Zimage Turbo - Foreskin

The goal is for this model to be able to fully understand the nude male.

Version 2 is rank 64 so the details are slightly better than version 1. It follows prompts better than prior versions. It is trained on:

• Nude Male

• Pubic hair (default is closer to a hairless body, you do not need to tag hairless but you can put pubic hair & pubes in the negative)

• Cumming/Ejaculating Cum works well

• Masturbating

• Inserting [object name] in foreskin

• Inserting finger in foreskin

• Grabbing foreskin

• Ass/ Ass hole (example: photo of nude male's rear focusing on the ass hole)

• Very Small/Small penis or Large penis

• Fellatio/inserting tongue in foreskin

• Fellatio/sucking foreskin

• Fully Retracted Foreskin. tag "fully_retracted_foreskin" at the beginning of your prompt

• Partially Retracted Foreskin. tag "partially_retracted_foreskin" at the beginning of your prompt

• Long foreskin - doesn't work as intended.

• Removed Foreskin (I did not train circumcised penis). tag "removed_foreskin" at the beginning of your prompt.

Pros of 2.0:

• Dataset was tweaked after 1.0 and 1.5 to fill in gaps for some concepts making them stronger

• Drawings/illustrations were recaptioned to avoid overwriting drawing styles as hard. The result are more realistic drawings.

• Trained with less aggressive optimizer parameters making it even more stable than Version 1.0

• It is more consistent with skin texture

• Dataset was tweaked to properly address leaning while sitting so that subjects no longer lean back as frequently

Cons of 1.5:

• Sometimes still ignores prompts

• It was not able to meet many of the goals

• Does not understand females well.

Version 1.5 is trained on:

• Nude Male

• Pubic hair (default is closer to a hairless body, you do not need to tag hairless but you can put pubic hair & pubes in the negative)

• Cumming/Ejaculating Cum works well

• Masturbating - works but the foreskin isn't well defined.

• Inserting [object name] in foreskin

• Inserting finger in foreskin

• Grabbing foreskin

• Ass/ Ass hole (example: photo of nude male's rear focusing on the ass hole)

• Fellatio/inserting tongue in foreskin

• Fellatio/sucking foreskin

• Fully Retracted Foreskin

• Partially Retracted Foreskin

• Long foreskin - doesn't work as intended.

• Removed Foreskin (I did not train circumcised penis)

• Very Small/Small/(default)/Large penis (does not work well)

The greens work more consistently. Yellow works sometimes. Red rarely works as intended, but you might get a good shot or two.

Pros of 1.5:

• Dataset was tweaked after 1.0 to fill in gaps for some concepts making them stronger

• Drawings/illustrations were recaptioned to avoid overwriting drawing styles as hard.

• Trained with less aggressive optimizer parameters making it even more stable than Version 1.0

Cons of 1.5:

• It is not consistent with the skin texture. The foreskin texture should be softer.

• subjects frequently lean back or are otherwise not vertical.

• Sometimes still ignores prompts

Version 1.0 is is trained on:

• Nude Male

• Inserting [object name] in foreskin.

• Masturbating

• Cumming/ Ejaculating cum

• Inserting finger in foreskin

• Grabbing foreskin

• Fellatio/ Inserting tongue in foreskin

• Fellatio/ sucking foreskin
  
  Going from green to red, the understanding gets worse

Cons of 1.0:

• Drawing styles were overwritten by my illustration style

Example Settings:

• Sampler: Res_2s

• Scheduler: beta57 (download RES4LYF nodes then close fully & restart comfyui)

• Steps - 8-12

• CFG - 1-2

• Shift - 3-5

• Lora Strength - .7-1.0

• Sampler: Euler

• Scheduler: Normal (my examples use "simple" but i think normal gives much better results)

• Steps: 8-12

• Cfg: 1.5 (may go down to 1 for speed)

• Shift: 3.5

• Lora Strength: 1.0

The original version prior to 1.0 is just a prototype! I no longer recommend using it because it does not behave consistently.

• My first version is a prototype. I used ~1200 images but I believe they were poorly captioned, so the resulting images are kind of mid to low adherance to prompts.

• My first version is overtrained. You will need to set the lora weight lower to like .5-.75.

• My first version does not work well with Anime. I decided to test the anime images I tagged for Noob/Illustrious training. It did understand some of the concepts but it adheres to them poorly and has failed to properly capture the anime style.