AI object remover: paint over it, and it is gone

What it does: removes a painted region of a photo and fills it in using the surrounding image as context — a person who wandered into the shot, a power line across a sky, a logo on a shirt, a stray object on a counter. You control exactly what gets touched by how you paint, not by how precisely you word a description, which is the main practical difference from a general instruction-based editor.

What it does not do: FLUX Erase does not read or use a prompt, so it cannot be told what to put in the erased area — it only reconstructs a plausible continuation of what's already there (more sky, more wall, more floor, more fabric). If the erased object was standing in front of something specific that needs to be recreated accurately — a sign with particular text, a piece of architecture, a pattern that has to match exactly — FLUX Erase's unguided reconstruction may guess wrong, since it has no instruction to work from beyond the mask itself.

It also is not a whole-subject isolator — that's a different job, covered by AI Background Remover, which produces a transparent cutout of an entire subject rather than erasing a region within a photo. And it's not a video tool: this page is about a single still photo. Removing something from a moving clip while keeping the rest of the footage intact is a separate capability (Object Swap / Motion Transfer on Genjutsu), not this one.

Both routes on VdoBloom use the identical mask-painting component: upload a photo, then paint over what you want to change with an adjustable brush that exports a black-and-white mask automatically as you draw. The difference is entirely in what happens after the mask is submitted. FLUX Erase (Black Forest Labs, via Runware) is purely mask-driven — no prompt field does anything, and the model's job is narrowly defined as "reconstruct the masked area from context." Qwen Image 2.1's mask mode (KIE's jobs API) is inpainting with instructions — the same white-marks-the-area convention, but the request also requires a text description of what should appear there, and the model uses that description to generate new content in the masked region rather than just extending what's already visible.

That difference decides which one to use. If the goal is genuinely to make something disappear with nothing distinctive put in its place — a person, a blemish, a cable, litter on a beach — FLUX Erase is the more direct tool: paint it, generate, done, no wording to get right. If the goal is closer to replacement — swap a sign for a different one, put a different object on the table, change what's in someone's hand — that's not really "removal" anymore, and Qwen 2.1's mask mode is the correct tool because it accepts a description of the replacement. Using Qwen's mask mode for pure erasure works too (describe "the empty wall" or "a plain sky continuing the existing gradient"), but it costs the same as any other Qwen 2.1 edit and requires writing that description, whereas FLUX Erase needs nothing typed at all.

Both routes share the same mask convention (image 1 = source, image 2 = mask, white = the area affected) and the same 10 MB per-file upload limit, but they're billed on entirely different schedules — FLUX Erase is a flat per-edit price with no resolution tiers, while Qwen 2.1 mask mode is priced the same way as its other resolution-based edits (1K/2K).

Cleaning up travel and event photos — a stranger who walked into frame, a tourist in the background of a landmark shot, a photobomb — where FLUX Erase's unguided reconstruction of "more of the same background" is usually exactly what's needed.

Removing small distracting elements from product or listing photos — a stray cable, a reflection artifact, a sticker or price tag left on — where the surrounding surface is uniform enough for the model to extend convincingly.

Taking a logo, watermark, or timestamp off a photo you have the rights to edit, where the area around it is simple enough to reconstruct (a plain wall, a sky, a solid-color garment).

Object replacement rather than pure removal — swapping what's on a table, in a hand, or on a wall for something specific — where Qwen 2.1's prompt-driven mask mode is the better fit, since it lets you describe the replacement instead of just erasing.

The two routes differ enough on cost and constraints that they're worth comparing directly — the numbers below come from the model configs and credit system, not estimates.

Paint generously around the edges of what you're removing, not just its outline — a mask that hugs the object too tightly can leave a faint ghost or outline behind, while a slightly wider brush stroke gives the model more room to blend the fill in.

For FLUX Erase, pick photos where the area behind the object is relatively simple — sky, wall, floor, uniform fabric. Busy or highly detailed backgrounds (crowds, patterned textiles, architectural detail) are harder to reconstruct convincingly without any guidance, since the model has no prompt to lean on.

If FLUX Erase's automatic reconstruction doesn't look right — the fill is blurry, mismatched, or guesses wrong about what should be there — try the same mask with Qwen 2.1's mask mode instead and describe what should be in that space (even something generic like "continue the plain wall") rather than leaving it to unguided inference.

Use the undo button in the painter if a stroke goes wrong rather than starting over — the mask painter keeps up to 20 prior states.

Work at a reasonable source resolution before uploading; the mask canvas caps at 2048px on the longest side, so a much larger source photo gets downscaled for painting and mask export regardless.