Intro
This time :
Trying to train the LoRa to make a girl looks like old man with beard.
Beard is a strong attribute of male, so it's hard to make a 1girl, 1boy image of that girl has beard and boy has not via prompting. So this LoRa aims to try if it is possible to make this easier.
Pure for learning purpose.. don't take it too serious.
NOTE : if it doesn't work, make the lora's weight higher (i.e. 1.2) and try.
NOTE 2 : dataset (with reg imageset) are included, pls see download file list
Later I will upload all 6 attempts with dataset and reg images. Perhaps with some of what I have learned.
KNOWN ISSUE :
Hands and arms will sometimes mess up
Findings
lab_6_6 : looks promising, however, there are still problems
problems
color :
tends to be brown toned?
color is not rich enough
image composition :
hands & feet is not well drawn
mouth :
malegirl tends to "closed mouth"
style :
feels like lora's got a FIXED style?
plan
maybe correct the dataset ( the solo version's got "lower" quality )
8 images are not enough for such task
Training params
For base params, pls check Lab Index (normall they are the same)
Following, it's the difference that I tried for each round
- round 6
- train on:
- dataset
- common token: lr-malegirl,
- with partial beard related token (to control the beard color... and help the model to understand beard stuff)
- 4p (20 times), 1girl solo,
- 4p (20 times), 1girl 1boy,
- reg - normal 1girl, 2 times each pic
- size: 1024 x 1152
- DIM: 16, 8
- keep N token: 1
- round 5
- train on:
- dataset
- common token: lr-malegirl,
- no beard related tokens,
- 4p (20 times), 1girl solo,
- 4p (20 times), 1girl 1boy,
- reg - normal 1girl, 2 times each pic
- size: 1024 x 1152
- DIM: 16, 8
- keep N token: 1
- round 4
- train on:
- dataset
- common token: lr-malegirl,
- no beard related tokens,
- 4p (10 times), 1girl solo,
- 4p (10 times), 1girl 1boy,
- reg - normal 1girl, 1 time each pic
- size: 1024 x 1152
- DIM: 16, 8
- keep N token: 1
- round 3
- train on:
- dataset
- common token: lr-malegirl,
- no beard related tokens,
- 4p (15 times), 1girl solo,
- reg - normal 1girl, 1 time each pic
- size: 832 x 1216
- DIM: 16, 8
- keep N token: 1
- round 2
- train on:
- dataset
- common token: oldmangirl,
- with beard tokens,
- 4p (24 times), 1girl solo,
- reg - same as dataset, 4p (15 times)
- size: 832 x 1216
- DIM: 16, 8
- keep N token: 0
- round 1
- train on :
- dataset
- common token: old man malegirl,
- with beard tokens,
- 4p (15 times), 1girl solo,
- reg
- same as dataset, 4p (15 times)
- size: 832 x 1216
- DIM: 16, 8
- keep N token: 0