Bing 3D for Flux is a primarily furry-oriented style LoRA trained on FLUX.1-dev using synthetic data sourced from publicly posted Bing Image Generator outputs, and a lesser amount of high quality Stable Diffusion outputs. It aims to improve the visual aesthetic of CGI style imagery of fluffy creatures with the look of the physically principled but stylized and artist-driven rendering approach of major animation studios.
The dataset is dual captioned, with one epoch consisting of a repeat with shuffled tags (e6 + booru) and a repeat using natural language captions from JoyCaption Pre-Alpha (2024-07-30a) by fancyfeast.
Please enjoy.
Description
This is a 2 epoch run (3704 steps), continued for 1 epoch (+1852), continued again for 1 epoch (+1852) for an effective running time of 7408 steps. This training run used the experimental Compasstic full BF16 optimizer by Lodestone Rock.