Zephyr is a stylistic LoRA trained on a highly opinionated, handpicked dataset, aiming to improve anime quality by muting the colors and exposure; increasing the detail; add a slight artistic feel to the images.
The versions indicate the iteration of the dataset and the order the models were made in, and although correlated, not necessarily indicative of quality. V3 is better than V4 for example. The order in which the models are listed is a better approximation of the quality.
Zephyr doesn't require trigger words, as it was trained without captions, with frozen text encoder weights, meaning that it's applying the aesthetic bias only through the UNet. Considering this, Zephyr shouldn't conflict with other LoRAs.
Zephyr is currently in rapid development, so you can expect many updates. Open to criticism and suggestions.
Description
What's new:
Cleaned up the dataset, removed duplicates and worse images, now it's down to 1,675 images from 2,219
Reduced LoRA rank from 64 to 32
What's next:
Still having trouble getting meaningful updates beyond the 100th step, after that the model just keeps drifting around the same loss/quality, I'll increase the gradient accumulation size