Qwen edit vs regenerate the still

When the still is already the product, Qwen Image Edit 2511 changes one locked file. Regenerating on GPT Image or Nano Banana is a new picture. Keep the file. Stop resampling it.

That split is a job pick, not a compliment. An instruction editor is allowed to be good at a pointed change. A text-to-image row is allowed to be good at a first plate. The mistake is treating "generate again" as a quieter edit. Same prompt. Same chat. New pixels. The face a human already signed does not return because the filename stayed.

This page is which job. The credit sticker is Qwen Image Edit 2511: change one locked still (3cr). How to write the sentence is Qwen Image Edit 2511 the instruction. Those are the meter and the phrasing. This is edit versus regenerate.

Edit is a pointed change

An edit starts from a file you would almost ship. One object is wrong, or one line of type, or one material. The rest of the frame is the product. You point. You name the verb. You name what must not move.

Qwen Image Edit 2511 is that contract on the Qwen shelf. Categories: edit-image and image-to-image. requires_image is true. A prompt-only call is not this slug. Live catalog on 19 September 2026: 2 credits on the discounted sticker, list 3, 50 percent off. max_reference is 6. No 4K badge. Aspects on the card: 1:1, 5:4, 4:5, 5:6, 3:2. Match the source. Do not ask an edit to become a 9:16 poster it was not given. Description on the row: instruction edits on a locked still (object change, style shift, or text fix) when you must keep the rest of the frame.

Siblings on the same job, different furniture:

• Qwen Image Edit is the 4K instruction card. Edit-image only. Requires an image. Max reference 1. Listed strength at changing or correcting text inside an image. Same live sticker: 2 credits, list 3. Open this when the hero is right and the letters in the pixels are wrong.
• Qwen Image 2 Edit is 3 credits, not discounted. Requires an image. Max reference 3. Aspects include 16:9, 4:3, 1:1, 3:4, 9:16. Use it when you need a later Qwen edit path and a slightly wider ratio list, not because the name has a 2.

Those three rows all refuse a blank canvas. That is the whole compliment. They are not generators wearing an edit label.

OpenAI has a true editor too. GPT Image 1 Edit is 2 credits, edit-image and image-to-image, requires_image true, max reference 6, 4K listed. Describe the change in plain English; the card says it repaints what you asked and preserves the rest. If the still is already the product and you specifically need OpenAI's instruction path, that is the GPT row. It is not GPT Image 2 text to image. People collapse those because both say GPT Image. One requires the file. One samples a new plate.

Bulk retouch that is not type surgery can sit on Pruna Image Edit: 1 credit, 4K, requires an image, max reference 5. Forty speckles do not need a Qwen turn. One wrong headline does.

The launcher for the edit job is the AI photo editor. The named agent job is edit a photo with AI. Both assume a still in the request. Neither is a first-picture aisle.

Write the instruction as a work order, not as a scene. "Remove the paper cup on the left. Keep the bottle, the label, and the crop." Not "premium bathroom, magazine light, same energy." Energy is a generate. The cup is an edit. One verb per run. Stacked wishes (new background, new cap, nicer hand, warmer grade) are how product edges melt and the lock you paid for becomes a cousin.

Edit is the wrong row when the composition is the defect. Wrong crop. Wrong person. Wrong camera height. Wrong era of lighting. Those are not objects you can point at. They are the picture. Pointing at "the vibe" is how an editor restyles the still you already signed.

Regenerate is a new picture

A generate samples a picture from a sentence. Even when you paste yesterday's prompt. Even when you attach last week's JPEG as a loose reference. The output is a new file. Identity is luck unless the row's contract is "keep this frame."

GPT Image 2 text to image is OpenAI's published generate card: text-to-image only, requires_image false, 4K listed. Live sticker: 9 credits at SD/HD (list 17), 17 credits at 4K (list 33), 50 percent off. Qualities 1K / 2K / 4K. Strong prompt adherence is the bet. It will not start from a locked still because the category list has no edit-image. Re-rolling GPT Image 2 because the bottle was almost right is how you buy a second bottle.

Nano Banana 2 is the Google still people treat as an editor because the card lists edit-image. Read the flags. Categories: text-to-image, image-to-image, and edit-image. requires_image is false. Live sticker: 4 credits (list 7), 50 percent off. Max reference 14. No 4K badge on max_output_resolution. Features listed: text_to_image, ultra_fast, lightweight. A prompt-only run on this slug is a generate. Attaching a file does not magically turn the call into Qwen. It can remix. Remix is not "keep the crop, delete the cup." Ultra-fast iteration is the product. A legal still is not an iteration toy.

Nano Banana is cleaner: text-to-image only, 4 credits, 4K listed, no image required, no edit category. Vibrant, cheap, a first plate. It cannot pretend to be a patch.

GPT Image 1.5 is the other dual-category trap. 2 credits, categories text-to-image, image-to-image, and edit-image, requires_image false, max reference 16, 4K listed. Conversational prompt following is the listed strength. A 2-credit generate that "almost" matches the approved still is still a cousin. Do not debug it as a failed Qwen instruction. It did what a generate does.

Qwen will also generate if you open a generate slug. Qwen Image is 1 credit (list 2), text-to-image, 4K. Qwen Image 3 text to image is 3 credits, 2K max, text-to-image, no image required. Those are first plates. They are not 2511. Family name is not a lock.

