How to Remove Objects From a Photo With an AI Assistant (and Recolor Them)

What erase actually does

Erase is inpainting: Mochify removes the pixels of the object you boxed and paints new pixels in their place, predicted from the surroundings. The result is a guess about what was behind the object, and a good guess on the right kind of object looks like nothing was ever there. Three things follow from that, and the rest of this guide keeps coming back to them.

It invents content. Tool pages like to say the AI replaces the object "with what is behind it". It cannot, because nothing behind the object was ever photographed. The grass behind the erased cone in the first run below is not grass that was photographed; it is grass the model believed should be there. That makes erase a cosmetic tool for tidying your own photos. It is never a way to hide something sensitive, because an invented patch can be read, compared and questioned in a way that a solid black block cannot. For a license plate, a face, an address or a document, use redaction instead; Erase or redact: which one you need draws the line.

Small and textured is easy; large and structured is hard. Mochify's inpainting model paints at a fixed 512 by 512, so a small distraction against pavement, grass, sky or a plain wall comes out clean, and a large removal, or one in front of a busy structured background, comes out softer. This is not a Mochify quirk. The research behind modern inpainting describes the same wall: these systems "often struggle with large missing areas, complex geometric structures, and high-resolution images" (Suvorov et al., "Resolution-robust Large Mask Inpainting with Fourier Convolutions", 2021). What goes wrong, and what to say next shows what "softer" looks like on a real car.

Something has to choose the box. Erase works on the REST API and on both of Mochify's MCP servers, where your assistant (or your own code) looks at the image and picks the rectangle. It is not in the mochify command-line tool, not a control in the web app, not a Magic Flow prompt, and not on any of the tool pages, for the same reason redaction is not: a text prompt cannot see the picture to place a box. Erase and recolor are on the paid plans, including the Day Pass (used from an assistant, on a free account you are signed in with); the Free plan's 25 images a month cover the image tools, redaction included, but not these two.

Why not let the assistant do the whole thing?

Because the photo a chatbot edits is not the photo you attached. When we measured what five AI chatbots actually receive when you attach a picture, between 80% and 94% of the pixels were gone before any model ran, removed by the upload step along with the camera data, the GPS position, the color profile and the HDR gain map; from a 5712 x 4284 iPhone photo, ChatGPT worked on a 2048 x 1536 copy, Claude on 2576 x 1932 and Perplexity on 1440 x 1080, and Gemini and DeepSeek returned no file at all (the full test is in our guide to what comes back when you ask ChatGPT to compress a photo). We tested compression in that piece, not object removal, so we will not claim what each chatbot does when asked to take a bin out of a picture. What we can say is that if it edits, it edits the shrunken copy, and what you get back comes out of its image model at the image model's size: a tidied listing photo at a fifth of its pixels, with no record of what else changed, or no file at all.

The MCP route separates the two jobs so neither one degrades the other. The assistant looks at the reduced preview, which is all it needs to decide where the object is, because the box is sent as fractions of the picture rather than pixels: a cone 78% of the way across a 1280-pixel original is 78% of the way across any preview of it, and the chat platforms we measured scale proportionally without cropping, so every fraction stays where it was. What the preview does cost is precision on small items, which is why the assistant draws with a margin and you check the result. Mochify, meanwhile, works on the original bytes: the file at the public URL you pasted, or the original you dropped on the one-time upload link. In the first run below the 1280 x 853 source came back at 1280 x 853 with one cone gone and nothing else changed; the bench photo came back at its full 1280 x 956 with a blue cushion and the same pebbles, floor and legs. The assistant's upload pipeline never touches the file Mochify receives, and that holds for any MCP client whose model can see a picture, whatever the model's own upload step does to it.

Three more things the split gives you that a one-step chatbot edit does not. The change is confined: only the pixels inside the outline Mochify finds are repainted or recolored, so the rest of the photo is your photo, not a regenerated likeness of it. The result tells you what happened: a count of what was erased or recolored, with the instruction to look at the spot, and on the REST API a header for how much of the frame was repainted. And the file is handled deliberately: the metadata and any HDR gain map are removed on purpose, because either could still carry the object, rather than lost as a side effect of an upload.

