Background Remover

Q: Are my images uploaded to a server?

Images stay on your device. AI inference runs entirely in the browser via MediaPipe (WebAssembly). The model + runtime (~0.24MB) are downloaded from the Google MediaPipe CDN on first use, then cached in the browser (offline thereafter). The site operator cannot see your images.

Q: Why is the first run slow?

On first use, the MediaPipe model (~0.24MB) and WebAssembly runtime are downloaded. The model is tiny so even the first run is fast, and subsequent runs start instantly from cache.

Q: What kind of photos work well?

People, pets, and product photos generally work well. Higher color/brightness contrast between subject and background = cleaner results. Fine details like hair / fur are limited by the small model (medium and large models are more accurate but heavier downloads).

Q: What if the result is unsatisfactory?

① Use photos with strong subject/background contrast (white background works great) ② if the subject is small, crop closer first ③ heavy shadows or reflections may leave residue ④ non-person subjects (pets, objects) can be weak with MediaPipe Selfie — a general model is under consideration ⑤ width under 1500px is fastest.

Q: What format is the result?

PNG (to preserve transparency). If transparency isn't needed (e.g. compositing onto a solid background), open it in any graphics tool, place it on the new background, and save as JPG. The Image Compress tool here can also convert PNG → JPG.

Q: Maximum image size?

No hard limit. On mobile, images wider than ~2000px may run out of memory. Desktop Chrome handles up to 4000px reliably. For large images, use the Resize tool to shrink to 1500–2000px width first — it's faster and safer.

Auto-removes the background → transparent PNG. People photos use AI (MediaPipe); white-background logos use color-based removal.

Images never leave your device. People background removal (MediaPipe Selfie Segmenter) runs in the browser via WebAssembly. The model + runtime (~0.24MB) are downloaded once from the Google MediaPipe CDN, then cached. White-background logos/graphics are processed instantly with color-based removal — no model.

⚠️ Note — Photo (people) mode downloads the MediaPipe Selfie Segmenter (~0.24MB) once and caches it. It's most accurate when a person is the main subject; fine details like hair have limits. For white-background logos/graphics, logo mode (color-based) is cleaner.

Drag an image here, or click to choose

JPG · PNG · WebP — width 1500–2000px recommended

Choose how to process

⚡ Auto (Recommended) Analyzes the uploaded image — white background + text → logo mode, otherwise photo mode, automatically. Best for most cases. 📷 Photo / People (AI) Separates the background of people/selfie photos (MediaPipe Selfie Segmenter). Most accurate when a person is the main subject. Inference 1–3s. 🎨 Logo / Graphic (Chroma Key) White background + text/logos. Pixel-level color removal preserves gray text too. Instant (<1s).

AI engine — MediaPipe Selfie Segmenter (Apache-2.0)

A single lightweight model (~0.24MB) separates the person from the background. No quality selection needed — it runs fast and the model is fetched only once.

Photos that work well vs. don't

Good results	Weaker results
Person or product on a white / clean background	Subject and background are similar in color
Subject fills 60%+ of the frame	Subject is far away / small
Sharp outlines (pets, objects)	Long hair / fur details (small-model limits)
Even lighting	Strong shadows / reflections / backlight
1000–2000px wide	Under 500px (low res) or 4000+ (memory)

How background removal runs inside your browser — on-device AI inference

Most online cutout sites upload your photo to a server, process it on their GPU, and send the result back. This tool works differently. The AI model (MediaPipe Selfie Segmenter) that finds the person/background runs directly inside your browser via WebAssembly. Your photo never leaves your device — privacy here is a structural property, not just a promise.

1. First load: download the model once

The first time you use photo mode, the model file (~0.24MB) and the WebAssembly runtime are downloaded once from the Google MediaPipe CDN. This downloads the program (model weights), not your photo. The assets are stored in the browser's IndexedDB / HTTP cache, so from the second run onward in the same browser it starts instantly with no extra download.

2. Inference: the model decides foreground vs background per pixel

The model takes your photo and computes, for every single pixel, the probability that it belongs to the foreground (subject), producing an alpha mask. This runs on your device through WebAssembly. So speed depends on your own hardware and the image resolution, not on a server.

Easy images vs hard images — and why

What separates an easy photo from a hard one comes down to how clear the boundary is.

