Two completely different technologies get called "background removal", and knowing which one you're using explains both the results and the failures.
Colour Keying (the Classic Method)
The tool picks a target colour and erases every pixel within a tolerance of it. It's fast, runs anywhere, and is entirely predictable.
- Works when: the background is uniform and clearly different from the subject — logos on white, products in a lightbox, scanned signatures, green screens.
- Fails when: the subject contains the background colour, the background is a gradient or a real scene, or the difference is subtle.
- Understands nothing. It has no concept of "person" or "product" — only colour distance.
Semantic Segmentation (the AI Method)
A neural network trained on huge numbers of labelled images predicts, for each pixel, whether it belongs to the foreground object. It's learned what people, products, animals and cars look like, so it can separate a person from a cluttered street scene where no colour rule could.
- Works when: there's a recognizable subject, even against a complex background.
- Fails when: the subject type is unusual (the model has never seen it), the subject blends into the background tonally, or multiple overlapping objects confuse it about which is "foreground".
- Understands objects, not colours — which is why it succeeds where keying fails, and vice versa.
Why Hair and Glass Are Hard for Both
A single pixel along a hair strand contains partly hair and partly background. The correct output is partial transparency — say 40% opaque — not a yes/no decision.
Producing that is a separate problem called alpha matting, and it's genuinely difficult. Good tools run a segmentation pass to find the object, then a matting pass to estimate fractional opacity along the boundary. Cheaper tools skip the second step, which is why cutouts often show:
- Hair rendered as a solid, wig-like outline.
- A halo of background-coloured pixels.
- Glass and smoke rendered as opaque or erased entirely.
Motion blur, fine mesh, feathers and fur all have the same underlying problem.
Choosing the Right Tool
- Uniform background (product on white, logo, signature) → colour keying is faster, fully predictable, and often *more* precise than a model.
- Complex background with a person or common object → segmentation.
- Hair, fur, glass, or a composite that must look real → a tool with explicit matting, and expect manual refinement.
- Sensitive images (ID documents, personal photos) → check whether processing happens on your device or on a server; segmentation models are large, so many services upload your image, while colour-based tools can run entirely locally.
Getting Better Results From Either
The biggest wins happen before processing:
- Increase contrast between subject and background — different colour, different brightness, more distance.
- Light the background evenly.
- Avoid thin, wispy edges where possible.
- Work at full resolution, then downscale, since edge quality degrades when you scale up an already-extracted cutout.
- Always inspect on a dark background before publishing — that's where fringing hides.