Mask resolution vs preview size

By Technical specialist

Mask resolution and preview size are related, but they are not the same measurement. Mask resolution describes the pixel grid that stores subject coverage. Preview size describes how large the editor draws the result on your screen. A 300-pixel-wide preview can represent a much larger image and mask; zooming that preview does not add new mask detail.

Confusing the two can lead to false conclusions. A cutout may look smooth when fitted to a small panel and reveal a halo at export size. It may also look jagged at extreme zoom even though those individual mask pixels are invisible in normal use.

Three sizes can exist in one edit

A browser cutout workflow can involve at least three coordinate spaces:

1. Source size: the decoded width and height of the chosen image.
2. Inference or working mask size: the grid on which subject coverage is calculated and edited.
3. Preview size: the scaled representation that fits the available editor area.

These sizes may match, but they do not have to. A large source can be reduced internally for model inference to control processing time and memory. The resulting mask can then be scaled into the working or export coordinate space. Finally, the browser scales the composite again to fit the screen.

Easy Background Remover can process very large images at a reduced internal resolution and show a warning. Export still targets the source dimensions. This keeps output dimensions useful, but it cannot invent edge detail that was absent from the internally calculated mask.

What a mask pixel stores

A mask is not necessarily a black-or-white stencil. Each mask pixel can hold an alpha value from background through partial coverage to fully retained subject.

At a hard object edge, the transition may occupy only a narrow band. At soft hair, fur, blur, or shadow, useful partial values may span a wider area. When the mask is resized, interpolation estimates values between the original mask samples.

Interpolation can make a transition visually smoother. It cannot recover the exact shape of a strand that disappeared during low-resolution inference. Smoothness and recovered detail are different qualities.

Why the fitted preview hides defects

When the editor fits a large image into a smaller panel, many source pixels contribute to each displayed pixel. Tiny holes, one-pixel fringes, and uneven alpha can be averaged away.

That is useful for composition: you can judge whether the subject is centered and whether the broad separation is correct. It is not enough for edge quality.

A reliable review uses more than one scale:

• fit view for the whole subject;
• moderate zoom for the outline and interior gaps;
• high zoom only for suspicious areas;
• intended-use size for the final judgment.

If the image will become a small avatar, a microscopic defect may never matter. If it will be printed large or placed over a dark background, the same defect can become obvious.

Why zoom can make a good edge look bad

At high zoom, the editor enlarges discrete pixels. Stair steps are expected on diagonal and curved boundaries because a pixel grid is not continuous. A soft transition may also look like a visible gray band when each pixel is enlarged.

The question is not whether pixels become visible at 800 percent. They will. The useful questions are:

• Is the boundary in the correct place?
• Does the edge look natural at the intended display size?
• Is there old background color mixed into partial pixels?
• Are important thin details retained?
• Does the alpha transition suit the subject?

Do not erase valid soft pixels just because they look gray against a checkerboard at extreme zoom. Test them over a realistic destination background first.

Source dimensions do not guarantee source-level detail

An exported PNG can have the same width and height as the source while its alpha mask originated at a lower resolution. Upscaling maps that mask across the full-size color pixels. The file dimensions remain correct, but boundary precision is limited by the information in the lower-resolution mask and any later refinement.

This is similar to drawing a simple outline on graph paper and enlarging it. The enlarged outline can be smoothened, but it does not acquire bends that were never drawn.

That does not make reduced inference useless. Many subjects have clear, broad edges, and a sensible internal size can provide a good balance between speed, memory use, and quality. It does mean that tiny details deserve special attention.

Brushes need coordinate mapping

When you paint Add or Erase strokes on a scaled preview, the editor has to map pointer coordinates back to the working mask. The brush radius also needs a consistent interpretation. A ten-pixel circle on screen should not accidentally become a completely different physical edit merely because the preview was resized.

Good mapping accounts for:

• the preview’s scale and offset;
• any zoom and pan transform;
• source-to-mask scale;
• device pixel ratio;
• brush radius and softness.

Small rounding differences are unavoidable, but repeated strokes should land where the pointer indicates. If a stroke appears offset, first check whether the canvas was panned or whether browser zoom changed the layout before assuming the model mask is wrong.

Light and dark backgrounds expose different pixels

Checkerboards communicate transparency, but they are not a universal quality test. Pale remnants from a white background can blend into light squares. Dark fringe can disappear against dark squares.

Review the same edge over:

• white or another light color;
• black or another dark color;
• a color near the intended destination;
• transparency if the export requires alpha.

An edge that works on all three backgrounds is more robust. If the final output always uses one fixed fill, optimize for that destination rather than chasing theoretical perfection everywhere.

Masks, color fringe, and false precision

Mask resolution controls where transparency changes. It does not directly remove old background color already blended into the source edge pixels. A correctly shaped mask can still reveal a pale or colored halo when the source was photographed against a strong background.

Shrinking the mask may hide that fringe by cutting into the subject. Expanding it may preserve more detail and more old color. The right correction balances shape and color at the destination size.

This is why a single numeric threshold cannot guarantee a clean export. Edge quality combines mask geometry, partial alpha, source color, output background, and viewing scale.

A practical export check

Before exporting an important cutout:

1. Confirm that the preview is showing the current refined mask.
2. Fit the image and check the overall subject.
3. Zoom into hair, fur, thin objects, gaps, and contact shadows.
4. Switch between a light and dark background.
5. Return to the intended use size.
6. Export and inspect the actual PNG in its destination when possible.

If the image triggered reduced-resolution processing, be especially cautious around narrow details. Manual Add and Erase refinement can correct intent, but it cannot recreate information that is not visible in the source.

For the automatic mask stage, see What an on-device cutout model sees. To understand why repeating settings can still require controlled inputs and versions, read Reproducible exports with the same settings.