Please consider starring ★ our GitHub Repo if you find this demo useful! 😊
Current Task: Depth
Runs the FLUX transformer in float16 (finer mantissa than bfloat16,
removes the bf16 depth staircase) and applies a per-channel
scipy.ndimage.median_filter (ksize=15) to the raw prediction
to strip the FLUX-VAE 16-px patch grid.