Qwen Image Edit 2511 prompting rules

Versely free tools and image edit handoff surface

Qwen Image Edit 2511 wants an instruction: the change, pointed at an object, on a still you already approved. It does not want a new scene prompt. A full "cinematic bottle on marble, dawn, 85mm" is how you restyle the face you already signed.

Search Console still shows "qwen image edit 2511" (22 impressions, position 14, seven days to 4 September 2026). The credit-meter page is Qwen Image Edit 2511: change one locked still (3cr). This page is how to write the sentence. The catalog row is /models/qwen-image-edit-2511.

The instruction is not a generate prompt

Qwen Image Edit 2511 is Alibaba's image editing model: precise modifications to an existing image based on textual instructions. Listed tasks: object removal, style transfer, content transformation. Categories: edit-image and image-to-image. Requires an image. 3 credits list, discounted on the published card. No duration. Audio empty. max_reference is 6. Aspects on the prompting surface: 1:1, 5:4, 4:5, 5:6, 3:2. Match the source. Do not ask an edit row to become a 9:16 poster it was not given.

Requires an image means there is no "describe a bottle from scratch" on this slug. Upload the locked still. Write the change. A text-to-image prompt is a different door. If you do not have a source file, you wanted text-to-image, not 2511.

The instruction has three parts, and only three:

Where. Left cup, red cap, the line of type on the label, the reflection in the window. Point. If two objects could match the noun, you did not point.
What happens. Remove, replace, recolour, correct the word, change the material. One verb.
What must not move. "Keep the label. Keep the face. Keep the crop." Say it. The model is an editor, not a mind reader. Unnamed regions are allowed to drift.

Bad: "Make it more premium, cleaner, luxury bathroom, soft light, magazine." Good: "Remove the paper cup on the left. Do not change the bottle, the label, or the crop."

Bad: "Woman wearing a blue dress, smiling, golden hour." Good: "Change the dress to matte navy. Keep her face, hair, pose, and the doorway."

Bad: "Add our new slogan somehow." Good: "Replace the headline on the poster with 'SHIPPED SAME DAY' in the same type style. Do not change the photo underneath."

Style transfer is still an instruction. "Apply the lighting and grade of reference image 2; keep identity and layout of image 1" is a transfer. "Make it like Wes Anderson" is a new generate with a famous name attached. Content transformation is a specified change: "swap the cap to matte black." It is not "make it cooler."

Extra references (up to six) are context for the edit, not a new identity hunt. Keep them on-model with the still you are changing: the same SKU from another angle, a colour chip, a type sample. A random mood board will pull the still toward the board.

Sentences that hold, and sentences that leak

Write as if a retoucher is on the other end of Slack. They cannot see your mood. They can see the file.

Removal. Name the object and the region. "Remove the stray hair over the left eye. Keep the rest of the hair." "Remove the price sticker on the lower right of the box." Do not say "clean up." Cleanup is twenty possible edits.

Recolour / material. Name the part. "Cap only, from gloss red to matte black. Label unchanged." Global "make it darker" will grade the face.

Type inside the image. This family is known for changing or correcting text in a still , a job most editors smear. Give the exact string. "On the front label, change '12 fl oz' to '250 ml'. Keep layout, colour, and the rest of the label." Do not say "fix the typo" if you have not said what the type should become.

Swap. "Replace the plant on the right with a white ceramic vase, same height, same shelf. Keep the bottle centre." A swap without a replacement becomes a removal plus invention.

Keep list. End with it. Faces, logos, crops, and legal claims belong on the keep list every time. If you do not say "keep the face," a dress change will sometimes become a new person. That is not a quirk you should romanticise. It is an underspecified instruction.

Do not re-describe the picture. "Photoreal bottle of serum on wet slate, 85mm, f/2.8" fights the pixels that are already there. The model tries to satisfy the description and drifts. Stage direction only.

One change per run when the still is expensive. If legal signed the face, do not also recolour the wall and add a plant in the same sentence. You will not know which clause moved the nose. Stack edits: run, check, run. 3 credits list is the meter for that patience. The sibling post is the bill. This page is why you do not spend those credits on a novel.

Match aspect to the source. If the still is 4:5, do not output 3:2 "for Instagram" on this row. Crop later, or lock a still at the delivery aspect before you edit.

After the instruction lands

The file is still a still. Animate it on image-to-video only when the object you pointed at is gone and the label you kept is the label. Multiple clips from one photo are I2V variants on that locked, edited plate , not a reason to skip the edit and hope the video model deletes the cup.

If the instruction failed, do not escalate to a new text-to-image of the whole campaign. Tighten the pointer. "The cup" → "the white paper cup, left third of the frame, sitting on the slate, handle facing camera." Then keep-list the bottle again.

If the instruction succeeded and you need a different motion pack, version the still: still_v02_qwen2511_no-cup.png plus the instruction text. That is AI asset versioning, a folder habit. I2V takes should hang off v02 so a buyer cannot pick a clip whose cup you already removed.

Do not use 2511 as a generator with a blank canvas. requires_image is true. The job will not run honestly without the plate. Do not use it as a lipsync, a video editor, or a layout tool for a 9:16 poster the source was not. Other Qwen image slugs exist on the Qwen roster; this page is 2511's instruction.

The photo editor is the launcher. Best image editing model is the category ranking. Neither writes the sentence for you.

A short list you can paste

Use these as scaffolds, then fill the brackets.

Remove [object] in [region]. Keep [identity list]. Do not change crop.
Replace [object] with [specific replacement, size, placement]. Keep [identity list].
Change [part] from [current] to [target material/colour]. Keep [identity list].
On [surface], replace the text "[exact old string]" with "[exact new string]". Keep type style, colour, and layout.
Apply the grade/lighting of reference [n]. Keep identity, pose, and layout of the base image.

If a scaffold needs more than one verb, split the run.

Qwen Image Edit 2511, the instruction: point, change, keep. Everything else is a generate you did not mean to buy.

Edit the locked still from the model page, then animate winners. Sign up when the plate is already approved and the next sentence should be a change, not a new scene.

Creator workflow that converts

Lock the still you are willing to defend (pack shot, board, UI plate).
Write one instruction that changes one thing.
Run Qwen Image Edit 2511 (see also Qwen Image Edit 2511 change one locked still).
Only then buy motion on Seedance / Kling if the still survives.
Burn captions after the video pass (caption styles).

Edit vs regenerate decision table

Situation	Use Qwen Image Edit 2511	Prefer a fresh generate
Locked product still, change one object	Yes	No
On-image label or typo fix	Yes	Often worse via regenerate
Brand-new scene from a prompt	No	Text to image
Then need motion	Edit first, then image-to-video	AI video generator

Open the model at Qwen Image Edit 2511 or start from AI photo editor. Credits are on the live model page.

FAQ

Why did my 2511 run restyle the whole photo?

The instruction described a scene instead of a change, or it skipped the keep list. Point at one object, name the verb, list what must not move. Re-describe lighting and lens only if the still is wrong and you accept a new picture.

Can I edit several things in one instruction?

You can. You should not, on a signed still. Stack runs so you can see which clause drifted the face. 3 credits list per run is the cost of that isolation.

Do I need reference images?

Only when the change needs a sample: a colour, a type treatment, another angle of the same SKU. Up to six. They are context for the edit, not a new cast. The base image is still required.