note - if you were one of the first 20 people to download the workflow, try re-downloading the T2I workflow, I messed up by using a condition zero out node. Don't use those. - use a regular negative clip encode node.
note - the Loras attached to this checkpoint ONLY work with checkpoint v1.5 or higher. Note that the current Loras are just prototypes. My main focus is on the checkpoint and I will come back to the lora later. the retracted prototype is not very detailed at all, it is just an experiment
This model is best used with Turbo/Distilled Lora, otherwise you are unlikely to get good results https://civarchive.com/models/2324315/klein-4b9b-base-to-turbo-lora .
This creates realistic images in 4 to 8 steps!!
This can do text to image & image to image (up to 3 images) since it is just a finetune of Klein.
This model is trained slightly differently from my lora https://civarchive.com/models/2343427/klein-foreskin
It is slightly better at image to image outputs than the lora, though there are some things the lora does better such as taking a drawing and making it into a photo.
It retains more knowledge and is more realistic than just using the lora.
Version 1.9 is just trained longer than 1.5
Version 1.9 is just trained longer than 1.5. This makes it follow the prompt better with some things, but at the expense of realism. It started to overbake after 1.9 due to unbalanced dataset. Many things are still underbaked. I will now work on a version 2.0 with the goals of rebalancing and improving what is already here.
Version 1.5 is a remake of Version 1.0
During training of Version 1.0, I had some dataset errors such as missing control images, mismatched control sizes, missing captions and more. This resulted in several I2I issues where the model was not properly following directions.
I wrote a script to help validate my dataset images all had captions and no controls were missing. I trained from the ground up for Version 1.5. I have tweaked a couple of things which result in following prompts more closely, however Version 1.0 may have slightly different outputs which makes it worth downloading too.
I have paused training to work on a circumcised/retracted foreskin penis lora (that pairs with this checkpoint),
I will also test the capabilities of 1.5 to determine areas of weakness and then continue training from here to push it further.
What can this model do?
This is designed to give easy to use control with natural language. It is trained on the following conceps. I am going to see how much ability I have to continue finetuning to improve the results:
works very well
1. nude male, pantless male, shirtless male, penis exposed through fly, pants pulled down,
2. penis, erect penis (needs some more training on erect)
3. foreskin, (i2i) add foreskin
4. masturbating (borderline needs more to training to finetune this more).
5. grabbing foreskin
6. inserting things in foreskin like fingers, objects
7. pubic hair (i2i) add/remove pubic hair/body hair
8. Cock rings
9. Cum/ Cumming/ Ejaculating Cum, Cum is on face
Version 1.5 is better at:
1. converting images to styles
Most of the way there and will often work
1. masturbating
2. erect penis, though I want to work on this more
partial training/ needs more work
1. (i2i) male penis erect/ make penis flaccid
2. fellatio, tongue in foreskin, sucking foreskin
3. (i2i) remove foreskin
4. long foreskin
5. glans
6. fleshlight
7. Camera views like point of view
training didn't stick
1. short_foreskin
2. make penis smaller
3. increase foreskin length
Items that were forgotten from base model
1. (i2i) make into a photo/ make into anime
2. text is sometimes messed up
no training yet
9. autofellatio
10. anal sex
11. circumcised
12. retracted
Plan
I will plan to make a lora for circumcised penises.
I will plan to expand my dataset and continue training 1.5 into 2.0 to fin