v1.0_____________________________________________________

I think I am happy enough to call this 1.0. Took longer that I anticipated but i learned a lot. Masked training, smooth loss, tensorboards... the works. Yikes.

About this version

Added trigger tokens back in -

B1GN4TZ - The main trigger, should be used in every prompt.

BNH34VY - Nudges towards a heavier, low hanging set

BN4P4RT - Nudges towards a wider-set pair.

Recommended LoRA Strength: 0.6 - 0.9

I may revisit at a later date if issues arise. On to the next one. I have some ideas.

v0.8_____________________________________________________

A little more consistent generation this version, with much better quality. Still needs a little more coaxing than I would like, but will try to improve for next version. Eliminated the unique token from previous version, and switched to using 'big natural breasts' for prompting.

Back to the drawing board for v0.9. Improving/Expanding the dataset and better captioning are the priorities. May re-visit unique trigger token if necessary.

v0.7_____________________________________________________

The LoRA is intended to help produce safe for work images of women with big natural breasts. I tried to use a variety of different clothing, ethnicities, and body shapes to achieve something acceptable.

I intend to improve on it, but wanted to share in order to get some feedback/ direction.

It seems to work best at a weight of 0.8 to 1.10, depending on which SDXL model is being used.