Works well — people, pets, and products with crisp outlines and strong color/brightness contrast against the background. The model confidently pushes the foreground probability toward 0 or 1, so edges come out clean.
Hard cases — (1) hair and fur, where foreground and background mix within a single pixel; (2) semi-transparent areas like glass, water, or smoke; (3) subjects whose color is similar to the background, making the boundary ambiguous; and (4) busy, cluttered backgrounds. Here the foreground probability hovers near 0.5 and the mask gets rough.

The MediaPipe Selfie model is tuned for people, so it's cleanest when a person is the main subject. For non-person subjects (pets, objects) or finer matting, a general model (a future option) may fit better. For simple images like a plain white background, logo mode (color-based removal) is faster and cleaner.

Related tools

Which engine removes the background?

Photo / people mode uses the MediaPipe Selfie Segmenter by Google (Apache-2.0). Both the code and the model weights are Apache-2.0, so it's free to use on commercial / ad-supported sites. Logo / graphic mode uses canvas color-based (chroma key) removal with no model.

Why download model files from the Google CDN?

Committing the model + WebAssembly runtime directly into the repo would make git clone / deploy heavy. Same pattern as OCR (Tesseract): the model (~0.24MB) and runtime are fetched from the Google MediaPipe CDN on first use and cached afterwards. Your images are never sent anywhere.

WebAssembly requirements

AI inference uses WebAssembly. Modern Chrome / Firefox / Edge / Safari (15+) all support it, and it runs single-thread without SharedArrayBuffer.

Tips for large photos

Inference time / memory scales with input size. 1500–2000px wide is the sweet spot. Use the resize tool to shrink 4000px+ photos first.

Does it work on pets / objects? / Switching engines

MediaPipe Selfie is people-focused, so it's most accurate when a person is the main subject. For pets, objects or finer matting, a general model is under consideration. The engine is isolated behind a code adapter, so swapping it is easy.

Frequently Asked Questions

Are my images uploaded to a server?

Images stay on your device. AI inference runs entirely in the browser via MediaPipe (WebAssembly). The model + runtime (~0.24MB) are downloaded from the Google MediaPipe CDN on first use, then cached in the browser (offline thereafter). The site operator cannot see your images.

Why is the first run slow?

On first use, the MediaPipe model (~0.24MB) + WebAssembly runtime download. The model is tiny so even the first run is fast, and subsequent runs start instantly from cache.

What kind of photos work well?

People, pets, and product photos generally work well. Higher color/brightness contrast between subject and background = cleaner results. Fine details like hair / fur are limited by the small model (medium and large models are more accurate but heavier downloads).

What if the result is unsatisfactory?

① Use photos with strong subject/background contrast (white background works great) ② if the subject is small, crop closer first ③ heavy shadows or reflections may leave residue ④ non-person subjects (pets, objects) can be weak with MediaPipe Selfie — a general model is under consideration ⑤ width under 1500px is fastest.

What format is the result?

PNG, to preserve transparency. If transparency isn't needed (e.g. compositing onto a solid background), open it in any graphics tool, place it on the new background, and save as JPG. The Image Compress tool here can also convert PNG → JPG.

Maximum image size?

No hard limit. On mobile, images wider than ~2000px may run out of memory. Desktop Chrome handles up to 4000px reliably. For large images, use the Resize tool to shrink to 1500–2000px width first — it's faster and safer.

References · sources

Last reviewed: 2026-06-24 / engine = MediaPipe Selfie Segmenter (Apache-2.0). Legacy option = @imgly/background-removal (AGPL/Commercial, switchable in code)

MediaPipe Image Segmenter (Selfie) — Google MediaPipe (Apache-2.0 · code + model weights). People background separation.
Model + runtime hosting — cdn.jsdelivr.net/npm/@mediapipe/tasks-vision + storage.googleapis.com/mediapipe-models/…/selfie_segmenter.tflite (~0.24MB).
Legacy engine (optional) — @imgly/background-removal (AGPL-3.0 / Commercial). One-line switch in code when a general model is needed.
WebAssembly — webassembly.org (W3C standard)

⚠️ AI results vary by photo and aren't guaranteed. For commercial or official material, review and touch up the result by hand. Use only on photos you own or have permission to process.