It is also usually cheaper. The edit is one tool call and one line of result text in the conversation, not an image generation; the finished file never enters the chat as pixels or as a base64 blob, so a batch of twenty listing photos costs twenty short tool calls, not twenty images' worth of tokens each way.

Four words this guide uses

• Box: a rectangle around the object, given as four numbers: left edge, top edge, width and height. Mochify takes them as fractions of the picture from 0 to 1, measured from the top-left corner of the photo as it displays upright, never as pixels. 0.775, 0.57, 0.07, 0.15 means "start 77.5% of the way across and 57% of the way down; cover 7% of the width and 15% of the height". Up to eight erase boxes, and up to eight recolor boxes, in one call. A box for erase should go around the whole object with a clear margin, shadow included; a snug box is the failure case, because anything left outside it can grow back.

• Erase: paint the boxed object out and fill the area with plausible background. Mochify finds the object's outline inside the box and repaints only the outline plus a margin, so as little of the scene as possible is invented; if no clear outline is found, the whole box is repainted.

• Recolor: change the color of the object inside the box and keep its shading. The outline is found the same way; the hue moves to the color you asked for and the lightness and saturation shift toward it, while every fold, highlight and texture stays where it was. Nothing is generated.

• Redact: replace an area with a solid block (or a blur) so its content cannot be read back. The privacy operation; a different tool with a different guide.

Order matters when you combine them: Mochify erases first, then recolors, then redacts, then crops to any region you set. So you can erase the garden light and recolor the bench in one call, and a crop can never bring an erased object back.

The fastest route: ask your assistant

Connect Mochify's hosted MCP server to an assistant that can see images (Claude Desktop, Cursor, or any client that supports remote MCP servers), attach the photo or paste a public link to it, and say what should go. The assistant finds the thing and measures the box; Mochify does the pixel work. The MCP server page has the one-time setup and the prompt for every capability; the hosted-versus-local guide goes deeper on the clients. If you run the local server (mochify serve) instead, the same instructions apply, with one difference in what comes back: the local server writes the finished file to disk and returns its path, with no download link and no pickup store.

Here is the first real run, exactly as it happened on October 19, 2026, on a 1280 x 853 street photo: an empty tree-lined road, a line of traffic cones along the curb, and one cone standing on its own where the grass verge meets the tarmac.

You say: "Remove the cone on the curb nearest the camera, the one on its own, and fill in the background."

The assistant does: looks at the image, measures the cone (about 78% of the way across and 57% down, a little under a tenth of the frame tall), adds a margin for its base and shadow, and calls Mochify's squish tool with one box:

eraseBoxes: [[0.775, 0.57, 0.07, 0.15]]
type: jpg

Mochify returns: Image processed successfully (image/jpeg, 226.5 KB), a download link, and the line Erased 1 area(s). Look at the result before calling it done: if part of an object, or its shadow, is still visible or has partly grown back, call again with a larger box around it. The cone is gone. The curb line and the edge of the grass run through the spot unbroken, the road markings beyond are untouched, and at normal viewing size there is no seam to find. The file came back at its original 1280 x 853.

The result line is advice, not a formality. Mochify can count the boxes it painted; it cannot judge whether the patch looks right in your eyes. Open the link and look at the spot before you move on. What goes wrong, and what to say next is about what you will sometimes see.

The link is short-lived. The hosted MCP server hands back a URL on files.mochify.app that expires in about five minutes; the assistant does not receive the image bytes, only that URL, so your conversation never fills up with a base64 blob. Click it, save the file, done.

If you attached the photo to the chat rather than pointing at a URL: the assistant sees the picture as pixels, not as a file, so it cannot forward your original by itself. It will call create_upload and give you a one-time upload link; drop the same photo on it and the assistant carries on with the returned upload id. The upload is deleted as soon as Mochify reads it, or after fifteen minutes if nothing does.

Ready to try it on one of your own? Connect the server, attach the photo, and name the thing you want gone.

What goes wrong, and what to say next

Every object-removal tool's marketing page shows the clean case. These are the three unclean cases from the same session, with the sentence that fixes each one.

