My humble attempt to capture a fraction of Hiroo Isono's majesty. I'm still not happy with this one but hoping folks can find optimal settings.

My best results are with:
- Distilled model
- CFG 3 (yes, with distilled)
- res3s sampler w/ beta scheduler
- 4 steps

This LoKR benefits from long prompts that specify the exact layout - see the gallery for examples.

It will work with ComfyUI and any other generators that support LoKRs.

Trained using perceptual depth anchors with ai-toolkit-perceptual.