Free Image-to-Image Generator Workflow: Keep the Product, Change One Thing

A free image-to-image generator works when you upload a real photo, name one change, and lock everything else. OpenAI launched ChatGPT Images 2.5 on September 8, 2026, and said it is better at keeping people and products from reference photos while editing only the requested part. That is the same rule that already works on GPT Image 2: change the background, the light, or the crop. Do not roll a new product from text. Review labels, faces, and rights before you publish.

Start from a still you already trust. Draft in FreeGPTImg, open the image to image generator, keep the GPT Image 2 loop on GPT Image 2, fall back to a new still only in the AI text-to-image generator, keep materials honest in AI product photography, flatten a listing hero in the white-background product photo tool, adapt a marketplace crop with the Amazon product photo generator, or recrop a winner in AI social media ads.

What image-to-image actually does

Image-to-image is an edit job, not a second text-to-image roll. You give the model a reference photo plus a short instruction. The reference is the product, the face, the bottle, or the character. The instruction is one noun: background, lighting, wardrobe, or crop.

Text-to-image invents the subject. Image-to-image should not. If you describe the SKU instead of uploading it, you are guessing. Marketplace main images, brand kits, and character sheets fail that way.

OpenAI said Images 2.5 keeps reference subjects more reliably and follows multi-turn edits more closely. API users pick GPT-Image-2.5 Flare for speed or GPT-Image-2.5 Sunburst for tighter stills. Both handle text-to-image and image editing. FreeGPTImg remains a GPT Image 2 workspace: 20 daily free credits, no credit card required. Use the model in front of you. Do not relabel a GPT Image 2 file as Images 2.5.

Pick the job before you upload

Weak edits fail because they try to be a catalog shot, a lifestyle scene, and an ad at once. Name the slot first.

Job	Start from	Change one thing	Keep	Wrong move
White-background catalog still	A real product photo	Background to a clean white field	Silhouette, color, label art	Inventing the SKU from a text prompt
Lifestyle scene	The same product photo	Table, window light, or room	Product geometry and materials	Adding accessories that are not in the box
Social ad crop	An approved still	Aspect ratio and an empty headline band	Object, lighting, brand color	Stretching 1:1 into 9:16
Transparent cutout	An isolated product	Empty background, crisp edge	Materials and outline	A drawn checkerboard that is not alpha
Character consistency	A face or turnaround you already reviewed	Pose, outfit, or backdrop	Face, hair, marks	A new identity on every roll

Amazon Seller Central still treats the MAIN image as a realistic photo of the actual product. Aim for about 85% frame fill and a pure white background, RGB 255, 255, 255, unless your category allows a lifestyle MAIN. Graphics, mockups, and extra text are the usual fail states. Image-to-image can clean a real photo. It does not turn a generated illustration into a legal MAIN image. Read the current help page for your category.

One-noun edit workflow

Use the same sequence in a GPT Image 2 image-to-image tab, ChatGPT Images 2.5, or any other editor that accepts a reference photo.

1. Start from a file you would ship. A phone snap of the real SKU beats a pretty fake. A reviewed character sheet beats a new face.
2. Name one change. “Background to matte seamless white.” Not “make it premium, lifestyle, and ad-ready.”
3. Upload that file as the reference. If the tool asks for a photo, give it one. Do not skip this step and hope the prompt remembers the bottle.
4. Write keep / change / do-not. Three short lines. If a line is missing, the model will invent it.
5. Keep the crop. Generate three options at the same aspect ratio. A new ratio is a new job.
6. Inspect the keep-list first. If the label warped, the logo melted, or the face drifted, stop. Go back to the better base. Do not stack a second edit on a broken one.
7. Overlay real words in an editor. Prices, dates, URLs, and legal lines never belong in the generator if they must be exact.
8. Export for the real slot. Flatten when the destination hates transparency. Recrop in the matching tool instead of stretching.

OpenAI’s own image-prompting guide makes the same point for API edits: ask for an isolated subject if you need a cutout, set a transparent background only when the file format can keep alpha, and repeat the keep-list on later turns. A drawn checkerboard is not transparency. JPEG will not save an alpha channel. PNG or WebP can.

Prompt formula

Image-to-image edit, not a new invention and not a collage.
Reference: the uploaded photo is the only source of truth for [product / face / character].
Keep: [exact silhouette, color, label art, face, hair]. Do not redraw [logo / barcode / eyes].
Change only: [background to matte seamless white / window light from the left / 4:5 crop with empty headline band].
Camera and lighting: [eye-level, soft studio / 3/4, overcast window].
Text in image: none, or leave an empty band at [top / bottom].
Do not add: extra props, fake seals, extra fingers, watermarks, store UI.