1. The large object comes out soft. In the second run the assistant boxed three things in one call: the lone cone again, two distant pedestrians by a lamp post, and a car on the road about 8% of the frame wide ([[0.8, 0.42, 0.085, 0.105], [0.88, 0.42, 0.04, 0.09], [0.775, 0.57, 0.07, 0.15]]). Mochify returned Erased 3 area(s). The cone and the pedestrians were clean. Where the car had been, the tarmac carried a soft gray patch, the right color and roughly the right brightness, but visibly smoother than the road around it. That is the documented limit: the model paints at 512 by 512, and a mid-size object in front of a textured surface is where the invention starts to show. There is no sentence that fixes this one. If the object is large, the honest options are to crop it out instead ("crop to the left two-thirds and keep the cones"), to accept a soft patch that will pass at thumbnail size, or to leave it. Do not expect a second call with a bigger box to help; the box was not the problem.

2. Part of the object grows back. The box was too tight, and the model treated the piece outside it as scenery worth preserving. The result line tells you what to do: "the top of the cone is still showing, make the box bigger". The assistant redraws with more margin and calls again; the second call costs one more image from your allowance, nothing else. Shadows are the usual culprit. A person on a sunny pavement casts one that runs well outside their outline, and a removed person with their shadow left behind looks wrong in a way the eye catches instantly. Say "include the shadow" up front.

3. The model finds an object you did not mean. Erase and recolor both work by finding an outline inside the box. On a plain background that works as described; on a strongly textured one, the model can find "an object" where you see none. Recolor an object and keep its shading has the run where a recolor box drawn deliberately over bare pebbles, to see whether it would be skipped, was not skipped: the model picked a cluster of stones and recolored them. The lesson transfers to erase: keep boxes as tight as the margin rule allows, and look at the result rather than trusting the count.

Across the five calls in this session, the allowance went from 249 to 244: one image per call, whether it carried one box or three, erase, recolor or both.

Listing photos: erase the clutter, never the flaw

Marketplace sellers are the people who will use erase most, because a clean photo sells and the background of a spare bedroom rarely cooperates. The rule for using it well is one sentence: erase things that are not the item, never anything that is.

The platforms draw the same line. eBay's picture policy prohibits "photos that don't accurately represent the item" alongside added borders, text and watermarks, and requires at least one photo 500 pixels on its longest side (eBay picture policy). Vinted tells sellers they "must provide clear, accurate photos showing the actual item as it is" and may not use any image "that doesn't depict the actual item being sold" (Vinted, commercial selling). Neither bans editing; both ban editing that changes what the buyer thinks they are getting.

So the bin, the cable, the cat's tail, the second pair of shoes that crept into the frame, the garden light behind the bench: fair game, and exactly what erase is built for. The scratch on the item, the stain, the missing button, the scuff on the sole: not. A buyer who receives a flaw that was painted out of the photo has a case, and the platform's own words are on their side. If a flaw is the problem, photograph it clearly and describe it; sellers who do that get fewer returns, which is the subject of the Vinted photos guide and the marketplace photo requirements guide.

Two practical notes for the same audience. First, erase is for your own photos: the ones you took of the things you are selling. Second, "remove" means two opposite things in this corner of the internet, and the search results mix them up: remove the object (erase it from the scene, this guide) and remove the background from the object (cut the item out to use on white or elsewhere). If what you want is the whole background gone, that is background removal, which is on every plan including Free and is a plain Magic Flow prompt ("remove the background and make it white"); erase is for a few objects in a background you are keeping.

Recolor an object and keep its shading

Recolor answers the listing question "do you have it in blue?" without a second photo shoot, and the design question "what would the chairs look like in green?" without a mock-up. The assistant boxes the exact part to recolor, you name the color, and Mochify changes only the object inside the outline: the hue becomes the color you asked for, lightness and saturation shift toward it, and every fold, button, highlight and texture keeps its place. It is not generative. No pixel is invented; existing pixels change color.

The run, on a 1280 x 956 photo of a red leather bench on steel legs, standing on a wood floor in front of a bed of white pebbles:

You say: "Make the bench cushion navy and keep the shading. Leave the legs and the floor alone."

