Most free background removers send your file to someone else's server. That is fine for a holiday photo, and less fine for an unreleased product or a client's image.
LabelOp Retouch removes the background on your own computer, inside the browser tab. This guide covers the whole process: the cutout, fixing the edges, choosing a new background and exporting the file. It also covers when the tool works well and where it falls short.
Short answer
Open retouch.labelop.com. The editor loads straight away, with no landing page and no sign-up. Drop in a photo, press R (Remove background) and approve the one-time model download. Touch up the edges with the Reveal and Hide brushes, pick a backdrop or keep the transparency, then export with Ctrl/Cmd+S.
The photo is decoded, processed and exported on your device. It is never uploaded.
What "without upload" means here
The background removal model is BEN2 Base (MIT license, released 2025). It runs through ONNX Runtime Web: on the GPU with WebGPU when your browser supports it, and in WebAssembly when it does not.
There is still some network traffic, so here is exactly what it is:
- Model files. The first time you use a tool, the editor asks before downloading its model. For background removal that is about 223 MB, fetched from Hugging Face. The browser caches it, so later runs work offline. The Models tab lists each model's source, size and license, and lets you remove cached models.
- Page analytics. The site uses Umami, a privacy-friendly page-view counter.
- Your photos. These are not in any request. Opening, editing and exporting all happen in the tab.
You can check this yourself. Open your browser's developer tools, go to the Network tab and run a cutout. The app's privacy page lists what stays on your device.
Before you start
Browser. Chrome and Edge 113 or later get full WebGPU support. Safari 26 and later currently use the WebAssembly fallback. Firefox is rolling out WebGPU and falls back to WebAssembly otherwise. The fallback works but is slower, and model compatibility varies by browser.
Device. Retouch is built for desktop first. It runs on phones, but mobile and touch input are not optimized yet.
File size. Files over 50 MB are refused, and so are images with a side longer than 16,384 px. Above 24 megapixels the editor asks whether to downscale. Past the browser canvas limit of about 67 megapixels, downscaling is the only option.
Step by step
1. Open the photo
Use Open in the header, drag a file onto the canvas, or paste an image with Ctrl/Cmd+V. A pasted image also stays in the browser.
2. Run Remove background
Choose Cutout in the toolbar's AI group, or press R. Before you run it, the panel shows one setting: Edge softness (0 to 10 px), which controls how soft the edge of the cutout will be.
Click Remove background. On first use, the inspector asks Download BEN2 Base · 223 MB? with Download and Not now buttons. Download progress, the running job and a stop button all appear in the same place.
When it finishes:
- the cutout becomes a new layer named after your photo with "cutout" added,
- the original photo stays in the Layers panel as a hidden backup, so nothing is lost,
- the editor switches to the Reveal brush and turns on the red overlay, so you can check the result right away.
Press C for a Before / after view and drag the split to compare.
3. Fix mistakes with Reveal and Hide
No automatic cutout is perfect.
- Reveal (B) brings back pixels the model removed. The stroke shows in green.
- Hide (E) removes leftover background. The stroke shows in red.
- Show hidden areas in red tints everything currently hidden, so gaps stand out.
Brush size goes from 2 to 300 px, and [ and ] change it while you paint. Hardness controls how soft the brush edge is. The layer header also has Invert and Clear for the mask.
The mask stays editable throughout. Hidden pixels are not deleted, so a Reveal stroke can always bring them back.
4. Refine the edge without rerunning the model
With the cutout layer selected, open Reveal or Hide to find Refine edges:
- Grow / shrink moves the mask edge by up to 20 px in either direction,
- Edge sharpness makes a soft edge crisper, or a crisp edge softer,
- Soften edges feathers the edge by 0 to 10 px.
The sliders preview live and do not rerun the AI model. Apply saves the result as one undo step, and Reset discards it. Apply or reset the preview before you export.
5. Remove color fringes
Cutouts often keep a thin line of the old background color along the edge. Remove color fringe (with a strength slider) and Clean edges fix this without an extra model download. The mask and transparency stay as they are, and one Undo reverses the cleanup.
6. Run it again if the first pass is off
If the cutout is wrong, for example because the edge is too hard, change Edge softness and use Remove background again → Run again. This replaces the mask on the same layer, which also discards your brush edits to it. Undo brings them back.
To keep the current cutout and try another one next to it, use New cutout layer instead.
