Pick the engine by what is in the frame
The three engines solve different problems. Running the wrong one on a photo wastes a download and gives a patchy mask, so match the engine to the subject before you press the button.
Portrait MODNet
A matting network trained on human portraits. The photo gets resized to 512 px on the short side before inference, which keeps runs quick on phones. Hair comes back with soft, partial alpha instead of a hard outline.
- Good at
- Headshots, selfies, group shots, full-body photos
- Gives up on
- Anything without a person. A sneaker on a table returns patchy or blank.
- Download
- 6.6 MB on CPU, 13 MB on WebGPU
Any subject BiRefNet lite
A salient-object model. The model sees a 1024 × 1024 copy of your photo and keeps the most prominent thing in the frame, person or not.
- Good at
- Products, pets, cars, furniture, plants, plated food
- Gives up on
- Busy scenes with no clear main subject, such as a full shelf, where the model guesses which object you meant
- Download
- 115 MB, once. Needs WebGPU to feel quick.
Flat backdrop No model
Reads the color along the photo's border, then floods inward until pixels stop matching within your tolerance. Enclosed gaps, such as the hole in a mug handle, stay filled until you erase them.
- Good at
- Packshots on a paper sweep, scanned signatures, logos on flat color
- Gives up on
- A white shirt against a white wall. The fill leaks wherever subject and backdrop share a color.
- Download
- None. Runs offline once the page has loaded.
From dropped file to finished PNG
- Decode. Your browser reads the file. Anything larger than 4096 px on the long side, or 12 megapixels in total, gets scaled down once so phones stay inside their canvas memory limits.
- Mask. A working copy capped at 2048 px goes to the engine. The engine returns one value per pixel: 255 keeps the pixel, 0 drops the pixel, and values in between become partial transparency along hair and motion blur.
- Edit. Edge shift, feather, and brush strokes change the mask, never the photo. Undo holds your last 12 strokes.
- Export. The mask gets scaled up to full size and applied to the original pixels, not the preview copy. Trim crops to the subject's bounding box plus the padding you set.
Settings for four everyday jobs
| Job | Engine | Background | Export |
|---|---|---|---|
| Amazon or Etsy main product image | Any subject | Color, white #FFFFFF | JPG, Trim on, padding 8% |
| Passport or visa photo | Portrait | White, or the shade your country names | JPG, then crop to the exact size |
| LinkedIn or team page headshot | Portrait | Blur, strength 18 | JPG, Trim off |
| Sticker, slide graphic, thumbnail cutout | Any subject | None | PNG or WebP, Trim on, padding 2% |
Amazon asks for pure white (RGB 255, 255, 255) behind the main listing image, with the product filling about 85% of the frame. Padding is measured against the subject's longer side, so 8% on each side puts the product at roughly 86%. For passport sizes, finish in the Image Cropper and check the rules your passport office publishes, since accepted shades and head sizes differ by country.
When the cutout looks off
A pale ring shows on dark backgrounds
The mask kept a pixel or two of the old backdrop. Set Edge shift to −1 or −2, then preview on the black swatch.
A hand or a bag strap went missing
Choose Restore and paint over the gap with a small brush. Ctrl+Z, or Cmd+Z on a Mac, reverses the last stroke.
Hair looks cut with scissors
Raise Feather to 2 or 3. On a Flat backdrop result, switch to Portrait instead, since the flood fill has no partial alpha.
The shadow under the product vanished
All three engines treat cast shadows as backdrop. Marketplace images on white rarely need one. For a lifestyle composite, add a soft shadow in your layout app.
The wall still shows between arm and body
Enclosed gaps survive more often with Portrait, where the input is only 512 px. Pick Erase and clear the gap by hand.
Limits worth knowing before you start
- Glass, smoke, veils, water. The engine outputs one alpha value per pixel with no color correction. A wine glass keeps the old backdrop's tint inside the bowl.
- Color spill. Green from a chroma screen, or red from a painted wall bouncing onto skin, stays in the edge pixels. Desktop editors with a decontaminate option handle this better.
- Batch jobs. One photo at a time. Two hundred catalog shots belong in a scripted pipeline or a paid API.
- HEIC files. Only Safari decodes iPhone HEIC photos. On other browsers, convert to JPG first.
- Low-memory phones. The Any subject model needs several hundred MB of RAM while running, and budget Android phones sometimes reload the tab mid-run. Portrait and Flat backdrop stay light.
- Edge detail ceiling. The mask comes from a 512 px or 1024 px model input and gets scaled up. Single hair strands follow the model's resolution, not your camera's.
This page or the Transparent Background Maker?
Toolexe has two cutout pages. The Transparent Background Maker works with color rules only, in color, edge, and low-texture modes, and saves PNG. Pick the Maker for logos and flat graphics when you want pixel-level control and zero download.
Stay here when the subject stands in a real scene, or when the goal is a new backdrop rather than transparency. After export, the Image Compressor trims a large PNG down before you upload the file to a store or a CMS.