The assistant does: boxes the cushion, not the whole bench, and converts "navy" to a hex value:

recolorBoxes: [{ box: [0.29, 0.37, 0.67, 0.47], color: "#1e3a8a" }]
type: jpg

Mochify returns: Image processed successfully (image/jpeg, 167.5 KB) and Recolored 1 area(s). The cushion is blue with every tufted button and every crease intact; the steel legs, the wood grain and the pebbles are untouched, and the color stops at the leather's edge rather than at the rectangle's. Two things to know from that result:

• The color lands near your target, not on it. The leather's own lightness carried over, so a dark navy came out as a strong royal blue. That is what "shift toward" means in practice: a bright object asked to become a dark color keeps some of its brightness, because the shading is being preserved. If the exact shade matters, say "darker" or "more muted" and let the assistant adjust the hex and call again.

• A box that stops short leaves a sliver. The first box ended a fraction above the cushion's lower front edge, and a thin red line stayed there. "There's a red edge at the bottom front, box a little taller" fixed it: the second call, with the box's height at 0.5 instead of 0.47, had no sliver. Box all of the part, with a small margin, and remember that a hem, a strap or a sleeve often runs further than you think.

The combined call in that second attempt also erased a garden light standing on the pebbles behind the bench (eraseBoxes: [[0.785, 0.26, 0.09, 0.15]]), and both result lines came back together: Erased 1 area(s). … Recolored 1 area(s). Erase ran first, recolor second, in one image from the allowance.

What to box, and what not to: recolor takes everything inside the outline it finds, so boxing a whole person to change their jacket also tints their face, their glasses and whatever they are holding. Box the jacket. For "make his suit navy", box the jacket and the trousers as two boxes, not the person. Sheer or skin-toned fabric right against bare skin can tint the skin beside it; if the result shows that, say so and the assistant will tighten the box.

And the skip. A box with no clear object in it is meant to be left unchanged and reported rather than painted as a rectangle, and the REST API has a header for exactly that. In this session, a box drawn deliberately over bare pebbles was not skipped: the model found a cluster of stones and recolored an oval of them green. On a textured background the skip is best-effort, which is one more reason the result is something you look at, not something you assume.

Erase or redact: which one you need

Both take a box. They answer different questions, and choosing the wrong one is the one mistake in this guide that can actually hurt someone.

The reason is not that erase is weak. It is that an inpainted patch claims something about the scene: that there was grass there, that the wall was blank, that the pavement was empty. For a photo you are tidying, that claim is harmless. For a photo someone might rely on, it is a fabrication, and a solid black block is the version you can stand behind because it claims nothing. Mochify's own tool instructions put it bluntly: erase invents what was behind the object, so never use it to hide anything sensitive; use redaction for plates, addresses and documents.

The two can share a call. "Hide the plate and remove the bin" becomes one redactBoxes box and one eraseBoxes box; erase runs first, redaction second, and the redaction is applied to the already-cleaned image. The face and plate guide covers the redaction side, including why a blur is not the same as hiding.

What happens to the file

An erased or recolored image is processed like every other image Mochify handles, in memory with nothing kept, plus two rules that exist because these operations need them. Here is the whole path for the hosted MCP route.

1. The assistant sends the image and the boxes to api.mochify.app. The image does travel to our servers over HTTPS; this is not an on-device operation, and no honest guide should tell you otherwise. It is streamed into the encoder in RAM, never written to disk, and never appears in a log.

2. The inpainting and outline models run on Mochify's own servers. No third-party vision or image-generation service sees the photo, which is also why erase and recolor are not affected by the Dashboard's third-party AI setting.

3. Erase is applied first, then recolor, then any redaction, then any crop, resize, rotation or format change.

4. The output always has its metadata stripped and any HDR gain map removed, even if the request asked to keep them. EXIF usually holds a thumbnail of the picture as the camera saved it, and an Ultra HDR file carries a gain map that is a second, lower-resolution image of the scene; either would still show the object you erased. If you need HDR or metadata on a file, do that job on a copy you are not erasing.

5. The original is wiped as soon as the encode finishes. On the hosted MCP server, the finished file waits in a pickup store under an unguessable name for about five minutes so the download link works, then it is evicted whether you fetched it or not. On the local MCP server (mochify serve) and the REST API there is no pickup store: the bytes come straight back to your machine and nothing is held at all.

And as with redaction, the meaning of the edit never reaches Mochify. The assistant knows it was removing a bin or recoloring a jacket. Mochify receives 0.775, 0.57, 0.07, 0.15 and, for recolor, a hex code. The same in-memory model is described on our architecture and data handling page, and the processor obligations behind it are in the Data Processing Agreement.

