Remove object from photo is sold as one job; it is two. We gave four AI editors one street photo and one instruction — delete every parked car, change nothing else — and checked what came back against the original. Nano Banana 2 at 1K was the only one that did all three: cars gone, face intact, same frame. Qwen image-edit did the first two and silently returned a different shape: 1194×1600 portrait in, 1280×960 landscape out, with the top of the frame cropped away. None of the three free sites we read that day says what it does to your input resolution.
Ask any tool to remove an object from a photo and you are sold two promises: the object disappears, and nothing else changes. On our run the first promise held. The second is the one that broke, and it broke silently.
So we tested the second promise on four remove object from photo models, with the original in hand.
Last updated: September 2026.
The remove object from photo test, in full
One source image, 1194×1600, of a woman mid-stride on a wet street with a row of parked cars behind her. One instruction, identical for every model:
Remove every parked car from the street behind the walking woman. Rebuild the road surface, kerb and pavement behind them so the street reads as empty. Keep everything else exactly as it is: the same woman with the same face, same hair, same camel coat, same jeans, same white trainers and the same mid-stride pose.
The person sits in the middle of the frame and is not the thing being deleted, which is what makes this a fair remove object from photo test. That is deliberate: it is the only arrangement in which you can see whether a model asked to edit the background quietly redraws the subject too.
Run 1 September 2026:
| Model | Published rate | Cars removed? | Face kept? | Frame kept? |
|---|---|---|---|---|
| Nano Banana 2 (1K) | $0.040/image | Yes, cleanly | Yes | Yes — 1792×2400, same 3:4 |
| Qwen image-edit | $0.030/megapixel | Yes, cleanly | Yes | No — 1280×960 landscape |
| Nano Banana edit (1st gen) | $0.020/image | task never returned | — | — |
| Recraft Remove Background | $0.003/image | wrong tool — cuts the subject out | — | — |
What "remove object from photo" got right
The deletion itself is a solved problem
Both finishing models removed every car and rebuilt the road, kerb and pavement underneath. No smeared patches, no ghost outlines, no half-car left at the edge of the frame. On this photo — ordinary geometry, ordinary light, a full row of cars — the remove object from photo step was not where either model struggled.
The face survived — and that is not a given
This matters because the same model failed the opposite test on the same day. We degraded a portrait to 2,983 bytes and asked Nano Banana 2 to restore it; it returned a sharp image of a different woman — narrower jaw, smaller mouth, the open smile closed. Same model, same day, opposite outcome. The full run is in unblur an image vs restore it.
The deciding factor is not the model and not the price tier. It is whether the pixels are still there. A model preserves what it can still see and invents what it cannot. A remove object from photo edit behind someone leaves her face untouched, so her face survives. Rebuilding a face from a ruined file leaves the model nothing to preserve, so it invents one.
What "remove object from photo" got wrong
Qwen changed the shape of your photo without saying so
Qwen image-edit did the remove object from photo job and kept the person — then handed back 1280×960 landscape from a 1194×1600 portrait input. The top of the frame, including most of the brick buildings, is simply gone. The composition is re-cropped.
If that output is going into a listing slot, a portrait crop, a print, or anywhere with a fixed aspect ratio, the edit succeeded and the deliverable is broken. Check the pixel dimensions of what comes back, not just the picture. This failure is invisible in every side-by-side comparison, because the comparison resizes both images to the same width.
Two of four never removed the object at all
The first-generation Nano Banana edit endpoint accepted the task and never returned a result. Recraft Remove Background is not a remove object from photo tool at all — it isolates the subject and drops the background, the opposite of what we asked for, and a common mis-click because both live under "remove" in every tool menu.
What free remove object from photo tools actually offer
The big free remove object from photo sites are large, and their terms are on their own front pages. Read 1 September 2026, with traffic from Columbus's July 2026 snapshot:
| Site | Monthly visits | DR | What its own page says |
|---|---|---|---|
| magiceraser.org | 1.1M | 40 | "Free AI Watermark & Object Remover (No Login)"; 3 credits per operation; claims 10 million users |
| logoremover.ai | 281K | 49 | free logo removal, "100% free", automatic detection, no manual editing |
| removerai.io | 200K | 13 | background removal; "No Watermark", "Free HD Download", "No Signup"; transparent PNG out |
All three lead with free and no signup, and none of them publishes what happens to your input resolution. That is the gap this page exists to fill: the thing worth checking is not whether the tool costs money, it is whether the file you get back is the same shape and the same person as the file you put in.
For reference, the rest of the shelf that can take this class of instruction, with published per-image rates read the same day: Nano Banana Pro $0.090 (1/2K), Seedream 4.5 edit and Ideogram v3-edit in the same band, GPT Image 2 image-to-image as the OpenAI option, and Topaz Image Upscaler at $0.050 if what you actually need is more pixels rather than fewer objects. None of these is dramatically better at deletion; they differ on what they do to everything you did not mention.
How to remove an object from a photo without losing the picture
- Name what stays, not just what goes. A remove object from photo instruction should spend one clause on the object and four on what must survive: ours named the coat, the pose, the trainers, the buildings and the light. Models drift on everything you leave unspecified.
- Keep the original file. It is the only way to catch a changed face or a changed crop, and both are invisible without it.
- Compare dimensions first, image second. Qwen's output looks correct until you read
1280×960. - Edit the background, not the subject, whenever you can. Asking a model to rebuild a face is a different and much riskier operation than asking it to rebuild a road.
- If identity matters, run it twice on different models. On Kavel both runs draw on one credit balance and the cost per model shows before you generate; the pricing page lists it per image.
The clean run: cars gone, road rebuilt, person and frame untouched.
Sources
Every figure was produced or read on the date named.
- Our own remove-object run, 1 September 2026: one 1194×1600 source, one instruction, four models. Outputs and dimensions as tabled above.
- magiceraser.org — "Free AI Watermark & Object Remover (No Login)", 3 credits per operation, 10-million-user claim. Read 1 September 2026.
- removerai.io — "No Watermark", "Free HD Download", "No Signup". Read 1 September 2026.
- logoremover.ai — free automatic logo removal. Read 1 September 2026.
- Published per-image rates for Nano Banana 2, Nano Banana edit, Qwen image-edit and Recraft Remove Background. Read 1 September 2026.
- Monthly visits and DR: Columbus AI-site index, July 2026 snapshot.
Remove object from photo FAQ
Can AI remove an object from a photo without leaving a mark?
On ordinary geometry, yes. Both models that completed our run rebuilt the road, kerb and pavement behind a full row of cars with no visible seam. The failure modes have moved elsewhere — to the crop and to the subject.
Does remove object from photo change the person in the photo?
It did not in our test, and that is the expected result: the subject's pixels were never destroyed, so there was nothing to invent. The risk appears when you ask a model to rebuild the subject itself, which is a different job.
Why did my remove object from photo result come back a different size?
Because some editors re-render at their own default aspect ratio rather than yours. Ours went from 1194×1600 portrait to 1280×960 landscape with no warning. Always read the dimensions of the returned file.
Which model should I use to remove an object from a photo?
On our run, Nano Banana 2 at 1K was the only one that removed the cars, kept the face and returned the same 3:4 frame, for $0.040 an image. Qwen image-edit did the first two for a similar price and changed the aspect ratio. If your output has to drop into a fixed slot, that difference matters more than the cent-level price gap.
What is the cheapest way to remove an object from a photo?
The free tiers listed above cost nothing and publish nothing about resolution. The wholesale rate for the same operation is $0.020 to $0.040 an image, which is what a paid credit is buying you — that, and the ability to check the result against the original.
