Trained at native 1024px without downscaling on an ethnically diverse watermark-free dataset, with equal buckets for 1024x1024, 1024x768, and 768x1024.
This Lora is effective at strengths as low as 0.1 - 0.2 on some checkpoints, and can scale upwards all the way to 1.0 from there, so feel free to experiment to find the right strength for any particular checkpoint you might be using.
All showcase images were generated directly as the resolution they're displayed at, without upscalung.