Creates realistic images focused on asian girls spreading their legs and/or using thier hands to spread their pussy.
Trained against Acorn Is Spinning for FLUX
It was trained against the DevFP8 V1.1 but i'm generating images on V1.1 Hyper 8-Step version without any problems.
Version 1
Size: W768 x H1024
Sampler: Euler (I use Euler with Beta and 30 steps)
Trigger word: as1anpu55y
- All the images were trained with flux style captions, plus the trigger word mentioned above.
Trained with 260 images, 15 repeats and 10 Epochs - for a total of 39000 steps.
Version 2:
Size: W768 x H1024
Sampler: Euler (I use Euler with Beta and 30 steps)
Trigger word: as1anpu55y
- All the images were trained with flux style captions, plus the trigger word mentioned above.
Trained with 315 images, 12 repeats and 10 Epochs - for a total of 37800 steps.
Version 3:
Size: W768 x H1024
Sampler: Euler (I use Euler with Beta and 30 steps)
Trigger word: as1anpu55y
- All the images were trained with flux style captions, plus the trigger word mentioned above.
Trained with 315 images, 16 repeats and 10 Epochs - for a total of 50400 steps.
Notes for each version:
Version 1 has a smaller dataset and is best used for the pose of girls laying on their back, with their legs open, usually spreading their labia/pussy with their hands. It does a better job of this than version 2 or 3, and provides the best quality and detail of this specific pose. It struggles to produce correct or detailed images in other poses. The captions for each image were AI generated to about 300-400 characters (without any human changes).
Version 2 is, in my opinion, the failed attempt of version 3. I changed the dataset from version 1 to reduce to chance of professional looking images, and reduce the chance very pale skin tones which might be the popular definition of beautiful in Asian girls, but doesn't represent reality and realistic photos. I also changed the dataset to allow for more poses, such as standing and squatting with legs spread apart or 1 leg raised. Version 2s failure is in the number of repeats... While it's not a total failure and it still produces some good images, 37800 steps on 315 images isn't enough (now I know).
Version 3 is just version 2's dataset, but with the right amount of steps for training. It can handle several standing, squatting and laying down poses to a good standard of quality. The captions were AI generated to about 1100 characters per image, and then I went through each one to customise and correct them. In comparison to version 1, due to the additonal poses/stances, it does lose a bit of quality and detail and has a higher chance of giving deformed images.
- words used in version 2 and 3's captions to help steer the image generation are: Japanese, Chinese, Southeast Asian, Asian, candid, unembellished, personal, private moment, unfiltered, amateur, professional, spread, apart, labia, pussy, vagina.
Summary:
For the best quality of girls laying down, using Version 1
For more versatility in poses/stances (but with a minor drop in quality/detail), and better a much better captioned dataset, use Version 3
Maybe I could go a little further with this model and tweak it some more, but I'm pretty happy with the result, so version 3 is probably the last version... Unless I decide to train it on a new checkpoint in future.