intro.

• 11-02 -- lab 11 is the series of LoRA, it is meant to be a "UTILITY" LoRA for me to fix and enhance dataset for training lab 10. however, I found it's ok to use alone, so share as well.

• 11-01 -- is a quick test on civit's training LoRA feature

how-to use

lab 11-02 ( better, can be used alone )

• model -- use janku v3

• weight 0.5 ~ 1, step > 20 ( won't produce too much buggy result )

• highres fix and adetailer is not a must

• positive prompt & trigger token could be :

  • lrfixme,

    • hmmmmmm, it's meant to be a utitliy LoRA to fix images for lab 10, so I use a token like this

  • sheer pantyhose,

  • sheer golden pantyhose,

  • sheer xxx pantyhose,

  • sheer bodystocking,

  • high detailed skin,

    • (NOT TESTED) it seems can improve a bit details?

• negative prompt could be :

  • pantyhose band,

    • personal taste, just for recomemdation

• dataset is attached as well, you can download and see :

  • how training dataset quality affects LoRAs, e.g. between 11-01 and 11-02

lab 11-01 ( bad, not recommended )

• weight 0.5 ~ 0.7 tested, step > 30 (step must be a bit higher, or I can see in some generations that ????? is rendered improperly)

• highres fix ( or high res img2img ) is recommended

  • adetailer might not be needed

note.

lab 11-01 : training in Civit

• hmmmm... I have no further comment as I haven't got a chance to test the result LoRA yet.

• Pro. However, the whole workflow is seamless ( from dataset preparing, all the way to publish a trained LoRA )

• Con. I didn't find a way to upload reg image set ?

• Con. There is no way to see the Loss over steps ( e.g. tensor board )

• Con. the training size is fix SQUARE, as said so, you can't specify something like 832 x 1216