Removal is really replacement. Something has to occupy the pixels the object was in, and the quality of the result is decided by how well you describe what that something is. OpenModels does this with a masked inpaint tool inside the image lightbox: paint over the area, write what belongs there, run it for ✦1. This page covers the tool, the prompt, and the cheaper approximation on the free lane.
You give the model two things: a black-and-white mask marking the region it is allowed to change, and a sentence describing what that region should contain. Everything outside the mask is left exactly as it was — that is the point of masking, and it is the difference between this and a prompt edit, which can and does redraw the whole frame.
Because the model has to render something into the masked area, "remove" is not an instruction it can follow literally. It will fill the space with whatever it thinks is most likely given the surroundings, and you get a much better result by deciding that yourself. If a bin is sitting against a brick wall, the prompt is "plain brick wall continuing the existing courses", not "remove the bin".
Which plan: Inpaint is a hosted call billed in ✦ gold, like every lightbox tool. There is no plan gate in the code, but green credits cannot pay for it and a new account is granted no gold — so it needs a paid plan or a credit bundle. The charge is held at submit, consumed on success and refunded if the run fails.
The mask editor is deliberately simple: paint, erase, a brush-size slider, undo and redo. Painting happens directly on the picture at the size it is displayed and the mask is scaled to the file when the job is submitted, so you are marking regions rather than pixels. There is no feathering control and no automatic selection, which means the accuracy that matters is your judgement about where the object's influence actually ends.
That boundary is nearly always wider than the object. A parked car occupies its own outline, its shadow on the road, the reflection in the window behind it and the sliver of pavement it hides. Mask the outline alone and the model dutifully fills a car-shaped hole and leaves the rest of the evidence in place. Mask generously and it has room to rebuild a scene that agrees with itself.
The other half of a good mask is what it does not cover. Painting over a face, a label or a straight architectural edge next to the object means those are re-rendered too, and re-rendered detail rarely matches its neighbours exactly. Erase back off anything you care about keeping.
| Situation | Weak prompt | Prompt that works |
|---|---|---|
| Bin against a wall | remove the bin | plain brick wall continuing the existing courses, same weathering |
| Person in a landscape | delete the person | empty gravel path receding to the treeline, same afternoon light |
| Logo on a shirt | remove the logo | plain navy cotton fabric with the same fold and shadow |
| Wire across the sky | remove the wire | clear overcast sky, unbroken, matching the gradient either side |
If you have no gold, there is a cheaper route that is worse at precision and free of provider fees: Qwen Image Edit (Local) on our own GPU, from ⚡1 an image. Attach the photo in the composer and write the instruction as an edit — "remove the bin from the left of the frame and continue the brick wall, keep everything else exactly as it is". There is no mask, so the model decides the boundary, and it may adjust things you did not ask it to. On simple backgrounds it often lands. On a busy frame it will not.
The trade is straightforward. The masked tool at ✦1 changes only the pixels you painted. The prompt edit at ⚡1 changes whatever it decides to, which is fine when the picture is simple and unacceptable when one region must be surgical. Both are on the same platform and there is no reason not to try the free one first.
Not with the masked tool, which is a hosted call billed in gold. The free approximation is Qwen Image Edit (Local) on our own GPU from ⚡1, driven by a written instruction with no mask — less precise, and often good enough on a simple background.
It needs both a painted mask and a prompt. Strokes with no text, or text with no strokes, leave it disabled on purpose — either one alone would spend a credit on a run that cannot do what you meant.
The mask stopped at the outline. Paint over the contact shadow and any reflection too, and re-run.
The tool has no idea what it is filling, but our Terms do. Use it on your own images. Removing our watermark from your own results is a plan setting, not an inpaint job — Creator and up can switch it off.
As many as you have gold for, at ✦1 each. Every run works from the untouched original, so a second attempt with a wider mask is a fresh start rather than an edit of an edit.