Background removal
What is background removal?
Background removal separates the subject of a photo from everything behind it, leaving the subject on transparency or a plain color. AI tools do it with a segmentation model that finds the subject and a matting model that refines its edges, so hair, fur and glass stay natural. Nothing is redrawn.
By the DreamBench team. Updated .
Background removal at a glance
At a glance
- Input
- One photo.
- Output
- The subject on a transparent background or a plain color, usually as a PNG.
- How
- A segmentation model finds the subject; a matting model refines its edges.
- Keeps
- The subject’s own pixels: nothing is redrawn.
- Struggles with
- Hair against a busy background, glass, several subjects, a subject the color of what is behind it.
Which background do I need?
Transparent, for a logo, a sticker or a design: A PNG on transparency, so the cutout sits on any color of a slide, a website or a video. The transparent background maker gives you one in one step; check it on a colored background, not the checkerboard.
Pure white, for a shop or marketplace listing: A cutout on plain white, which is what most marketplaces ask for on a main image and what sites that flatten transparency onto black can’t spoil. The white background maker does it in one step.
Blurred, to keep the place but lose the clutter: The same photo with its background softened and the subject sharp, the way a portrait lens does. The blur background preset keeps every pixel of the subject.
A new scene behind the subject: A generative edit rather than a cutout: the subject is kept and a new background is drawn to match its light. The AI background changer does it from a description; expect the subject’s edges to be redrawn slightly.
How background removal works
A model trained on many labeled photos looks at the whole picture, finds the main subject and outlines it. The result is a mask, a map of the pixels to keep. A matting model refines the mask at the edges, so a strand of hair can be half see-through instead of cut off.
Segmentation and matting are the two steps people mean by “AI background removal”. Segmentation answers in or out for every pixel; matting adds how much, for the pixels on the edge.
Mask
Every pixel is in or out. Fine enough for a solid object, blocky on hair, smoke or motion blur.
Matte
Every pixel has a share of transparency, from 0 to 100%, so loose strands keep their softness.
Unlike a generative edit, background removal redraws nothing: the subject’s pixels are the photo’s own. That is why it keeps faces, labels and fine text exactly as they were.
Four ways to remove a background
| What it needs | Edges | Good for | |
|---|---|---|---|
| By hand | A pen or lasso tool and time | As good as your patience | A few important images, hard edges |
| Chroma key | A subject filmed against an even green or blue screen | Good, if the screen was lit evenly | Studio photos and video |
| AI segmentation | Any photo | Clean on solid shapes, blocky on hair | Products, logos, quick cutouts |
| AI matting | Any photo | Soft edges kept, strand by strand | People, pets, hair, fur, glass |
Where people use background removal
- Product photos for shops and marketplaces, often on white.
- Profile pictures and team photos on one background.
- Logos, stickers and cutouts for designs and slides.
- A clean subject to animate, or to place in a new scene.
Where background removal struggles
- Hair and fur against a busy background, where the strands and the scene are hard to tell apart.
- Glass, smoke and water, which are partly see-through to begin with.
- A subject the same color as what is behind it.
- Several subjects, when the model has to guess which one is the subject.
- Shadows: a cutout usually loses the subject’s own shadow, so it can look pasted on.
A pale fringe or halo around a cutout is the old background kept along a soft edge, where the matte is partly transparent: check the edge at full size on a dark color, and start from the least-compressed copy you have, because the blocks of a heavily compressed JPEG follow into the cut. A white shirt on a white wall is the hardest edge of all. The matte has no contrast to find, so part of the shirt can go with the wall or part of the wall stay; crop to the subject, or photograph it again against a different color.
How DreamBench removes a background
The AI background remover runs BiRefNet, a segmentation model its authors apply to tasks including matting and background removal. It works on photos up to 2048 pixels on the long edge and returns a PNG on transparency or plain white.
A cutout uses 7 credits. Presets make a transparent background, a white background or a blurred background in one step, and a cutout can open in the composer as a reference for a video.
| Figure | What it means for your photo | |
|---|---|---|
| Input | JPEG, PNG, WebP or HEIC, up to 40 MB | A photo straight from a phone is fine. A picture over 2048 pixels on its long edge is scaled to 2048 before the run. |
| Model | BiRefNet, one pass | A segmentation model its authors apply to matting: every pixel of the subject is the photo’s own, and the model only decides how much of each edge pixel to keep. |
| Output | A PNG at the upload’s size, up to 2048 pixels, on transparency or pure white | PNG keeps the alpha channel; white is flattened to #ffffff for marketplaces that ask for it. |
| Price | 7 credits a cutout, with any plan | The button shows the price before you press it, and a run that fails is refunded. |
| Time | A few seconds | A matte is one pass of a model, not a generation that draws the picture again. |
| Your upload | Removed from the tool’s storage after 7 days, with the result | It is used to make that result and nothing else, and not used to train a model. |
To take one thing out of a photo and keep the rest of the scene, the job is inpainting, which the object remover does from a description; to do the same in a clip, it is video inpainting. A cutout is the other way round: the subject stays and the scene goes.
