Refining edges

A soft matte leaves the old background mixed into every semi-transparent pixel, so compositing the original pixels onto a new background shows a halo of the old one — most visible on hair, fur and motion blur.

refine_foreground solves for the foreground color those pixels would have had, using the blur-fusion estimator of Approximate Fast Foreground Colour Estimation (two passes, coarse then fine).

In cutout

processor.cutout(image, alpha, refine=True).save("output.png")

That is the whole feature: the image's colors are replaced with the estimate before the alpha channel is attached.

Standalone

It is pure torch, so it runs wherever its inputs live — keep the tensors on the GPU and the refinement stays there too:

alpha = processor.post_process_alpha_matting(
    outputs, target_sizes=[(image.height, image.width)]
)[0]
foreground = processor.refine_foreground(pixel_tensor.cuda(), alpha.cuda())

The return type matches image: a PIL image in, a PIL image out; a batched tensor in, a batched tensor out, with the input dtype and device preserved. The estimate itself is always computed in at least float32 — the divisions inside overflow in half precision, which shows up as black patches.

Choosing r

r is the coarse blur radius, and the default of 90 is tuned for the resolutions these models run at. Larger reaches further for a color to borrow (better on wide, wispy edges) at a roughly linear cost; the second, fine pass always uses a fixed radius of 6.

When it matters