iteration of Stable Diffusion 1.5, modestly adapted for more refined generation of human figures, hands, and text. The training, while not groundbreaking, was conducted on a reasonable setup of four NVIDIA 3090 GPUs and spanned a modest 16 hours for 8 epochs.
Its capabilities are somewhat specialized, being more adept at creating images of people and textual elements, and less so with animals. This selective improvement makes it a suitable, though not exceptional, tool for tasks requiring detailed human figures or textual accuracy.
The training process incorporated a set of 13,100 unique examples, leading to a dataset of 131,000 images. Each epoch dealt with 31,000 examples, and the model was trained with a batch size of 40. The optimization steps totaled 26,200, with a consistent gradient accumulation, emphasizing gradual and steady learning.
The improvements, while not radical, aim to address common issues in image generation such as blurriness and disproportion. The goal was to achieve clearer, more anatomically coherent results, although the advancements are more evolutionary than revolutionary.
