Remove Image Background & Transparent Cutout
AI cut-out + touch-up brush Full-size transparent .PNG (up to 16.7 MP)
How Remove Image Background & Transparent Cutout Works
This tool cuts the subject out of a photo and gives you a transparent PNG, or puts it on a white, ID-blue or custom background. Its default mode is an AI segmentation model (BiRefNet lite) that runs inside your browser, so it handles people, products and pets on busy backgrounds. A second, non-AI mode removes a single backdrop colour, which suits logos and signatures. The model file is downloaded once; your photo is never uploaded.
How to remove an image background
- Drop a .png, .jpg or .webp image onto the workspace.
- In “AI (people, products, pets)” mode, press “Download model, ~66 MB, once” the first time. That is the model file (60 MB) plus the ONNX Runtime from cdn.jsdelivr.net; the tool names the exact download hosts before you start. After that both load from your browser’s cache.
- The AI finds the subject and the preview shows the cut-out on a checkerboard. Use “Cut-off” to remove more or keep more, “Crisp edges” for hard objects such as products, and “Feather” to soften the edge by a set number of pixels.
- Fix any mistakes with the brush: choose “Erase” to remove leftover background or “Restore” to bring back part of the subject, set the brush size, and paint on the preview. “Undo stroke” removes the last stroke. Switch the preview between “Result”, “Original” and “Mask” to check the edges.
- Choose a new background (Transparent, White, ID blue, or Colour with your own colour), pick .png or .webp, and export.
AI mode or plain backdrop mode?
- “AI (people, products, pets)” works on real photos with cluttered backgrounds: a person in a room, shoes on a carpet, a dog on grass. Fine details such as stray hair, fur and thin straps are the hardest part; check them in “Mask” view and use the brush where needed.
- “Plain backdrop (logos, signatures)” does not use AI. It removes one colour: the tool picks the most common colour along the border, or you set it with the colour swatch or “Pick from photo”. “Tolerance” sets how far from that colour still counts as background, and “Edge blend” softens the cut.
- In plain backdrop mode only areas connected to the border are removed, so a white shirt in front of a white wall survives. Turn on “Remove enclosed areas” to also clear holes, such as the inside of a ring or the counters of letters in a logo.
- If the AI model cannot download or run (blocked network, very old browser), the tool says so and switches to plain backdrop mode.
Speed, size and privacy
| Item | Detail |
|---|---|
| First use | About 66 MB downloaded once: the BiRefNet lite model (60 MB) and the ONNX Runtime WebAssembly (6 MB, cdn.jsdelivr.net). Every file is checked against a fixed checksum before it is used |
| Later visits | The model loads from this browser’s cache, with no new download |
| Speed | With WebGPU on a recent laptop, about one second per photo; on the WebAssembly (CPU) fallback, roughly 4–6 seconds on a laptop and longer on phones |
| Your photo | Processed in your browser tab and never uploaded |
| Output size | Same width and height as your photo, up to 16.7 megapixels; larger photos are worked on at 16.7 MP and the tool shows the size it used |
| .png | Keeps transparency; best for logos, signatures and product cut-outs |
| .webp | Keeps transparency in a smaller file; good for websites |
About the AI model
The model is BiRefNet lite by Peng Zheng and colleagues (MIT licence), run at 512 × 512 with 8-bit weights so it fits in a browser. The mask it produces is scaled back up to your photo’s full size. It runs through ONNX Runtime Web on your graphics card (WebGPU) or, where that is not available, on your processor (WebAssembly).
A note on passport and ID photos
Official photo rules often ask for an unedited picture. The U.S. Department of State requires a plain white or off-white background and says photos must not be digitally altered, so retaking the photo against a white wall is safer than replacing the background. Check the rules of the authority you are applying to before using an edited photo.
Questions
- Is my photo uploaded?
- No. The only downloads are the AI model and its runtime, once. The photo is decoded and processed in your browser tab, and the result is saved straight to your device.
- Why is the first run slower?
- The first time, the browser downloads about 66 MB and prepares the model. After that it loads from the cache. Without WebGPU the model runs on the CPU, which takes several seconds per photo but gives the same result.
- Part of the subject disappeared, or some background is left. What do I do?
- Lower “Cut-off” to keep more or raise it to remove more. For small areas, use the “Restore” or “Erase” brush on the preview. “Mask” view shows exactly what will be kept (white) and removed (black).
- Why is there a halo around my subject?
- In AI mode, raise “Cut-off” a little or turn on “Crisp edges”. In plain backdrop mode, shadows on the backdrop are a slightly different colour: raise “Tolerance” or “Edge blend”, or pick the backdrop colour again closer to the subject.
More in Photo & Visual
Portrait Skin Retouch & Blemish Brush
Smooth skin with an edge-preserving filter through a mask you control: skin-coloured areas are pre-selected, then paint Smooth or Protect to fix hair or background, and heal spots with a brush that copies nearby skin texture. Exports at full size up to 16.7 megapixels. No face detection or AI.
Crop, Rotate & Resize Image (1:1 / 4:5 / 16:9 / Passport)
Drag the crop box or its handles, lock a ratio (1:1, 4:5, 3:4, 16:9, 9:16) or an exact size such as a 35×45 mm passport photo, type the output width × height in pixels, rotate 90°, straighten up to ±45° and flip horizontally or vertically.
Portrait Relighting – Light Direction & Colour
Add a simulated studio light to a portrait: drag the light across the photo, set its strength, colour temperature and softness, and export at full resolution. A 2D effect driven by the photo's own brightness, with no depth or face model.