Intro to Beholder Vision

30/11/2024

I'm happy to announce version 1o (and as of the time of writing, 1.1o) of Beholder Vision! With an even more thoroughly curated dataset and reworked training parameters (the specifics of which will be uploaded if there's interest), I can safely say that it's a substantial improvement to the 0.9beta.

Following in the theme of experimentation, 1o comes as a GLoRA. Its successor, 1.1o, returns to Beholder Vision’s LoCon roots, featuring a marked improvement over its parent at a fraction of the file size.

• Now natively supports other SDXL models (such as NoobAI-XL Epsilon-pred 1.0-Version, used for the images above).

28/11/2024

My first foray into LoCon and LyCORIS, trained on a dataset of 126 well-curated images of aesthetic anime art.

Beholder Vision is designed with the idea of "eye-candy" in mind, and can be very flexible to suit your needs.

Recommendations

30/11/2024

Set the strength between 1.0-1.2 (having done a lot of testing, I personally recommend sticking with just 1.0). DPM++ 2M SDE + Karras. 25-35 steps and CFG 5-7.