Action Auto Generates. Don't Let It Invent the SKU.

OpenAI’s Responses API image-generation tool can choose whether to generate a new image or edit one already in the conversation. The optional action parameter controls that choice. Leave it on auto and the model decides. Set generate and it always mints a new file. Set edit and it is supposed to rewrite the still already in context. Auto is not a lock. A chat that “makes the kettle sharper” can return a cousin bottle. Lock catalog and campaign stills from a photograph you already trust. Leave type empty. Overlay real words last. FreeGPTImg is a GPT Image 2 workspace: 20 daily free credits, no credit card required. Do not stamp Flare, Sunburst, or action:edit on a GPT Image 2 JPEG, and do not export the auto-generated cousin as the product.

Draft in FreeGPTImg. Stay on GPT Image 2 when that is the label on the screen. Recrop a feed frame in AI social media ads. Keep materials honest in AI product photography. Put one unit on white in the Amazon product photo generator. Change only the field in white background product photo. Change only one noun from a locked hero in image to image. Start a net-new concept in text to image only when you do not yet have a photo.

What the job actually is

OpenAI’s image guide still splits the work: generate from a prompt, or edit an existing image, including partial edits. If you only need one still from one prompt, the Image API is the documented path. The Responses API is for conversational, multi-step image work. You pick a supported mainline model at the top of the request — the current examples use gpt-6-astra — and you name gpt-image-2.5-sunburst or gpt-image-2.5-flare inside the image-generation tool. Sunburst is the edit-precision pick. Flare is the everyday generate. FreeGPTImg currently exposes GPT Image 2 on this loop. Write the name on the screen into the filename. Do not relabel a GPT Image 2 still as Flare, Sunburst, or Responses.

action is the unused handle. Keep action: "auto" and the tool model chooses generate or edit. Set action: "generate" and it always creates a new image. Set action: "edit" to force editing when an image is already in context. You can keep a thread alive with previous_response_id, or by feeding prior image-generation outputs — including the image ID — back into context. A long thread is a mood board. It is not a catalog. Ten turns that “refine the glaze” are ten chances to mint a cousin SKU.

The same guide’s gift-basket example is the other trap. Four reference photos — lotion, soap, bath bomb, incense — become one new basket you may not sell. Reference images are a mood board. They are not four cartons plus a ribbon SKU. n can still be more than one. Multiple files from one chat are a concept set, not a listing gallery.

Quality on 2.5 includes low, medium, high, xhigh, max, and auto. Both 2.5 models default quality to auto. Pin the word if the still might ship. auto is not a grade on the pack. Recommended sizes are 1024x1024, 1536x1024, and 1024x1536. Custom WIDTHxHEIGHT must be multiples of 16, aspect between 1:3 and 3:1, edges at most 3840 px, total pixels 655,360–8,294,400. Above 2560x1440 is experimental. Transparent background needs png or webp. Complex prompts can take up to two minutes. Recheck the live docs; quality tiers and model IDs move.

Amazon’s product image help still treats MAIN as a realistic photo of the item you sell, on a pure white field (RGB 255, 255, 255), without extra graphics or type. An auto-generated cousin still is not MAIN. Meta’s image-ad guide for Facebook Feed currently recommends JPG or PNG, a 4:5 frame, and 1440×1800 as a working resolution, with a 600 px minimum width. That is a crop target, not permission to bake a chat CTA into pixels.

The cheap failure today is not “I did not read the Responses page.” It is “I left action on auto, the thread minted a new bottle, and now the grid disagrees with the box.”

Name the still before you generate

Weak sets fail because one Responses thread tries to be a concept still, an auto generate, a forced edit, MAIN, a 4:5 ad, a four-reference basket, and the signed headline. Pick one job.

Job	What you export	Keep	Change one thing	Wrong move
Concept still	A mood board or placeholder	One idea	Nothing else	Feeding it back as the product
Locked catalog still	The SKU you already ship	Shape, color, label, logo, pack size	Nothing, or light only	Starting from a generated JPEG
White MAIN draft	Same SKU, square, opaque white	The upload	Field only	Chat chrome, extra unit, restyled colorway
4:5 feed still	Same SKU, empty lower or top band	Silhouette, palette	Crop and empty band	Asking auto to “make it sharper”
9:16 story still	Same SKU, empty lower third	The object	Crop only	Status bars, prices, “Responses” badges
Forced edit	Same locked photo, action: "edit"	Everything the prompt does not name	One noun	Leaving auto to decide
Forced generate	A net-new concept, action: "generate"	The idea only	Nothing about a real SKU	Shipping it as the carton
Offer overlay	Don’t generate it	Empty type	Editor, last	“SALE”, “$29”, “Shop now” in the prompt

A text-to-image pass invents. An image-to-image pass should borrow. If you already paid for a pack shot that matches the listing, that file is truth. A Responses thread is a conversation. It is not a second product.

