Some model combinations blend nicely together and some do not. This is due to a number of factors, such as the tags (descriptive words) used to identify parts of the pictures in the training sets, but it also matters which version of Stable Diffusion the models were initially made from. For example, F111 was based on Stable Diffusion 1.4 and blends better with other models made from Stable Diffusion 1.4, while F222 was based on Stable Diffusion 1.5 and tends to blend better with other SD1.5 models, etc.

Sometimes mixing two models will result in getting the benefits of the tags used in one model, with some of the style and coherency of the other. Other times, the results are something along the lines of the failed ‘Ripley’ experiments in ‘Alien: Resurrection’ and should probably be burned with a flamethrower.

That said, when combining any two models and observing the blends from 0-100%, there can be combinations in which visual coherence occurs and produces striking new styles, even if most of the time the outputs for a given combination are total trash.

I have made some custom Python scripts that generate images based on a selection of my favorite prompts using a combination of any two .ckpt files in my folder, at a random blend ratio (from 100%A and 0%B all the way to 0%A and 100%B), with a random seed and a random CFG between 16 and 18. I’ll run this overnight and look at the results in the morning, and any images that resulted in unique coherent styles I copy into a specific folder.

I then run another Python script that looks at all the images in this folder and counts the number of times a model was used to make these promising combinations.

The following is the current results:

Experiments

 "['Anything']": '44',
 "['BerryMix']": '50',
 "['CandyMis']": '47',
 "['EimisSem']": '21',
 "['F111.ckp']": '12',
 "['F222.ckp']": '30',
 "['MMD V1-1']": '26',
 "['Multiple']": '31',
 "['RD1412p.']": '4',
 "['Systemy-']": '13',
 "['The Defi']": '3',
 "['animeful']": '3',
 "['iridium.']": '5',
 "['nai-anim']": '23',
 "['samdoart']": '14',
 "['sd-v1-4.']": '5',
 "['v1-5-pru']": '20',
 "['wd-v1-2-']": '7',
 "['wd-v1-3-']": '1',

Based on these results, it would appear that ‘BerryMix’, ‘Anything-V3.0-pruned’ and ‘CandyMissionBerryF222-Hassan_v1-2’ produce the most frequent “good” combinations.

I will regularly run this script as new models are released and post the results here.

(Last updated 11/24/22)

F111 : 36% | BerryMix: 64%

Combinations to try:

Systemy-yohan1754 : 60% | v1-5-pruned-emaonly: 40%

sd-v1-4 : 66% | Anything-V3.0-pruned: 34%

MMD V1-18 MODEL MERGE ALPHA : 75% | nai-animefull-final-pruned-model: 25%

Anything-V3.0-pruned : 67% | F222: 33%

BerryMix : 59% | F222: 41%