For the main LoRa Lab series

please visit : https://civarchive.com/models/1366934/lora-lab-index

and remember to use the model : ( WAI-NSFW-illustrious-SDXL v12 , as LoRa are train on this model and I didn't try the compatibility ) https://civarchive.com/models/827184/wai-nsfw-illustrious-sdxl

Thanks to WAI's awesome model

Goal : lift your skirt and show me!

• girl's state : a bit nervous, trembling?

• easier to show cum trails on mouth & there... hmmm.. you know...

• better leg wear effects

• should not give half body ( a complete face to thigh area is the goal )

• easier to trigger bdsm effect (neck collar, collar leash, shibari, shibari over XXX)

• SHOULD not logically break the image, and give funny color patch

Notice ( about training params )

NOTE : For comparison made easier, I would normally go with a same set of training params ( unless I'd like to test some/one of them ). Nevertheless, I will upload all the dataset and reg images, so be free to try any different settings by yourself.

So for other params, ref to : https://civarchive.com/models/1362442/lora-lab-1-trying-to-have-better-pantyhose-see-through-maturity

Round 3 : could the training process works on other base model? -> seems ok? but more tests need to be done

R3 : settings

• base model :

  • iLustMix ( thanks to the author's great model! it works really awesome! )

  • https://civarchive.com/models/1110783?modelVersionId=1456068

- 30 pics ( 10 times x 8 epoches )
- image tags : add additional : anime, cartoon (for differentiating from ilust realistic style)
- ... same as R2

R3 : finding

• seems ok on performance,

  • Some original image variety gets lost & girls' appearance becomes more "anime asian girl style"

  • A further distilling process could be done to make it even better?

Round 2 : OK ? ( still testing, but so far, it seems to meet some reqs )

Retro from R1

• (TRY) It might get trained too less, so the effect does not show up

• (TRY) Extra token related to the goal can be set to guide AI ?, e.g. bdsm ?

• (FUTURE) The reg images are merely plain daily girls ( a bit sexy, but ok ), nothing to do to reg the bdsm topic???

• (FUTURE) image quality??? image quantity???

R2 : settings

- 30 pics ( 20 times x 8 epoches, 9600 steps in total )
- trigger token : showme, bdsm
- with reg ( full girls, tagged, but 4 repeats for each image as I don't have time to prepare more for now )
- w: 832, h: 1216
- DIM 24 / 12 (remain the same)

R2 : finding

• This time it shows me something

  • It feels like, training is :

    • a) re-distributing the probability of the effects ( aka. knowledge of the model ) and re-link to the prompt tokens

    • b) learn some pattern, textures, shapes, compositions in LoRa training ( but I doubt how much it can learn within so less the dataset )

    • c) the reg image set will constraint AI , in case it goes nut ( as I didn't see much over-cooked things as a lot of post told me, that over certain amount of steps, AI will over-cooked. So I reckon the reg image set helps?? ) <--- to be confirmed later

  • quality of dataset is important

  • proper trigger token is IMPORTATNT ?? as a) says, we are more trying to re-use existing concepts of the model, instead of creating our own.

    • However, R2 I changed both the steps (longer) and add bdsm token, so this remain uncertain to me

• AGAIN... however, I reckon that R2 is still not ideal as "lift and show me" concept, as I try 1boy, solo, and...

  • AI : dude, do you really want a boy?

  • me : yes, do it

  • AI : I can make lovely porr.. oh, ... corn pic for you, instead...

  • me : no, do boy

  • AI : boys... don't have breasts though

  • me : do it, without female breasts

  • AI : and ... how about ...

  • me : HECK!! DO IT !!!!

  • AI : alright, aright, don't get mad, i'll do it ... let's start... with a boy....

  • me ( waiting )

  • AI : hmm... a boy is good... but maybe I ... hey dude, here's a girl, ... nevertheless, I made the breasts absolutely flat, and add a stick to the proper position. we are happy now?~~ 😌 I'm pretty happy with that

  • me : ...

  • AI ( dead )

• Small tips... machine can hibernate when you want it to train over night

  • what you think is : machine doesn't sleep

  • what machine think is : zzz.z.zz.ZZZ.z.z.z.

Round 1 : FAILED??

R1 : settings

- 30 pics ( 4 times x 8 epochs, 1920 steps in total )
- trigger token : showme
- with reg ( full body girls, tagged )
- w: 832, h: 1216
- DIM 24 / 12

R1 : finding

• WITHOUT this LoRa, the base model EVEN performs better 😂😂😂