The generate door is text to image. The named agent job is generate an image from text. If the brief starts with "a photo of," you are here. If it starts with "change" and you already have the JPEG, you are not.

Regenerate feels like an edit because the chat is the same. You liked take three. Take four used the same words. Take four is a new sample. The label kerning that passed legal on take three is not a layer you can keep. It is a drawing the model may or may not draw again. Teams lose weeks this way: a hero that was signed, a "small tweak" in the generate box, a new face, a new kitchen, a new lawyer review.

Unpublished rows in this snapshot (cite the catalog, do not invent /models URLs): Qwen Image Edit 2509 is 3 credits, requires an image, max reference 6, prefer 2511 when you want the newer edit path. Nano Banana Edit is 2 credits, requires an image, 4K listed. Nano Banana 2 Edit is 4 credits, requires an image. GPT Image 2 image-to-image is 9 to 17 credits discounted and requires an image. If you operate those in-studio, they are still not "hit generate again on the T2I card."

When the still is already the product

"Already the product" is a human yes. Someone would print this file, run it in an ad, put it on a pack, or send it to image-to-video as frame one. The remaining note is a change, not a new picture.

Typical yes:

• The packshot is the SKU. The cup in the background is not. Edit. Do not resample the bottle.
• The talent is cleared. The jacket should be navy. Edit. A new generate is a new person wearing a navy idea.
• The poster headline is two letters off. Qwen Image Edit exists for in-image type. Regenerating the kitchen to fix a word is how you lose the kitchen.
• Four colourways from one hero. One verb each: "change the cap to matte black, keep everything else." That is a family. Four GPT Image 2 rolls are four families.

Typical no:

• You do not like the composition. The product sits too low. The crop is 1:1 and the placement is 9:16. That is a generate (or a reshoot), then a lock, then maybe an edit.
• The first plate was a sketch. Nobody would print it. Iterate on Nano Banana 2 or roll GPT Image 2 text to image until a freeze-frame would pass. Then stop generating.
• The lighting is the wrong decade. You cannot point at "decade." You need a new plate.

The cheap test is the sentence you would send a retoucher. If it names a region and a verb, edit. If it describes a scene, generate. "Make it more premium" is not a retoucher note. "Delete the cup, keep the label" is.

Do not mix the jobs to "see." An edit that drifted, followed by a generate "to get the magic back," is two bills and no lock. A generate you almost like, followed by a Qwen pass, is the honest sequence: lock the plate, then point. Animating the pre-edit file because the edit "mostly worked" is how the cup you removed comes back as motion.

After the still is signed, motion is a different row. Image to video takes the file you kept. It will not rescue a regenerate that almost matches last Tuesday's hero. It will faithfully move whatever you hand it, including the wrong letters.

Article 50 can apply to a realistic still of a real person. Read Article 50. A 2-credit edit does not exempt the file. A 9-credit generate of a new face is a new disclosure problem, not a fix for the old one.

The test

List the last ten stills you actually shipped.

If eight of them were files a human had already said yes to, and the note was a cup, a cap, a word, or a colour, you wanted Qwen (or GPT Image 1 Edit, or Pruna). Regenerating those eight on GPT Image 2 or Nano Banana was how the lock dissolved. Run Qwen Image Edit 2511 for the pointed change. Run Qwen Image Edit when the pixels contain words. Stop opening the generate aisle because the chat was already open.

If six of them were first plates (no file, wrong crop, sketch energy), regenerate was the honest job. Text to image is that door. GPT Image 2 text to image when prompt adherence and a 4K generate are the brief. Nano Banana 2 when you are iterating fast and you accept that a prompt-only call is a new picture. Nano Banana when you want the older 4-credit 4K generate with no edit category to confuse the ticket. Do not open 2511 without a still. It will not start honestly.

If the slate is mixed, write the job on the ticket: "Qwen edit, locked still" or "GPT Image 2 generate, no file yet." "Use Qwen" and "use Banana" are how a generate spends an edit credit, or how an editor is asked for a scene it was not given.

The AI photo editor stays the edit surface. The generate surface stays text-to-image. One studio can hold both. Living in one box is how the jobs collapse.

Pick the job, then sign up and pin the slug. Qwen keeps the still. GPT Image 2 and Nano Banana make a new one. That is the whole pick.

FAQ

Can I edit by prompting GPT Image again?

Not on GPT Image 2 text to image. That card is text-to-image only. A second sample is a new picture. If you already have the file and you need an OpenAI instruction path, GPT Image 1 Edit is the row that requires the image. Dual-category GPT Image 1.5 will run without a file; treat a prompt-only call as a generate, not as a patch.

Does Nano Banana 2 edit because the card lists edit-image?

It lists edit-image, image-to-image, and text-to-image, and it does not require an image. A prompt-only run is a generate. Fourteen references can lock a loop. They do not turn the slug into Qwen. When the still is already the product and the note is one object, Qwen Image Edit 2511 is the keep-the-frame job. Banana 2 is the fast iterate/generate still at 4 credits (list 7).

When is regenerate the honest job?

When there is no locked file, or the file would not print. Wrong composition, wrong person, wrong ratio, sketch. Generate on text to image until a freeze-frame would pass. Then stop. Further pointed changes are edits. Re-rolling GPT Image or Nano Banana after a human said yes is how you buy a cousin and call it a tweak.