Background Remover
AI background removal that runs entirely offline.
Remove image backgrounds with a small AI segmentation model that runs entirely in your browser. The model downloads once (around 4 MB), then works offline for every photo after that — no uploads, no API keys, no usage limits.
About the Background Remover
AI background removal used to mean uploading photos to a paid online service and waiting a few seconds for a cloud GPU to do the work. That model made sense when running neural networks in a browser was a curiosity. Today, segmentation models have gotten small enough and WebAssembly has gotten fast enough that the entire workflow can live inside the browser tab you have open. The result is a tool that's free, private, and just as accurate as the paid services for the vast majority of photos.
Under the hood, the background remover uses a U-Net-style segmentation model trained on millions of foreground/background examples. Given an input image, the model produces a per-pixel probability map — for each pixel, how likely it is to belong to the subject — and the tool uses that map to build an alpha channel. Pixels with high foreground probability become opaque, low-probability pixels become transparent, and the soft transition zone (typically along hair, fur, edges of clothing) gets a smoothed alpha that produces clean composites.
The model is around 4 MB compressed. On first use, your browser downloads it from a CDN and caches it. After that, every removal happens entirely on your device using WebAssembly — no network requests, no usage caps, no servers involved. On a modern laptop a removal typically takes 2–6 seconds. On a recent smartphone it's closer to 5–15 seconds. The tool runs the inference in a Web Worker so the page stays responsive while the model is working.
Like every segmentation model, results depend heavily on the input. A portrait shot against a clearly different background — a person against a wall, a product against a tabletop — typically segments perfectly. Photos with motion blur, very fine hair detail, glass, fishing line, or busy backgrounds where the subject blends into the surroundings are harder. The mainstream paid services don't do meaningfully better on those edge cases, because the limitation is in the segmentation task itself, not the specific model. Soft edges around hair will sometimes need touch-up in an editor for a polished final result.
The output is a PNG with a transparent background, ready to drop into any compositing workflow: an e-commerce listing, a presentation slide, a profile picture, or a layered design in any image editor. Since the file is produced locally, you can keep iterating — re-run the model with a different setting, compose against different backgrounds, run the result through the compressor — without ever sending the image anywhere.
How to use the Background Remover
- 1
Drop a photo with a clear subject
The model works best on photos with a single clear subject — people, products, pets — separated from the background by lighting or contrast.
- 2
Wait for the first-time model download
On your first use, the browser downloads a ~4 MB AI model. This happens once and gets cached, so future removals start instantly. The progress bar shows how far along the load is.
- 3
Let the model run
Processing typically takes 2–6 seconds per photo on a modern laptop, longer on phones. The page stays responsive because the model runs in a Web Worker.
- 4
Download a transparent PNG
The result is a PNG with a transparent background, ready to drop onto another image or paste into a presentation.
Features
- AI segmentation that runs locally with WebAssembly
- No upload, no API key, no per-image cost
- Transparent PNG output ready for compositing
- Works offline after the first model download
- Web Worker keeps the UI responsive during processing
- Free for personal and commercial use
Common use cases
- Cut out a product for an e-commerce listing
- Remove the background of a portrait for a profile picture
- Isolate a subject before pasting it onto a different scene
- Prepare logos or illustrations for use on any background color
- Build hero images for a landing page from raw product photography
- Generate transparent thumbnails for app store screenshots
Tips and best practices
Use photos with good subject/background contrast
The model relies on visual separation between the subject and the background. A person in dark clothing against a dark wall is harder than the same person against a bright wall. If you have control over the input photo, shoot against a contrasting backdrop.
Mind the hair edges
Fine hair detail is the hardest case for every segmentation model, paid or free. If hair edges look slightly fuzzy or feathered, that's expected — sometimes a small amount of feathering produces a cleaner composite than a harder edge would.
Combine with the cropper for tighter compositions
Background removal isolates the subject; cropping reduces the empty space around it. Running the cropper after background removal often produces a much tighter, more useful final image.
Save the PNG, then compress
The transparent PNG output is uncompressed by design (PNG is lossless). For web use, run the result through the format converter to produce a smaller WebP that also supports transparency.
Composite against any color
Once you have a transparent PNG, you can place it on any background — solid color, gradient, another photo. Use the Collage maker or any image editor to assemble the final composition.
Technical details
How the model runs in your browser
The model is loaded via WebAssembly and runs entirely on the CPU (or GPU where available via WebGPU). No data ever leaves your machine — you can verify this by checking your browser's Network tab during a removal. After the initial model download, the tool runs even with the network disconnected.
Quality compared to paid services
On standard portrait, product, and pet photos, results are visually comparable to leading paid services. Both approaches use similar U-Net or transformer-based architectures and similar training data. The main differences appear in edge cases (very fine hair, fishing line, glass) where no current model is perfect.
Performance and battery considerations
Inference is CPU-intensive. On a laptop expect roughly one removal every few seconds with the fan spinning up briefly. On a phone, removals are slower and warm the device noticeably. For large batches, plug in or process in shorter sessions.
Frequently asked questions
How accurate is the AI background remover?
Quality is comparable to mainstream paid services on photos with a clearly separated subject. Complex hair, glass, and motion blur are still the hardest cases for any model — including server-side ones — and may need touch-up afterwards.
Why does the model take a few seconds to load?
The model is roughly 4 MB and runs on WebAssembly. Once cached, subsequent loads are instant. If your network is slow, the first load may take a minute or two; after that, removals are local.
Can I use the output commercially?
Yes. Whatever rights you have to your input image, you keep in the output. We do not claim any rights to anything you process here.
Does the model send my photo anywhere to process it?
No. After the initial model download from a CDN, every removal runs entirely on your device. You can verify this by opening your browser's developer tools and watching the Network tab during a removal — there are no outbound requests carrying image data.
What if the result has rough edges?
Rough edges usually indicate the model couldn't find a clear boundary — a subject that blends into the background, motion blur, or very fine hair detail. Try shooting against a more contrasting background, or touch up the alpha channel in any image editor (the eraser and brush tools work great on the PNG output).
Does it work on non-human subjects?
Yes. The model is trained on a wide variety of foreground subjects: people, animals, products, vehicles, food, and more. It does not require a specific subject type — anything visually distinct from the background should segment cleanly.
Why is the result a PNG and not a JPG?
JPG does not support transparency. PNG and WebP do. We export PNG by default for maximum compatibility, but you can run the result through the format converter to get a transparent WebP if you need a smaller file.