Doing it in bulk: the REST API

If you have your own vision model, or a folder of listing photos and a script, erase and recolor are two query parameters on POST /v1/squish, in the same fraction convention as redact and region. Your code picks the boxes; Mochify receives numbers.

# Erase one object (box as fractions x,y,w,h), output JPG
curl -X POST "https://api.mochify.app/v1/squish?type=jpg&erase=0.775,0.57,0.07,0.15" \
  -H "Authorization: Bearer $MOCHIFY_KEY" \
  -H "Content-Type: image/jpeg" \
  --data-binary @road.jpg -o road-clean.jpg -D headers.txt

# Erase a garden light and recolor the cushion navy in one call, output WebP
curl -X POST "https://api.mochify.app/v1/squish?type=webp&erase=0.785,0.26,0.09,0.15&recolor=0.29,0.37,0.67,0.5:%231e3a8a" \
  -H "Authorization: Bearer $MOCHIFY_KEY" \
  -H "Content-Type: image/jpeg" \
  --data-binary @bench.jpg -o bench-navy.webp -D headers.txt

Several boxes are separated with ;. The # in a hex color is URL-encoded as %23. Read the response headers, not just the body; they are the detail the MCP result line summarizes:

• X-Mochify-Erased: how many boxes were painted out.

• X-Mochify-Erase-Outlined: of those, how many were repainted along a found outline rather than as the whole box. A 0 means every box was repainted whole, which is your cue to look closely at the result.

• X-Mochify-Erase-Coverage: the share of the frame that was repainted, from 0 to 1. A pipeline can use it as a threshold: above a few percent, route the image to a human.

• X-Mochify-Recolored: how many boxes were recolored.

• X-Mochify-Recolor-Skipped: how many were skipped because no clear object was found inside them. Redraw those tighter and call again.

Rules the API enforces: up to 8 boxes each for erase and recolor; every value is a fraction from 0 to 1; the operations need a paid plan (a 403 otherwise), and recolor returns a 503 if the outline model is unavailable, so retry rather than treat it as a bad request. Each request is one image from your plan's allowance. The full parameter table is on the API documentation page.

For a seller with fifty photos and no code, the assistant route handles the batch the same way: "remove the bin from all of these", one call per image, one result line per image to read through.

When you do not need any of this

You do not need erase when the object is the point, when the background is going anyway (background removal is the tool for that), or when a crop solves it: if the stranger is at the edge of the frame, "crop them out" is a cheaper and more honest edit than painting over them.

You do not need Mochify for a single casual photo on a recent phone. Apple's Clean Up does object removal inside the Photos app on iPhone 16 models, iPhone 15 Pro and Pro Max, iPads with an A17 Pro or M1 chip and later, and Macs with M1 and later, running iOS 18.1, iPadOS 18.1 or macOS Sequoia 15.1 and later (Apple, Requirements to use Clean Up in Photos); Google Photos has Magic Eraser; Photoshop has had a remove tool for years. If one of those is in your pocket and the photo is for you, use it. (If your iPhone is a 14 or earlier, Clean Up is not on it, whatever a tutorial says; the most common thread in Apple's own support community on this subject is someone looking for a button their phone does not have.) The route in this guide earns its keep when the photos are part of a job (listings, a property set, a client's gallery), when there are many of them, when you want the edit done from the same conversation that is already writing the listing, or when you need to know exactly what happened to the file afterwards.

And it does not cover video. Mochify's video tools run in your browser and do not include erase or recolor; removing an object from a video is a different job and this guide does not claim to do it.

Mochify workflow: erase and recolor from your assistant

Privacy note for this workflow: the photo travels to api.mochify.app for the edit and the encode, is processed in RAM and wiped immediately, and the finished file sits in the hosted MCP pickup store for about five minutes; the assistant only ever receives the download URL. The inpainting and outline models run on Mochify's own servers with no third-party service. What the box contained is known to your assistant and to you, never to Mochify.

Want to see the other things the same connection can do? The MCP server page lists every capability with the prompt that triggers it, and plain compression and conversion work in the browser at mochify.app with the same in-memory handling.

Cheat sheet

FAQ