Prompt examples

Catalog background. “Keep this bottle exactly. Change only the background to matte seamless white. Keep the label artwork, cap color, and liquid level. No props, no drop shadow that reaches the frame edge, no extra bottles.”

Lifestyle scene. “Keep this speaker exactly, including the fabric texture and logo stamp. Change only the setting to a oak desk near a window, soft daylight from the left. No extra gadgets. No invented cables.”

Social crop. “Keep this product and lighting. Recrop to 4:5. Leave the top 20% empty for a headline I will add later. Do not write prices or slogans.”

Character lock. “Keep this original character’s face, hair, and jacket marks. Change only the backdrop to a plain studio sweep. Do not age the face. Do not change the eye color.”

If the first batch moves the keep-list, shorten the prompt. One noun. Then retry. Free retry exists on FreeGPTImg when a generation misses. Use it on the same brief, not a new fantasy.

Review the pixels before you reuse the file

Zoom to 100%. Thumbs hide errors a laptop will show.

• Identity. Same bottle, face, and stitch line. If you cannot match it in three seconds, discard it.
• Type and marks. Labels and logos stay readable or they come off in an editor. Melted type is not close enough.
• Hands and edges. Extra fingers, fused straps, and hair halos fail ads and cutouts.
• Background. Off-white that looks white on a phone can still fail RGB 255, 255, 255. A drawn checkerboard is not alpha.
• Provenance and rights. OpenAI still uses C2PA metadata and an invisible watermark on GPT Image outputs. Downstream platforms may keep, strip, or flag that data. You also need rights to the photo, the product design, and any likeness. A better edit is not a license.

Limitations you should not skip

Image-to-image is not a studio replacement and not a compliance stamp. Amazon, Shopify, Meta, and Google publish their own image rules, and those rules change by category. Generate the still. Read the current page. Overlay only claims you can stand behind.

ChatGPT Images 2.5 is not GPT Image 2. Flare and Sunburst are API names. ChatGPT users get one Images 2.5 experience. On FreeGPTImg you are in a GPT Image 2 image-to-image loop. Write the model you used into the file name.

Better multi-turn edits are not infinite edits. If turn four rewrites a face, revert. Do not stack adjectives. Do not use a photo, celebrity, or mascot you do not own as a style reference. If a channel requires disclosure or forbids photorealistic people, follow that policy.

FAQ

What is a free image-to-image generator?

It is a tool that takes an uploaded photo plus a short instruction and returns an edited still. The photo is the subject. The instruction is one change. FreeGPTImg includes an image-to-image tab on GPT Image 2, with 20 daily free credits and no credit card required.

When should I use image-to-image instead of text-to-image?

Use image-to-image when the product, face, or character already exists and must stay recognizable. Use text-to-image when you are inventing a new still and have nothing to preserve. If a store, a brand kit, or a game bible cares about the object, upload it.

Did ChatGPT Images 2.5 change this workflow?

It made the same workflow more explicit. On September 8, 2026, OpenAI said Images 2.5 is better at preserving reference subjects and at editing one element while keeping composition and brand treatment. The creator rule did not change: one noun per edit, then review.

Can I use an image-to-image result as an Amazon main image?

Only if the file still meets Amazon’s current MAIN-image rules for that category. Seller Central asks for a realistic photo of the actual product, about 85% frame fill, and a pure white background unless your category allows a lifestyle MAIN. A generated illustration is the wrong file. Inspect RGB values, crop, and accuracy yourself.

Why did the product change even though I uploaded it?

The prompt asked for too many changes, or it described a new object on top of the photo. Shorten it. Name what must not move. Generate three options on the same crop. If the keep-list still drifts, the reference is too small, too dark, or the job needs a real reshoot.

Should I regenerate every catalog photo after Images 2.5?

No. Regen when the current file fails a real job: warped label, dirty background, or a crop that will not recrop. Keep files that already passed review. A new model name is not a reason to throw away a working still.

Sources

• OpenAI: Introducing ChatGPT Images 2.5
• OpenAI: Image prompting guide
• OpenAI: Create image edit API reference
• OpenAI: Images API reference
• Amazon Seller Central: Product image requirements
• Amazon Seller Central: Non-white background

Keep editing with a reference on FreeGPTImg

Upload one photo. Write one change. Keep the crop. Open FreeGPTImg, run the edit in image to image, stay on GPT Image 2 when that is the model in front of you, and send a clean still to AI product photography or AI social media ads only after the keep-list still matches the reference.