FreeGptImg
返回博客

AI Background Remover: Keep the SKU, Cut Only the Field

Upload one real product and remove only the background. Keep the SKU. Not a new pack shot, not a field swap, not an Amazon main image.

2026年10月14日FreeGPTImg TeamFreeGPTImg Team
AI Background Remover: Keep the SKU, Cut Only the Field

An AI background remover is a cutout, not a new pack shot. Upload the real item. Keep shape, color, label, logo, and pack size. Remove only the field behind it. The result is not a lifestyle scene, not a pure-white listing photo, and not a restored photo of letters that were never readable. OpenAI’s Image API lets you request background: "transparent" and export PNG or WebP on GPT Image 2.5 Flare and Sunburst. The GPT Image prompting guide still lists transparent PNG and WebP assets as a preview path on GPT Image 2. ChatGPT Images 2.5, launched 8 September 2026, is stronger at keeping a subject from a reference photo. FreeGPTImg is a GPT Image 2 workspace: 20 daily free credits, no credit card required. Overlay real words last.

Draft the still in FreeGPTImg. Run the keep-and-cut pass in the AI background remover, lock a listing-style white field in the white-background product photo tool, change only the scene in the product background changer, keep materials honest in AI product photography, cut a marketplace crop in the Amazon product photo generator, polish from a locked hero in image to image, and stay on GPT Image 2 when that is the label on the screen.

What the job actually is

The file you already have is the SKU. The field is the variable. If the bottle gains a new cap, a new flavor stripe, a marble slab, or a second product, you did not remove the background. You rolled a cousin.

OpenAI’s image-generation docs treat edits as a separate path from generating a new scene. You can ask for a transparent, opaque, or automatic background. Transparent output needs PNG or WebP; JPEG does not keep an alpha channel. The prompting guide treats product cutouts, icons, stickers, and presentation graphics as reusable plates, not as finished listing photos.

Amazon’s product image guide still treats the main image as a realistic photo of the item you sell, on a pure white field (RGB 255, 255, 255), without extra graphics or type. A checkerboard cutout is not MAIN. A white halo around a bottle is not MAIN either.

FreeGPTImg currently exposes GPT Image 2 on this loop. Write the name on the screen into the filename. Do not stamp a 2.5 or 2.1 badge on a GPT Image 2 file.

Name the cut before you generate

Weak files fail because they ask for a transparent PNG, a new white sweep, a sale badge, and “make it more premium” in one pass. Pick one job.

Job What you export Keep Change one thing Wrong move
Isolated cutout Same SKU, empty field Shape, color, label, logo, pack size Background gone New packaging, extra props, a second SKU
White listing field Different file, different tool The SKU A visible pure-white studio field Treating alpha as Amazon MAIN
New lifestyle field Different file, different tool The SKU Scene only Using the remover as a background changer
Light and texture polish Different file, different tool The SKU and the field already in the photo Lighting or sharpness only Cutting the field while “enhancing”
Ad-ready plate Same SKU, empty offer band The product silhouette Field gone Prices, “50% OFF,” stars, URLs, copyrighted characters
Forensic restore Do not export this Nothing useful Nothing honest Asking the model to invent missing detail

A cutout and a white listing photo are not the same file. The remover deletes the field. The white-background tool puts a visible white sweep behind the SKU. The background changer replaces the field with a named scene. Amazon MAIN still has to look white, not empty, not beige paper, not a faint checkerboard.

Leave prices, ratings, and headlines empty if those words must be exact. Models still invent letters. Overlay in an editor after the silhouette matches the upload.

Clean generation workflow

Use the same sequence in the AI background remover, a GPT Image 2 tab, ChatGPT Images 2.5, or a Gemini session. The rule is the lock, not the logo on the chat.

  1. Upload the real product. Front-of-pack, complete object, readable label, one subject. A tiny crop, a screenshot of a competitor, or a photo of three SKUs is the wrong source.
  2. Write five lines. SKU identity (shape, color, label words you must keep), the one job (cutout), crop, model name on the screen, and “no type.”
  3. Write keep / change / do-not. Keep the upload, including materials and print that already exist. Change only the field: remove it. Do not add a second product, a new backdrop, a logo you do not own, a person, or a price.
  4. Optional look reference is look only. Lighting direction may come from a second image. The product, logo, and packaging still come from the upload. Do not paste someone else’s listing as the hero.
  5. Generate three options on the same crop. Switching from 1:1 listing to 16:9 lifestyle between option A and option B is a new job.
  6. Inspect at 100%. Check melted labels, shifted logos, extra caps, white halos, leftover table edges, fake checkerboards baked into the pixels, and unreadably invented type.
  7. Route the file. Cutout, listing white, gallery lifestyle, and campaign overlay are four downloads. Do not reuse one JPEG for all four.
  8. Overlay real words last. Offer, date, and URL never belong in the generator if they must be exact.

Prompt formula

Product cutout, not a new SKU and not a new scene.
Model: [GPT Image 2 / ChatGPT Images 2.5 / Nano Banana 2.1 / name on screen].
Source: uploaded photo of [one real product, pack size, label color].
Keep: exact silhouette, materials, label layout, logo, cap, colorway.
Change only: remove the background. Empty field. No studio sweep unless this file is a white listing photo.
Crop: [1:1 / 3:4 / 16:9], product large in frame.
Text in image: none. Empty price and headline bands.
Do not add: extra products, people, pets, new backdrops, fake certifications, watermarks, sale badges, copyrighted characters, restored letters that were not readable.

