Background removal guide
How to remove a background from a photo, and get clean edges
Removing a background used to mean an hour with the pen tool. Today an AI model does the first 95% in seconds, and the difference between a cutout that looks pasted on and one that looks real is almost entirely in the last 5%: the hair, the fur, the soft edges. This guide explains what happens under the hood, where free tools cut corners, and how to get professional results without uploading anything.
What "removing a background" actually produces
A cutout is your photo plus an alpha channel: a second, grey-scale image that says, pixel by pixel, how opaque the subject is. White means fully kept, black means fully removed, and every shade in between means partly see-through. A PNG or WebP file stores that channel next to the colour, which is why those formats can have a transparent background and JPEG cannot.
Most of an alpha channel is easy: the middle of a person is clearly the person, the far wall is clearly the wall. The hard part is the thin band where they meet. A strand of hair is often narrower than a pixel, so the pixel holds a blend of hair and background. Getting that band right is what separates a clean cutout from one with a jagged or haloed outline.
How the AI finds the subject
Clipout uses BiRefNet, an open segmentation model released under the MIT licence. It looks at the whole image at 1024 × 1024 pixels (512 × 512 on the CPU engine), decides what the main subject is, and predicts an alpha value for every pixel. It was trained on high-resolution datasets built for exactly this task (dichotomous image segmentation), so it handles people, animals, products, plants and objects without being told what to look for.
The model runs with ONNX Runtime Web inside your browser. When your browser supports WebGPU (current Chrome and Edge on a computer), the work happens on your graphics card and takes a few seconds. Otherwise a 512-pixel version of the model runs on the CPU through WebAssembly, because the full-size one does not fit a browser's CPU memory; the full-resolution edge pass makes up most of the difference. iPhones and iPads always use the CPU engine, because WebGPU there tends to crash the tab on large models.
Why cutouts from free tools look pasted on
The model sees a 1024-pixel square. Your photo is probably 12 megapixels (4032 × 3024) or more. Most tools simply stretch the small mask up to the photo's size, which produces two visible faults:
- Soft, blurry edges. A mask stretched three or four times cannot follow a strand of hair; it smears it into a fuzzy outline.
- A halo of the old background. The see-through pixels at the edge still carry the background's colour. Put a subject shot against a green wall onto white, and a thin green fringe appears around the hair.
Many online services also limit the free download to a small preview (remove.bg's free tier is 0.25 megapixels, roughly 625 × 400), which hides both problems until you pay for the full-size file.
The two steps Clipout adds
Edge refinement at full size. After the model, Clipout runs a colour guided filter (He, Sun and Tang) that fits the mask to the real edges of your full-resolution photo. Confident areas keep the model's answer; only the uncertain band along the outline is redrawn. Hair and fur keep their shape instead of turning to fog.
Colour decontamination. Clipout then estimates the true colour of the subject in every see-through pixel, using the multilevel foreground estimation method of Germer and colleagues (the one in the pymatting library), and removes the old background's tint from the fringe. The correction is computed at working size and added to your full-size photo, so fine texture is untouched.
Measured results
We tested on 40 images from AIM-500, a public matting benchmark with hand-made alpha mattes: 16 solid subjects, 14 with see-through or very fine detail, and 10 without a single salient subject. Lower is better for SAD, MSE and gradient error; higher is better for IoU. RMBG-1.4 is included because many free in-browser tools use it.
| Pipeline | SAD (×10³) | MSE (×10⁻³) | Gradient error (×10³) | IoU |
|---|---|---|---|---|
| RMBG-1.4 (1024), plain mask | 41.9 | 9.16 | 89.1 | 0.857 |
| BiRefNet lite 512, plain mask | 48.5 | 11.58 | 109.4 | 0.845 |
| BiRefNet lite 512 + Clipout (CPU engine) | 42.1 | 9.06 | 81.8 | 0.869 |
| BiRefNet lite 1024, plain mask | 40.0 | 8.03 | 88.1 | 0.879 |
| BiRefNet lite 1024 + Clipout (GPU engine) | 34.6 | 6.24 | 61.1 | 0.900 |
On the GPU engine the refinement lowered MSE by 22% and gradient error (the metric that tracks how natural an edge looks) by 31% against the same model's plain mask, and the result has 32% lower MSE than plain RMBG-1.4. On the CPU engine, which runs a 512-pixel version of the model to fit a browser's memory, the refinement brings the result level with RMBG-1.4 at 1024 pixels. Measured on 16 September 2026 with the same code the site runs; the benchmark images were scaled to at most 1600 pixels, so on larger photos the gap in edge sharpness is wider.
Step by step: removing a background with Clipout
- Open [the tool](/#app) and drop an image, paste one with Ctrl+V (⌘V on a Mac), or tap Choose images. You can add up to 50 at once.
- Wait a few seconds. The first visit downloads the model once (115 MB for the GPU engine, 94 MB for the CPU engine); after that it is cached and starts immediately.
- Check the result with Compare, which slides between the original and the cutout.
- Touch up if needed: Erase removes leftovers, Restore brings back parts the AI dropped. Undo and Redo work as usual.
- Pick a background: transparent, white, black, any colour, a blurred copy of the original, or a photo of your own.
- Export as PNG or WebP (transparent) or JPEG, at full size or cropped to the subject, or copy the image to the clipboard.
Tips for the best cutouts
- Contrast helps. A subject that stands out from the background in colour or brightness gives the cleanest edges.
- Use the largest original you have. The edge refinement works at full resolution, so more pixels mean finer hair.
- One clear subject works best. Crowds or scenes without an obvious subject may need the brush.
- For marketplaces, export JPEG on white. Amazon asks for a pure white background (RGB 255, 255, 255) on main product images.
- For design work, keep PNG. It stores the full alpha channel without compression artefacts.
Private by design
Upload-based removers send your photo to a server, process it there and send the result back; the better ones promise to delete it within an hour. Clipout has no server step at all: the model and every processing step run inside your browser tab, and the site has no upload endpoint. You can confirm it in your browser's network panel, or load the page once and switch your connection off: the tool keeps working.
That is also why it is free without a catch. There is no GPU bill to pass on, so there are no credits, daily limits, accounts or watermarks.
Ready to try it? Remove a background now, or read the frequently asked questions.