Qwen Image Edit: 12 Prompts That Change One Thing
Twelve copy-ready Qwen Image Edit prompts grouped by edit type, from text swaps to new camera angles, and a quick guide to choosing between the six Qwen edit models on each::labs.
By Eftal Yurtseven · Nov 17, 2026 · 9 min read

You asked for a red mug instead of a blue one. You got a red mug, a new table, a different hand and a window that wasn't there before. That's the real problem with image editing: changing one thing is easy, and changing only one thing is hard. Qwen Image Edit is Alibaba's answer to that second problem, and the prompts that work best with it read less like art direction and more like a work order. Say what changes. Say what stays. Stop there.
This is a collection, not an explainer. Below are 12 copy-ready prompts grouped by the kind of edit you're making, plus a quick way to choose between the six Qwen edit models on each::labs.
Which Qwen Image Edit Model to Pick
Alibaba's Qwen team ships edit models fast, so sort them by job, not version number.
Qwen Image 3.0 Pro Image Edit is the flagship. Alibaba Cloud's Model Studio docs list the 3.0 series as accepting one to three input images, processed in the order you send them, and the each::labs endpoint follows that: one to three images, an optional negative prompt, up to six outputs per run, and an output aspect ratio taken from the last image in your list.
The 2.1 family is the wide-reference option. Alibaba's model card for Qwen-Image-2.1 says it supports up to 10 reference images, preserves identity for people and products, and handles native transparency, meaning it can create and edit transparent layers and pull a subject out of a photo. On each::labs you get three flavors. Qwen Image 2.1 Pro Image Edit is the careful one, with a thinking mode switched on by default. Qwen Image 2.1 Turbo Image Edit is the fast one for drafts and batches. Qwen Image 2.1 Edit exposes guidance scale and lets you save PNG or WebP, the two formats that keep transparency.
Qwen Image Edit Plus runs Qwen-Image-Edit-2509, the release whose model card introduced multi-image editing (person plus person, person plus product, person plus scene) and text edits that change font, color and material, not just the words. And Qwen Image Edit 2511 Multiple Angles is a specialist: it takes no text prompt at all. You hand it an image and dial in a camera position with numbers.
Short version: 3.0 Pro for finals, 2.1 Pro when you need more than three references, 2.1 Turbo for volume, 2.1 Edit for transparent assets, Edit Plus for text and product composites, 2511 for new angles. If you've been writing prompts for generation rather than edits, our guide on writing better Qwen Image 3.0 prompts covers the other half of the job.

The Three-Part Edit Prompt
Every prompt below follows the same skeleton. First the change, stated as a verb: replace, add, remove, restyle, move. Then the anchor, which is the thing that must survive untouched. Then, if needed, how it sits in the scene.
The anchor is the part people skip, and it's the part that matters. Alibaba's own Model Studio example for a portrait edit opens with "Perfectly preserve the facial features and smooth black long hair of the young woman in the input image" before it asks for anything new. That sentence tells the model where its freedom ends.
When you send more than one image, name them by position. Alibaba's Qwen image editing guide says Image 1 is the first input, Image 2 the second and so on.

Object Swaps: Name the Thing and Its Neighbors
An object swap fails when the model can't tell where the old object ends. So describe it by location as well as by name, and say what touches it.
Prompt 1: product swap on a desk. Works well on 3.0 Pro or 2.1 Pro.
Replace the blue ceramic mug on the left side of the desk with a matte black steel travel tumbler of the same height. Keep the hand holding it, the keyboard, the notebook and the window light exactly as they are. Match the tumbler's reflections to the warm light coming from the right.
Prompt 2: remove and fill. Alibaba's guide shows removal working with a plain instruction ("Remove the hair from the plate"). Add what should replace the gap and you'll avoid a smudge.
Remove the parked bicycle leaning against the brick wall. Continue the brick pattern and the shadow line of the wall where the bicycle was. Do not change the door, the street sign or the pavement.
Text Edits: Quote the Words Exactly
Text is where Qwen's edit models earn their reputation. The Qwen-Image-Edit-2509 model card says it can modify text content and also edit fonts, colors and materials, and the Model Studio guide shows the same with examples like replacing the words on Scrabble tiles and turning a logo's lettering into glass.
Two rules. Put the new text in quotes, exactly as it should be spelled, with the capitalization you want. And tell the model what the text sits on, because a word printed on fabric, painted on a wall and stamped into metal should look different.
Prompt 3: sign replacement. Edit Plus or 3.0 Pro.
Change the text on the café chalkboard from "SOUP OF THE DAY" to "OPEN LATE FRIDAYS". Keep the hand-drawn white chalk lettering style, the same size and the same slight tilt. Leave the chalk dust and the wooden frame untouched.
Prompt 4: material change on lettering. This mirrors the "change the material to glass" pattern from Alibaba's docs.
Change the material of the word "NORTH" on the storefront to brushed copper with soft reflections of the street. Keep the font, letter spacing and position identical. Do not alter any other text in the image.

