Hair is the hardest part of background removal. Every other edge in an image is either hard or soft. Hair is both at the same time, which is why it separates good tools from bad ones.
A strand of hair is thinner than a pixel at normal viewing sizes. When you look at a photo of a person, the hair around their head is not a solid shape. It is a collection of individual strands that overlap and let light through.
A segmentation model that produces a binary mask has to decide, for each pixel, whether that pixel is hair or background. But a single pixel at the edge of hair might contain a strand of hair, some background, and the shadow of the hair on the background. It is a mixture.
The correct representation for this pixel is a partial alpha value that reflects the mixture. The model needs to estimate that value rather than making a binary decision.
The best hair preservation is done with alpha matting. After the initial segmentation pass creates a rough mask, the matting pass analyzes the boundary region and estimates alpha for each pixel. On hair, this produces a soft edge where individual strands are partially transparent.
The quality of the result depends on several factors. The resolution of the input image matters because hair detail is lost when the image is downscaled. The contrast between hair and background matters because the matting algorithm needs to see the difference. The colour of the hair and the background matter because similar colours confuse the algorithm.
A portrait with dark hair against a light background is the easiest case. Light hair against a light background is the hardest. Dark hair against dark hair is somewhere in between.
If the original photo is low resolution, the hair detail is not there to preserve. The model cannot invent strands that were not captured.
If the hair colour is close to the background colour, the algorithm cannot distinguish them. A blonde person against a cream wall is a difficult case.
If the hair is backlit, the strands become bright and the background becomes dark. This is unusual for a portrait but it happens in outdoor photos and it confuses the model.
Use the manual editor. Every background removal tool has some way to refine the result by hand. A brush tool that adds or removes alpha in specific regions can fix the worst problems.
Shoot against a contrasting background. If you are in control of the photo, a blue or green screen behind the subject makes hair preservation much easier. The algorithm has more signal to work with.
Increase the resolution of the source image. Hair preservation is directly limited by input resolution. A 4000 pixel image preserves more detail than a 1000 pixel image.
Use a model that specifically handles hair. BiRefNet and ISNet are the best choices. U2-Net is fast but loses strands. Silueta is for video and should not be used for still portraits.
A binary mask produces a hard edge at the boundary of the hair. This looks like a helmet because the soft transition of individual strands is missing. You need a tool that uses alpha matting.
Partially. You can add back some transparency around the edges with a brush, but you cannot recover strands that were never detected. The tool is a starting point, not a final result for difficult hair.
BiRefNet is the best for hair preservation as of 2026. ISNet is second. Both use matting. U2-Net does not and produces worse results on hair.
Yes, significantly. High contrast between hair and background makes preservation easier. Low contrast makes it harder. If you control the shoot, use a contrasting background.