_diffusion.ckpt

• b_redshift_diffusion.ckpt

• c_openjourney.ckpt

• d_f222.ckpt

• e_minillbgr.ckpt

• h_eimisanime.ckpt

• i_portrait.ckpt

• m_vintedois.ckpt

• q_unreal.ckpt

• r_hunter69.ckpt

• u_analog.ckpt

• w_wavyfusion.ckpt

• x_uberRealisticPornMerge.ckpt