Questions
What is the difference between segmentation and matting?
Segmentation decides, for each pixel, whether it belongs to the subject or the background. Matting also estimates how much of each edge pixel belongs to the subject, so soft edges such as hair keep their partial transparency.
How does AI background removal work?
AI background removal works in two steps. A model trained on many labeled photos finds the main subject and outlines it as a mask: in or out for every pixel. A matting model then refines the edges, giving hair and other soft edges partial transparency, and everything outside is removed.
Why does the background remover cut off hair or leave jagged edges?
A background remover cuts off hair when it uses segmentation alone: every pixel is either subject or background, so strands become a hard outline. A matting model keeps partial transparency at the edge instead, which is what DreamBench’s remover uses. Small or heavily compressed photos also lose the edge detail the model needs to find the strands.
Does background removal change my photo?
Not when it is done by a segmentation or matting model, which only decides which pixels to keep, so a cutout loses no quality of its own: quality is lost only if the photo was small or heavily compressed to begin with. A generative tool that repaints the background may change the subject slightly, which is why DreamBench makes its cutouts with a matting model.
What file does a background remover give me?
A PNG, because PNG keeps transparency and JPEG can’t. DreamBench returns the cutout on transparency or on plain white; pick white for marketplaces that flatten transparent images onto black. The cutout keeps the photo’s own size up to 2048 pixels on the long edge.
What background do marketplaces want for product photos?
Most ask for the main image on a plain white background with the product filling most of the frame, so check the marketplace’s own image rules before you upload. The white background maker gives a cutout on pure white in one step.
Why does my cutout look pasted on?
Because it lost its shadow and the light on it no longer matches the new background. Add a soft shadow, or put it on a background lit from the same side.
Can DreamBench remove a background from a video?
No. DreamBench’s background remover works on photos. Removing a background from a video takes a model that follows the subject from frame to frame.
Can I remove the background from several photos at once, or on my phone?
One photo per run: DreamBench has no batch upload, so a set of product photos is done one at a time. The remover works in a phone browser, with nothing to install, and takes a photo from your camera roll.
How long does background removal take?
Background removal takes a few seconds on DreamBench. The matting model makes one pass over the photo, at up to 2048 pixels on its long edge, and the PNG is ready to download when it finishes; nothing is generated and nothing is retouched afterward. A photo larger than that is scaled down before the run, and a run that fails is refunded.
What happens to the photo I upload?
The photo you upload stays in your browser until you run the cutout. Then it is used to make that result and nothing else: DreamBench doesn’t use uploads to train or fine-tune AI models, and the photo is removed from the tool’s storage after 7 days, along with the result. A result you keep in your library stays as long as you keep it. Upload only photos you have the right to use.
What is the difference between cropping a photo and cutting out the subject?
Cropping keeps a rectangle of the photo and discards the rest, background included; the subject still sits on whatever was behind it. A cutout keeps the subject’s own outline and makes everything outside it transparent or white, so the subject can sit on any background. The two go together: crop to the subject first when a busy scene has several things the model could take for the subject, then remove the background.
Does background removal cost anything?
It depends on the app. On DreamBench you create a free account and choose a plan, and each cutout uses 7 credits.
Related
DreamBench tools
- AI background removerCut the subject out of any photo as a transparent PNG, edges and all.
- Transparent background makerMake any background transparent and download a PNG with see-through pixels, edges and all.
- White background makerPut the subject of a photo on plain, pure white, edges blended and nothing redrawn.
- Blur background toolBlur the background of a photo and keep the subject sharp, cut out pixel for pixel.
- Logo background removerCut a logo, an icon or a signature out of its background as a transparent PNG.
- AI background changerPut the subject of any photo in a new scene, with the light to match.
Terms
- Transparent backgroundPixels with no color of their own, stored in an alpha channel, so the page behind shows through.
- InpaintingFilling a chosen area of a picture with new content, the way AI removes objects and text.
- Video inpaintingFilling a region of a video so it fits every frame, which is how AI removes objects and people.
Sources
- BiRefNet on GitHubBiRefNet, read on
DreamBench is an independent platform and isn’t affiliated with or endorsed by the companies named on this page.