NewSeedance 2.5just landed. A full 30 seconds, with sound, in one take.30 seconds, with sound, one take.Try it immediatelyTry it
Guides

4 min readMarkdown

Cut a subject out onto transparency, properly

Remove background returns a lossless image with a real alpha channel. What it handles cleanly, what hair and glass do to it, and where the cut-out goes next.

Remove background separates the subject from everything behind it and returns the subject alone, on transparency: a real alpha channel, not a white rectangle pretending to be one.

No prompt, no model, no selection to paint. One image in, one cut-out back.

Why the file format matters here

The result comes back as lossless WebP, and both halves of that are deliberate.

Lossless, because a cut-out is usually not the finished thing. It goes into a composite, a listing, a slide, a video. Every lossy re-save on the way there compounds, and the first place compression shows is the edge you just paid to get right.

With alpha, because the alternative is a lie. An image “on white” is not cut out; it is cut out and then pasted onto white, and the moment you place it on any other colour the seam appears. What you get here has genuine transparency, so it sits on a dark page, a coloured card or another photograph without a halo.

If your next step is a tool that cannot read transparency, flatten it there, once, deliberately, rather than starting flattened.

What it handles, and what it struggles with

The cut is decided by the model, and the model reads contrast and shape. So the difficulty is a property of your photograph rather than of your instructions.

Clean, every time:

  • A product on a plain or plainly-lit ground, the case this is most used for.
  • A person or an object with a defined silhouette against a contrasting field.
  • Anything already shot on a sweep, a table or a backdrop.

Where it gets hard:

  • Hair, fur and fine fringe. These are the classic failure and they fail softly: you get a good edge with a few strands lost. At small sizes nobody sees it; at full size on a dark background, everyone does.
  • Glass, smoke, water, anything translucent. Transparency in the subject is a different problem from transparency in the background, and cutting glass out cleanly usually means keeping some of what was behind it.
  • Subject and background the same colour. A white mug on a white sweep is ambiguous to a model for the same reason it is ambiguous to you at a glance.
  • Motion blur and shallow depth of field. A blurred edge has no single correct place to cut, so the model picks one.

The practical read: spend your effort on the input rather than on retries. A second photograph against a contrasting ground beats five runs against the first one, and it costs less.

What to do with it next

A cut-out is rarely the destination. Three roads it takes here:

  • Back into a generation as a reference. The result is a ready image like any other, so it can be a reference for a new picture, and a subject on transparency is a better reference than the same subject in a room, because a reference with no stated boundary donates its background to whatever you are making.
  • On to Vectorize, if the subject was flat art rather than a photograph. Removing the background first is what stops the tracer describing the desk the logo was sitting on.
  • Out of the studio, into a listing, a deck or a layout, where the alpha does its job.

When it goes wrong

There is a faint outline of the old background

The classic halo: a rim of the colour that used to be behind the subject, surviving in the semi-transparent pixels at the edge.

Fix: it is worst when the old background was bright and the new one is dark. Compositing onto a mid-tone hides it; re-shooting or re-generating the source against a ground closer to the destination removes it.

Part of the subject was cut away

Usually a section that reads as background: a gap between an arm and a body, a hole in a handle, a dark area that merges with a dark field.

Fix: raise the separation in the source. This is a contrast problem, and the model gets the same information you would give a person doing it by hand.

It kept a shadow I wanted, or dropped one I needed

Contact shadows sit exactly on the boundary between subject and ground and are resolved either way depending on how attached they look.

Fix: do not rely on the tool for the shadow. A composite reads better with a shadow you place deliberately in the destination than with one inherited from a photograph shot under different light.

Every tool is one press inside the studio.

No prompt, no model to choose. Pick the tool, drop the file in, and the finished version lands in your library beside everything else.

Start picturing