Hugging Face's logo

ShinCore
/
MMDv1-18

Text-to-Image
stable-diffusion
stable diffusion
Model card Files Community
2
MMDv1-18 / README.md
ShinCore's picture
ShinCore
Update README.md
b59fdc3
preview code
|
raw
history blame contribute delete
5.68 kB
metadata
license: creativeml-openrail-m
tags:
  - stable-diffusion
  - stable diffusion
  - text-to-image

MEGA MERGE DIFF (MMD) VERSION 1-18. 18 MERGED MODELS IN ONE

ANNOUNCEMENT:

  • DUE TO THE FACT THAT I CANNOT SEEM TO CATCH A BREAK AND GET SOME TIME TO ASSUAGE MY OWN INSECURITIES ABOUT THE QUALITY OF THIS MODEL, I AM JUST GOING TO RELEASE IT.

FIRST MODEL RELEASE: MMD V1-18 MODEL MERGE ALPHA:

SUMMARY: MMD V1-18 A MEGA MERGE OF SD 1.5 AND 17 OTHER MODELS. IT IS INTENDED TO BE A GENERALIST MODEL, NOT FOCUSED ON ANY SINGLE GENRE OR CATEGORY OR STYLE OR SUBJECT. THERE ARE ALREADY A PROLIFERATION OF GREAT MODELS OUT THERE, COVERING A BROAD SPECTRUM OF CONTENT. HOWEVER, THAT ALSO CAUSES A PROBLEM IN THAT WE HAVE A PROFLIERATION OF GREAT MODELS OUT THERE THAT DO ONE OR TWO THINGS REALLY WELL, BUT THATS KINDA IT. OTHER THAN THOSE ONE OR TWO THINGS THAT HAVE BEEN ADDED TO THE BASE SD MODEL, ITS NO DIFFERENT THAN ANY OF THE OTHER MODELS.

MMD WAS CREATED TO ADDRESS THE ISSUE OF DISORGANIZED CONTENT FRAGMENTATION ACROSS HUGGINGFACE, DISCORD, REDDIT, RENTRY.ORG, 4CHAN, AND THE REMAINDER OF THE INTERNET. IT ALSO TRIES TO ADDRESS THE ISSUES INHERENT WITH THE BASE SD 1.5 MODEL. NAMELY, PROBLEMATIC ANATOMY, LACK OF RESPONSIVENESS TO PROMPT ENGINEERING, BLAND OUTPUTS, ETC.

THE CURRENT SET OF MERGED MODELS ARE A CROSS SECTION OF MODELS THAT I FEEL IMPROVE AND ENRICH THE BASE MODEL. IN MY TESTS (WHICH YOU CAN TAKE A LOOK AT THE LOGS IN MY #EXPERIMENTS CHANNEL IN THE DISCORD I WORK OUT OF), THE MERGING OF A SPECIFIC SET OF MODELS HAS SHOWN TO IMPROVE HUMAN ANATOMY COHERENCY, INCREASE CREATIVITY AND DETAIL IN BOTH FORE/BACKGROUNDS, AND CAN BE MORE RESPONSIVE TO PROMPTING (PLEASE SEE PROMPTING NOTE BELOW).

  • DOWNSIDE: TRIGGER TERMS ASSOCIATED WITH SPECIFIC MODELS HAVE A DRASTICALLY REDUCED EFFECT. IF USING A TRIGGER TERM ASSOCIATED WITH A SPECIFIC MODEL, YOU MUST INCREASE THE STR TO SEE ANY EFFECT. THE MODEL CAN ALSO BE MUCH MORE SENSITIVE TO THE SETTINGS THAT YOU USE. I HAVE LISTED SOME RECOMMENDATIONS BELOW.

IMPORTANT: I DO NOT, IN ANY WAY, SHAPE OR FORM, CLAIM THAT THIS MODEL IS SUPERIOR TO ANY OTHER MODEL OUT THERE. NOR DO I FEEL THAT I AM SOMEHOW SOME KIND OF SD GURU AND AM AN EXPERT OF ANY KIND. MY INTENTIONS IN CREATING THIS MODEL IS FOR MY OWN PERSONAL GOAL OF USING IT, AS WELL AS OTHER AI TOOLS, TO CREATE A STREAMLINED WORKFLOW PIPELINE THAT WILL ENABLE INDIE SOLO GAME DEVS TO CREATE GAMES WITH GREATER EASE AND EFFICIENCY. I DREAM OF SOMEDAY BEING A SOLO INDIE GAME DEV, AND THIS IS MY WAY OF HEADING TOWARDS THAT GOAL IN AN INDIRECT FASHION.