Style Changes Without Losing the Subject
A restyle is the biggest edit you can ask for, so it needs the strongest anchor. Name what has to survive the trip into a new look: identity, pose, composition, the product's shape. Both the 2509 and 2511 model cards put identity preservation front and center, and 2511 specifically describes reduced image drift and better character consistency.
Prompt 5: photo to illustration.
Redraw this photo as a flat gouache illustration with visible brush texture and a limited palette of ochre, teal and off-white. Keep the woman's face, hairstyle, pose and the composition of the frame the same. No outlines, no added text.
Prompt 6: relight, don't repaint. 2511's model card says popular LoRAs are now built into the base model, with realistic lighting control as one example.
Relight the scene as late golden hour, with low warm sunlight entering from the left and long soft shadows across the floor. Keep every object, color and texture the same; change only the light and shadows.
If you're on 3.0 Pro or 2.1 Pro, use the negative prompt field for the things that keep creeping in ("text, watermark, extra people"). On 2.1 Turbo there's no negative prompt field, so fold the exclusions into the main prompt as plain "do not" sentences.
Background Swaps and Scene Moves
Moving a subject is mostly a lighting problem. Say where the new light comes from and the edit stops looking like a sticker.
Prompt 7: studio packshot to lifestyle.
Place the perfume bottle on a weathered marble ledge in a sunlit Mediterranean bathroom, with a blurred window and green leaves behind it. Keep the bottle's shape, cap, label and liquid color exactly as in the original. Add a soft reflection of the bottle on the marble and match its highlights to daylight from the right.
Prompt 8: clean ID-style background. Alibaba's guide lists "Change the image to an ID photo with a blue background" as a subject-consistency example. Here's a fuller version.
Change the background to a plain, evenly lit light gray studio backdrop. Keep the person's face, hair, glasses and clothing unchanged, and keep the head and shoulders framing. Even, soft front lighting with no harsh shadows.
Need the subject on nothing at all? That's a job for the 2.1 family's transparency support. Our post on Qwen Image 2.1's transparent layers and ten references walks through it.

Multi-Image Reference: Count Your Images Out Loud
This is where a Qwen Image Edit prompt starts to look like a recipe. You give the model a person, a product and a place, then you say which ingredient goes where. Alibaba's guide example is the cleanest pattern there is: "The girl in Image 1 wears the necklace from Image 2 and carries the bag from Image 3 on her left shoulder."
Pick the model by how many images you have. Three or fewer fit 3.0 Pro and Edit Plus (the 2509 card says it performs best with one to three). Four to ten means 2.1 Pro or 2.1 Turbo. And on 3.0 Pro, put the image whose shape you want for the output last, because the each::labs endpoint takes the output aspect ratio from the final image.
Prompt 9: model plus product.
The man in Image 1 is holding the headphones from Image 2 in his right hand at chest height, in the city street from Image 3. Keep his face, beard and jacket from Image 1 and the exact color and shape of the headphones from Image 2. Match the lighting to the overcast sky in Image 3.
Prompt 10: two people, one frame. 2511's model card says it can fuse two separate person images into a coherent group shot, and the 2509 card lists person plus person among its supported pairings.
Combine the woman in Image 1 and the man in Image 2 into one photo: they stand side by side on a wooden pier at sunset, both looking at the camera and smiling. Keep both faces, hairstyles and outfits exactly as in their source images. Natural shared lighting and matching color grade.
Camera Angles: Set Numbers, Not Adjectives
Qwen Image Edit 2511 Multiple Angles is the odd one out, and that's why it's so reliable. There's no prompt box. You upload the image and set three numbers: horizontal angle (0 to 360 degrees), vertical angle (minus 30 to 90 degrees) and zoom (0 to 10, default 5). The model returns the same scene from that new position. The 2511 model card describes generating new viewpoints with the base model as one of the abilities folded in from popular LoRAs, and this endpoint pairs it with a dedicated Multiple Angles LoRA.
So the "prompts" here are settings. Copy them as they are.
Prompt 11: three-quarter product turn.
horizontal_angle: 45
vertical_angle: 15
zoom: 5
negative_prompt: distorted proportions, extra objects
Prompt 12: the side view for a turnaround sheet.
horizontal_angle: 90
vertical_angle: 0
zoom: 4
Run prompt 12 again at 180 and 270 and you have a four-view turnaround from one photo. Fix the seed across the batch so only the angle changes between frames. If you'd rather describe the angle in words, the Model Studio guide shows short viewpoint prompts like "Get a rear view." and "Face left." that work on the general edit models.
Edit Like an Editor, One Pass at a Time
The fastest way to ruin a good edit is to ask for four changes at once. Make one change, check it, feed the result back in as the new input, then make the next.
Mind prompt rewriting too. Prompt extend is on by default on 3.0 Pro, 2.1 Pro and 2.1 Turbo, and Alibaba's guide recommends it for short prompts and suggests turning it off when your prompt is already detailed. For precise text edits, switch it off so your quoted words reach the model untouched. Then set a seed once you like a direction, so the next run varies only what you changed.
Once a sequence of edits works, don't rebuild it by hand every time. Chain the steps in each::labs flows (background swap, then relight, then angle set) and run the whole pipeline on every new product photo.
Frequently Asked Questions
What is Qwen Image Edit best at?
Targeted changes that leave the rest of the image alone. Alibaba's docs and model cards highlight text editing (including font, color and material), adding and removing objects, keeping a person's or product's identity through a scene change, and combining several input images into one result.
How many images can I send in one Qwen edit?
That depends on the model. Qwen Image 3.0 Pro Image Edit takes one to three, and so does the practical sweet spot of Edit Plus. The Qwen Image 2.1 edit models on each::labs accept up to ten, which is what you want for multi-product layouts or several character references at once.
Does the 2511 Multiple Angles model need a prompt?
No. It works from your image plus horizontal angle, vertical angle and zoom values, with an optional negative prompt. That makes it predictable: the same numbers and seed give you a matching view every time.
Should I write Qwen Image Edit prompts in English or Chinese?
Either works. Alibaba's Model Studio documentation states support for both Chinese and English. Pick whichever you can write most precisely, and keep any text you want rendered inside quotes in the exact language and spelling it should appear in.



