When a background removal tool produces a cutout with jagged edges or missing hair, the problem is almost always the masking approach. Most tools use a binary mask. The good ones use alpha matting.
A segmentation model outputs a mask that says, for each pixel, whether it belongs to the foreground or the background. This is a binary decision. A pixel is either in or out.
The result is a hard edge. On a box or a car or a piece of furniture, a hard edge is correct because those objects have hard edges. On hair, fur, or a leaf, it is wrong. The edge of hair is not a line, it is a gradient from opaque to transparent over several pixels.
Binary masks on hair produce the classic cutout artifact: a jagged outline where strands should be, or a halo of background colour left around the head.
Alpha matting estimates a partial alpha value for each pixel in the transition zone between foreground and background. Instead of 0 or 255, a pixel can be 127, meaning it is half transparent.
The result is that fine strands of hair are preserved as semi-transparent pixels. When composited over a new background, they blend naturally. The edge looks like a photograph rather than a cutout.
The tradeoff is compute. Matting is a second pass on top of segmentation. It analyzes the boundary region pixel by pixel and solves an optimization problem to estimate alpha. This takes time.
Matting matters when the subject has soft edges. Hair, fur, feathers, foliage, smoke, glass, lace, and any object with a fuzzy boundary benefits from matting.
Matting does not matter for hard-edged subjects. A book, a phone, a box, a car. These have clean boundaries. A binary mask is correct and matting would add processing time for no visible improvement.
The practical implication: a tool that uses matting on everything is wasting time. A tool that never uses matting produces bad results on hair. The ideal is a tool that detects when matting is needed and applies it selectively.
Remova runs matting when the image is classified as containing a human or animal subject, or when the router detects complex edges. On product photos and simple graphics, it uses a faster binary mask.
The X-Hair-Refined header on the API response tells you whether matting ran. If it is true, the tool applied an additional refinement pass. If it is false, the image did not need it.
Not through the default interface. The router decides. If you are using the API, there is no parameter to force matting because it would double the processing time on images that do not need it.
Matting is a second inference pass. The first pass creates a rough mask. The second pass analyzes the transition zone and estimates alpha for each pixel. Two passes take roughly twice the time of one.
In terms of compute, yes. It uses more CPU or GPU time. For a free service, this is a meaningful cost that has to be balanced against quality.
Yes. Models like MatAnyone and the newer BiRefNet variants can produce a soft alpha directly without a separate matting pass. They are newer and less widely deployed.