I AM NOT A GENIUS. I AM NOT EVEN A GODDAMN CODER/PROGRAMMER/MATHMETICIAN. I AM COMPLETELY OUT OF MY DEPTH. I SOMETIMES FEEL LIKE THAT ZOOLANDER MEME, HOOTING AND POKING AT A COMPUTER THAT IS BEYOND MY LIMITED COMPREHENSION. I AM NOTHING MORE THAN AN OLD, TIRED, LAZY, AND GRUMPY BASTARD WHO IS SPENDING WHAT LITTLE FREE TIME I HAVE TRYING TO FIGURE THIS CRAP OUT SO THAT I DONT HAVE TO LEARN HOW TO CREATE A GAME FROM SCRATCH. I JUST WANT AN AI TO DO IT FOR ME.

NOTES ABOUT USAGE:

  • MODEL CAN BE A BIT HARSH AND RIGID, BUT PUT OUT SOME AMAZING GENS. 2ND RELEASE WILL BE LESS EXTREME.

RECOMMENDED SETTINGS:

  • IMG HEIGHT/WIDTH MUST BE SET TO A MULTIPLE OF 128
  • SET HIRES FIX TO ON. SET FIRST PASS HEIGHT/WIDTH TO HALF OF IMG HEIGHT/WIDTH.
  • PLAY WITH DENOISING STR, I SET MINE TO .69, YMMV.
  • MODEL IS SENSITIVE TO CFG. I USUALLY USE 12.5, BUT OTHERS HAVE REPORTED BETTER OUTPUTS AT LOWER/HIGHER VALUES. TRY THEM OUT. DONT BE AFRAID OF GOING REALLY HIGH OR REALLY LOW.
  • SET RESIZE SEED OPTION TO 512X512. (DDIM RESPONDS BETTER TO THIS SETTING. OTHER SAMPLERS MAY NOT. TURN ON AND OFF AND COMPARE)

PROMPTING:

  • MERGED MODELS USE A COMBO OF STANDARD SD BLIP/CLIP AND DANBOORU TAGS. I USE BOTH IN MY PROMPTS. TRY USING BOTH CLIP AND DANBOORU INTERROGATOR ON IMAGES, GET THE RESULTS FROM BOTH, AND USE THEM IN YOUR OWN PROMPTING. THEY SEEM TO REINFORCE EACH OTHER, THOUGH THIS IS DIFFICULT TO SCIENTIFICALLY VERIFY. I HAVE SEVERAL THEORIES THAT I INTEND TO TEST OUT, AS TIME PERMITS
  • USE NEGATIVE PROMPTING WITHOUT RESERVATIONS. YES, I HAVE SEEN STATEMENTS TO THE EFFECT THAT NEGATIVE PROMPTING IS LIKE A PLACEBO. ALL I CAN TELL YOU IS THAT, WHEN I REMOVED ALL OF MY NEGATIVE PROMPTS, MY GENS TURNED INTO A HORRORSHOW.

I COULD NOT PUT THEM BACK ON FAST ENOUGH.

2ND RELEASE, MMD V1-18 MODEL MERGE (TONED DOWN) ALPHA:

  • THIS IS THE SAME AS THE FIRST RELEASE,BUT I MERGED BACK IN 25% OF SD 1.5. MORE FORGIVING, LESS EXTREME EDITION. EVERYTHING ELSE IS THE SAME.

FINAL NOTE: IT HAS BEEN POINTED OUT TO ME THAT I AM USING ALL CAPS, AS IF I WAS SOMEHOW NOT ALREADY AWARE. AS I HAVE PREVIOUSLY POINTED OUT, I AM AN OLD, TIRED, LAZY AND GRUMPY BASTARD. MY PREVIOUS AND CURRENT PROFESSION HAS GOTTEN ME INTO THE HABIT OF USING ALL CAPS WHEN WORKING. DONT ASK ME THE REASONS, THEY ARE STUPID. AS IN, OTHER PEOPLE THAT I DEAL WITH. I AM PERFECTLY CAPABLE OF USING PROPER CAPITIALIZATION. I CHOOSE NOT TO WHEN I AM FOCUSED, GRINDING AWAY AT SOMETHING, HASTILY TRYING TO FINISH SOMETHING, OR AM IN A BAD MOOD BECAUSE SOMETHING IS IRRITATING ME.

SO BASICALLY, PRETTY MUCH ALL THE TIME.

I APOLOGIZE, BUT YOUR ARE GOING TO HAVE TO DEAL WITH IT.