7. Choose a background, or keep transparency
Under New background, you can put one of these behind every layer:
- White, Black or Gray,
- a Gradient,
- a Custom color from the color picker,
- your own picture via Image…, scaled to cover the canvas.
The backdrop becomes its own layer at the bottom of the stack. Picking another one replaces it. To export with transparency, just skip this step.
8. Export
Press Ctrl/Cmd+S or click Export.
| Format | Transparency | When to use it |
|---|---|---|
| PNG | Kept | Best quality, design tools, further editing |
| WebP | Kept | Smaller files for the web |
| AVIF | Kept | Smallest files, if your browser can encode AVIF (some apps cannot open it) |
| JPEG | Turned white | Opens everywhere |
The Size menu keeps the original proportions (with 100/75/50/25% buttons) or uses a preset: Instagram post (1080²), Instagram story (1080×1920), Amazon main (1000², white), Etsy listing (2000²), YouTube thumbnail (1280×720) or X post (1600×900).
For presets, Advanced settings has Pad (fit), which fits the image inside the frame and fills the rest, and Crop (fill), which fills the frame and crops the overflow. Choosing the Amazon preset also switches to JPEG with a white pad. Whether the result meets a marketplace's current image rules is still up to you to check.
Every export is re-encoded from the canvas pixels. Camera EXIF data, including GPS location, is therefore removed.
Removing backgrounds from many photos
Open Batch in the left sidebar, choose Remove background, and add PNG, JPEG, WebP or AVIF files. The images are processed one after another on your device. You can download each result on its own or click Save ZIP. Batch cutouts are saved as transparent PNGs.
Batch has no brush step. If an image needs manual touch-up, open that one in the editor.
When it works well
Background removal models of this kind decide on their own what counts as the foreground. In general they do best with:
- one clear subject, such as a product, person, pet or object,
- a background that differs from the subject in color or brightness,
- a subject that is in focus.
Busy scenes with several possible subjects are harder. So are subjects that blend into the background, and see-through materials like glass or veils.
The model works on a 1024×1024 version of the image, and its mask is then scaled back to your photo's full size. On large photos, very fine detail such as single hair strands may need a few Reveal and Hide strokes.
If you want one specific object rather than "the foreground", use Select object (S) instead. Click the object, add Include or Exclude points (Alt+click or right-click also excludes an area), then click Accept selection. This uses EdgeTAM, a separate model of about 21 MB.
Limits to know first
- First run takes longer. The model download is about 223 MB. After that it is cached.
- Speed depends on your hardware. The WebAssembly fallback is noticeably slower than WebGPU.
- No metadata round-trip. Exports never carry EXIF data. That is useful for privacy, but a problem if you rely on embedded camera data.
- Recovery saves a flat image only. The editor autosaves a flattened copy of your work in the browser. If the tab closes, it offers to restore that image, but not the layers, masks or undo history. Export when you are done.
- No text layers, vector layers or cloud sync. These are left out on purpose for now.
Where Retouch fits in the LabelOp family
Retouch is one of several LabelOp tools that run in the browser. Tracker annotates video. Moments finds a scene in a video when you describe it.
If you are cutting out objects for a computer-vision project, keep in mind that a cutout is a picture, not a label. Retouch gets the pixels ready. Use the LabelOp labeling workspace when you need segmentation labels in a versioned dataset. For the dataset side of cutouts, read preparing object images for a computer vision dataset.
Try it
Open LabelOp Retouch, drop in a photo and press R. You do not need an account. After the first model download, it also works offline.
FAQ
Is my photo uploaded when I remove the background?
No. The photo is processed in your browser tab, on your device. Network requests are limited to the one-time model download from Hugging Face and privacy-friendly page analytics. Your image is not part of any of them.
Why is the first background removal slow?
The first run downloads the BEN2 Base model, which is about 223 MB, after you approve it. The browser caches the model, so later runs start right away and work offline.
How do I get a transparent PNG?
Skip the New background step and export as PNG, WebP or AVIF. All three keep transparency. JPEG has no transparency, so transparent areas turn white.
Can I remove backgrounds from many images at once?
Yes. Open Batch in the sidebar, choose Remove background, add your images, then download the transparent PNGs one by one or as a ZIP. Images that need manual touch-up should be finished in the editor.
Which browsers are supported?
Chrome and Edge 113 or later get full WebGPU support. Safari 26 and later use a slower WebAssembly fallback, and Firefox uses WebGPU where it is available. Retouch is designed for desktop first.