Checklist before anything leaves the thread

Use this sequence. Do not paste a campaign brief onto action: "auto" and expect the export to be the SKU you used to ship.

1. Name the job. Concept still, catalog still, MAIN draft, feed still, story still, forced edit, forced generate, or overlay. One file, one job.
2. If it is a catalog, ad, or listing still, start from a photograph. The locked JPEG of the item you sell. If you do not have that file, stop and photograph it. Do not invent a bottle because the thread is pretty.
3. Write keep / change / do-not in five lines. Keep silhouette, color, label, logo, and pack size. Change only one noun: crop, field, empty band, or one edit. Do not add a second SKU, a restyled colorway, or a percent.
4. Name the engine on the screen. If FreeGPTImg shows GPT Image 2, write gpt-image-2. If an API log says gpt-image-2.5-sunburst or gpt-image-2.5-flare, write that ID. If you set action, write edit or generate into the filename too. Do not relabel either file auto-approved.
5. Lock the still outside the chat first. Run the keep / change / empty-type loop on GPT Image 2. Only then send that PNG into a Responses thread, and only with action: "edit" if the file must stay the SKU.
6. Do not leave action on auto for a shipping still. Auto lets the model decide generate or edit. Generate is a new bottle. If the photograph is already truth, force edit. If you do not yet have a photo, force generate and keep that file in the concept folder.
7. Ask for empty type. Prices, percents, star ratings, “Shop now,” legal lines, and model names never belong in the generator if they must be exact. Overlay them in an editor after a person signs the words.
8. Leave production room. Product in the middle third. Quiet field. Empty band where the headline will sit. Do not fill every edge with chat chrome, arrows, or the last ten prompts.
9. Generate one still, then multiply. One locked pass first. File names like kettle-4x5-empty-band-gpt-image-2.png. Never kettle-final-auto.png for a listing. A four-reference basket is a concept. previous_response_id is a thread, not a SKU history.
10. Inspect the subject before the overlay. A rewritten mark, a floating “auto”, a ribbon that says “SALE”, or a tag that says “4.9” are stop conditions. Hold the still next to the real pack. If a stranger would think this brand shipped a chat edition, discard it.
11. Only then overlay. In an editor, set the real headline and only the claims a manager signed. Recrop to the live placement size after the pixels match. Do not ask auto to spell the price. Do not ask Sunburst to invent a lid that “reads cleaner at 4K.”

Prompt formula for the locked still

GPT Image 2 catalog or ad still, not a Responses generate and not MAIN unless named.
Start from the uploaded product photo.
Keep: exact silhouette, color, label, logo, pack size.
Change only: [crop / field / empty headline band / one noun].
Empty type. No prices, percents, ratings, model names, watermarks, or chat chrome.
Do not add: extra SKUs, restyled colorways, “Shop now”, celebrity faces, children.

Prompt examples

Locked catalog still. “Image-to-image: keep this exact matte-black kettle, spout curve, handle, and printed mark. Even studio light. Empty type. Not a Responses reshoot, not MAIN unless named.”

White MAIN, not an auto file. “Image-to-image: keep this exact SKU. Change only the field to opaque pure white. No badge, no extra unit, no chat arrows.”

4:5 feed, empty lower band. “Image-to-image: keep this exact kettle. 4:5 crop, product in the upper two thirds, empty lower band for later overlay. Quiet kitchen, even light. No prices, no Responses chrome, no second kettle.”

9:16 story, empty lower third. “Image-to-image: keep this exact mug glaze and printed mark. 9:16, empty lower third. No status bar, no ‘Shop now’, no model watermark.”

Forced edit after the still is locked. “Use this product photograph exactly as uploaded. action: edit. Change only the lid finish to the same matte black as the body. Do not restyle, recolor, or redraw the rest of the SKU. Empty type.”

Do not generate this. “action auto, make it sharper and more premium, new copper glaze, ten colorways, gift basket of the last four stills, Instagram CTA Shop now $29, 4.9 stars, guaranteed Amazon approved, cousin lid that looks cleaner at max.”

What to tell an explicit 2.5 Responses edit after the still is locked

If you already have the GPT Image 2 PNG, a Responses pass on gpt-image-2.5-sunburst can still be useful. Keep the instruction short and negative about the product. Set action: "edit". Pin quality. Put the photograph in context first.

Use this product photograph exactly as uploaded. Do not restyle, recolor, relight, or redraw the SKU.
Keep labels, logo, silhouette, and pack size accurate.
Change only: [one noun].
Leave all type empty. Do not invent prices, percents, or CTAs.
action: edit
quality: high

That is an edit pass. It is not a second generate of the object, and it is not a license to let auto invent a colorway.

Review the pixels before they leave the overlay

Zoom to 400%. A chat still hides a rewritten label. An overlay will not hide it either.