Prompt examples

Isolated bottle, empty field

Keep this exact matte ceramic pour-over kettle, brushed steel handle, true product color, same spout and logo.
Remove only the background. Isolated subject, empty field, no table, no wall.
Square 1:1, kettle large in frame.
No new kitchen tools, no steam, no type.
Do not change the handle finish or invent a second kettle.

Portrait cutout for a profile plate

Keep this exact person from the upload: same face, hair, clothing, and pose.
Remove only the office background.
No new outfit, no beauty filter, no extra jewelry, no type.
Empty field behind the subject.

Logo mark, no campaign scene

Keep this exact brand mark from the upload, same geometry and color.
Remove only the paper or screen behind it.
Do not redraw the letterforms, add a drop shadow slogan, or invent a tagline.
Empty field. No type overlay.

Image-to-image, one noun

Image-to-image: keep this exact product, label, and cap color.
Change only the background: remove it.
Do not add type, props, or a second SKU.

Review the pixels before you upload

Zoom to 100%. A thumbnail hides a melted barcode.

  • SKU honesty. Same silhouette, colorway, pack size, and label. If a shopper would think it is a different flavor, discard it.
  • Field honesty. The background is gone, not replaced. A new marble slab, beach, or loft is a different tool.
  • Edge honesty. Hair, glass, steam, mesh, and clear bottles fail first. A white halo, a chewed spout, or a fake contact shadow glued to the object is not a clean cut.
  • Alpha honesty. JPEG cannot store transparency. If you need an empty field, export PNG or WebP and inspect it on a dark and a light ground. A checkerboard drawn into the pixels is not alpha.
  • Type honesty. Prices, stars, “new,” and URLs stay out of the pixels. Existing label words must match the pack you sell.
  • Rights honesty. Your product. No competitor packaging, no celebrity hands, no copyrighted holiday mascot.
  • Route honesty. MAIN, gallery, ad, and cutout are different files. Amazon’s main-image rules still want a realistic product on pure white, not a floating bottle on empty.

Limitations you should not skip

FreeGPTImg is a GPT Image 2 workspace with 20 daily free credits and no credit card required. The AI background remover is AI editing. That is not a claim that FreeGPTImg runs ChatGPT Images 2.5 or gemini-nano-banana-2.1, that every phone snap becomes a studio cutout, that missing print will be recovered, that you own commercial rights in every territory, or that a marketplace will accept the file. Match the label on the page you opened.

The remover is not a background changer and not a product photo enhancer. OpenAI’s API can request background: "transparent" on supported GPT Image models, with PNG or WebP output. That is a different export path from “make me a marble kitchen.” Keep the jobs apart.

Do not upload a cutout as Amazon MAIN because the kettle looks isolated at 200 pixels. Read Amazon’s product image guide for the category you sell. Leave type off MAIN. Keep sale treatments for other slots.

Models still invent extra objects when you ask for “more premium,” “gift-ready,” or “complete the kitchen.” Shorten the brief. One SKU. One cut.

Hair, smoke, water, and frosted plastic are hard edges. If the upload is busy, crop tighter or reshoot on a simpler field before you spend another pass.

FAQ

Is a background remover the same as a background changer?

No. A remover deletes the field and keeps the product. A changer replaces the field with a named scene and keeps the product. FreeGPTImg’s AI background remover does not promise a new backdrop or a pixel-perfect transparent PNG.

Can I use the cutout as an Amazon main image?

Only after you place it on a pure white field and inspect it against Amazon’s product image guide. MAIN still wants a realistic photo of the item you sell, on RGB 255, 255, 255, without extra graphics or type. An empty field is not that white sweep. AI output is a draft until a person checks it.

Will it recover a blurry label?

It may keep contrast on print that is already there. It is not a forensic restoration tool. If the letters were not readable in the upload, do not publish invented type. Reshoot, or overlay the real line in an editor.

Should I ask the model for the price and the headline?

No, not if those words must be exact. Leave price and headline bands empty. Overlay the real offer in an editor after the SKU matches the upload.

Does FreeGPTImg run GPT Image 2.5 or Nano Banana 2.1?

Use the label on the page you opened. FreeGPTImg currently exposes GPT Image 2 on this loop. Write that name into the file. Do not stamp a 2.5 or 2.1 badge because a docs page used that id.

Can I use someone else’s listing photo as the upload?

Do not. Upload a photo you have the right to use. A competitor pack shot is not a lighting reference. Look references, if you use them, control light only.

Sources

Cut one SKU on FreeGPTImg

Write five lines: the real SKU, the one job (cutout), the crop, the model name on the screen, and empty type. Open FreeGPTImg, upload in the AI background remover, lock listing white in the white-background product photo tool, swap a field only in the product background changer, keep materials honest in AI product photography, cut a marketplace version in the Amazon product photo generator, polish from a locked hero in image to image, and stay on GPT Image 2 when that is the loop in front of you. Download the image. Overlay the real offer last.