fg-bg-cadences

# Context bootstrapping with fg/bg independent diffusion cadences

Infer foreground/background using salient object detection (e.g. U2-net), then set diffusion cadence out of sync, so the foreground and background 
alternate taking turns informing how to infill each other. 

Conversely, could potentially be some interesting effects setting diffusion cadences independently, 
so e.g. the foreground objects change more dynamically whereas the background would be forced to be more stable, or vice versa

https:/xuebinqin/U-2-Net