• Subject honesty. Same silhouette, color, label, and pack size. If a stranger would think this brand shipped a Responses edition, discard the still.
• Job honesty. A concept still stays in the concept folder. An auto export is an auto export until you prove otherwise. MAIN stays white, one unit, no graphics.
• Engine honesty. Write the model you used into the file name. A GPT Image 2 still is GPT Image 2. A Sunburst edit still is Sunburst plus action:edit. Do not stamp Responses on the FreeGPTImg JPEG.
• Empty type. Prices, percents, ratings, store chrome, and model names stay empty of anything you need people to read. The thread is a conversation, not a headline.
• Handle honesty. edit or generate. Not auto on a shipping file. Four reference photos are four references, not a gift-basket SKU you do not sell. If you needed a new object, you did not need edit.
• People and file. No recognizable generated faces you cannot clear. No children. PNG or JPEG, RGB, quiet field, no QR. Preview at feed size and at the size the catalog actually renders.
• Provenance. A 200 from the Responses API is not a license, not a marketplace approval, and not a signed offer. Recheck Amazon’s product image help and Meta’s ads guide before you tell a teammate the file is ready.

Limitations you should not skip

An auto export is not a signed discount, not a commercial-rights grant, and not MAIN. FreeGPTImg does not run the Responses API, action, Flare, or Sunburst. The default loop here is GPT Image 2 stills. Do not generate medical seals, “#1 2026,” store logos, or sale badges unless those claims are already on the real label and a manager signed them.

OpenAI’s own limitations still apply: complex prompts can take up to two minutes; text in the image can still miss placement and clarity; recurring brand marks can drift across generations; structured layouts are harder than they look. Treat those as reasons to lock the photograph first, not as a dare to chat longer.

OpenAI’s image guide is explicit about the two jobs: generate from a prompt, or edit an existing image. The Responses tool adds a third handle: let the model pick. Treat auto as a concept switch, not as a promise any model will honor an unsigned price. Quality tiers, background modes, and model IDs move. Recheck the live docs.

Models invent letters, digits, and product cousins when you mention Instagram, a dollar figure, “sharper,” “more premium,” or “make ten.” Shorten the brief. One SKU. One noun. Empty type. previous_response_id is not a backup of the PNG you already had. If the locked raster is honest, recrop that. If it is not honest, rephotograph. Do not ask a chat to remember a pack you never locked.

FAQ

What does action do on the Responses image tool?

It tells the tool whether to generate a new image or edit one already in context. auto lets the model decide. generate always creates a new file. edit forces an edit when an image is in the conversation. Recheck the live image-generation guide.

Should I leave action on auto for catalog stills?

No. Auto is a decision, not a lock. If the photograph is the SKU you sell, set action: "edit" and tell the model not to restyle the pack. If you do not yet have a photo, set generate and keep that file in the concept folder.

Does FreeGPTImg run action, Flare, Sunburst, or the Responses API?

No. FreeGPTImg is a GPT Image 2 workspace: 20 daily free credits, no credit card required. Use it to lock the still. Write the name on the screen into the filename.

Can I use an auto still as Amazon MAIN or a Meta ad?

Not if it changed the product, and not because the thread returned 200. MAIN is a realistic photo of the item you sell, usually on pure white, without extra graphics or type. A restyled glaze, a “premium” reshoot, or baked “Shop now” type fails that job. Lock the SKU first. Empty type. Overlay last. Recheck Amazon’s product image help and Meta’s ads guide.

Why did four reference photos become a gift basket?

Because the documented reference path can generate a new image from other images. Four stills in context are a mood board. They are not a SKU you ship. Do not feed four locked catalog files into a thread and export the ribboned basket as the product.

Should I lock stills in the Responses thread or outside it?

Outside it, if the file might ship as catalog, MAIN, or an ad. Photograph the item. Run the keep / change / empty-type loop on GPT Image 2. Then, if you still need a 2.5 edit, send that PNG with action: "edit" and tell the tool not to touch the rest of the SKU.

Sources

• OpenAI: Image generation
• OpenAI: Conversation state
• Amazon: Product image requirements
• Meta: Image ads guide

Lock the still before you leave action on auto

Pick the PNG that already matches the item you sell, not the JPEG a Responses thread invented because auto chose generate. Write five lines: keep, one noun, empty type, overlay later. Open FreeGPTImg, stay on GPT Image 2 when that is the label on the screen, recrop a feed frame in AI social media ads, keep materials honest in AI product photography, put one unit on white in the Amazon product photo generator, change only the field in white background product photo, run the edit in image to image, start a net-new concept in text to image only when there is no photo, and only then overlay the signed words — or send the locked PNG into a Responses pass with action: "edit", never auto on a shipping file, and tell the tool not to touch the SKU.