Qwen Image 2.1 Pro prompts: 5 scenes you can copy straight to an image

Qwen Image 2.1 Pro can be summed up in one sentence: with no reference images it does text-to-image, with reference images it does editing — the same model ID qwen-image-2.1-pro handles it all. The 5 scenes below are its real outputs; each comes with the complete prompt and parameters, ready to copy and use. It's live on HiAPI — open the model page to generate directly; new users get 200 credits on sign-up. Each image costs $0.05, so $10 gets you 200 images.

1. Furniture placement: two reference images, each with one job

Parameters: 16:9 · 2688×1536 · 2 reference images

Place the oak armchair from image 1 into the living room in image 2, with the chair on the patch of floor light in front of the floor-to-ceiling window, keeping the chair's orientation consistent with image 1, and add cast shadows and contact shadows that fit the scene for the chair. Keep the walls, wood floor, baseboards, window frames and the view outside the window of the room in image 2 completely unchanged, and unify the whole image under warm-white natural side light in the afternoon.

Generate with this prompt →

The two reference images are an oak armchair on a studio background and an empty living room: image 1 supplies only the chair, image 2 supplies only the room — the division of labor is fixed in the first sentence. Clicking the link only prefills the prompt; you upload the two reference images yourself on the model page. After keep ... completely unchanged, the walls, floor, window frames and so on are listed item by item — that's how the model avoids adding furniture on its own.

2. Chinese event poster: three lines of text spelled out line by line

Parameters: 3:4 · 1728×2368 · text-to-image · prompt_extend on by default

A minimalist Chinese event poster with a cream paper background; in the center of the frame, typeset only the following three lines of Chinese text, all horizontally centered: the first line is large black serif text "春日茶会" (Spring Tea Gathering), the second line is medium dark gray text "三月二十八日 下午两点" (March 28, 2:00 PM), the third line is small gray text "城南茶室 免费入场" (South City Tea Room, Free Admission). Even letter spacing, ample white space around all sides, and apart from these three lines there is no other text, graphic, line or decoration in the frame; matte print texture.

Generate with this prompt →

The font size, color and position of the three lines are written separately, and the last sentence stops the model from adding borders, seals and small icons on its own. This case uses the default prompt_extend; when you need strict layout matching the original text, set it to false so the expansion won't alter your fixed copy.

3. Perfume bottle makeover: change the material, not the bottle shape

Parameters: 1:1 · 2048×2048 · 1 reference image

Based on the perfume bottle in image 1, do an e-commerce product image rework: change the bottle's glass to a frosted semi-transparent texture, swap the silver-gray metal cap for brushed brass, change the liquid inside to deep amber, place the bottle on a dark green marble countertop, change the background to a warm gray wall, and use side-backlight from the upper left to outline the bottle's silhouette. Keep the bottle shape, proportions, cap height and label area shape from image 1 unchanged.

Generate with this prompt →

The reference image is a clear glass perfume bottle with a blank label area on the front; after uploading it on the model page, split the requirements into two lists: the items after change to / swap for are the changes, and the items after keep are the locked shapes. Leaving the label area blank is intentional — brand name and volume get overlaid in post, which makes revisions easier.

4. Line drawing to 3D toy: let the reference image handle the silhouette

Parameters: 16:9 · original generation 2688×1536 (displayed at 1600×914) · 1 reference image

Turn the line-drawing robot in the reference image into a polished 3D cartoon toy render: matte white shell with coral-orange joints, ears and antenna, a soft pink glow from the heart on its chest, bright blue eyes, keeping the original design, proportions and standing pose, soft studio lighting, clean light gray-white background, keep only the robot itself.

Generate with this prompt →

Line-drawing coloring fits the edit mode well: the silhouette, antenna angles and stance come from the reference image, and the prompt only determines material and color, so the design rarely drifts. Upload the line drawing on the model page and run; if you want a transparent background, don't rely on the prompt — both of this model's generations came back with opaque backgrounds, so hand matting off to a background-removal model.

5. Realistic portrait: natural skin in window light

Parameters: 3:4 · 1728×2368 · text-to-image · prompt_extend off

A realistic portrait photograph: an Asian woman around thirty sitting on a wooden dining chair by the window, body slightly turned toward the camera, wearing a light beige knit cardigan over a white crew-neck top, dark hair naturally tucked behind her ears. Soft afternoon window light comes in from the left of the frame, creating a natural light-to-shadow transition on her cheek and collarbone, with a clear window-shaped catchlight in her right eye. The background is a softly blurred light gray-green wall and a potted plant, 85mm lens, f/2.0, shallow depth of field, skin retains real texture and fine vellus hair, no skin smoothing, no plastic look. The subject occupies the center-right of the frame, gaze toward the camera, relaxed expression. No text, logos, watermarks or extra hand gestures in the frame.

Generate with this prompt →

Fix the light direction: window light from the left and a catchlight in the right eye keeps the natural light-to-shadow transition on the face. Then put "no skin smoothing, no plastic look" and "no extra hand gestures" at the end, and the two hard parts — skin texture and hands — stay under control.

General approach

• Write in the order "subject → environment → lens → lighting → details", and end with a sentence listing what should not appear.
• For editing, first assign each reference image a job, refer to them as "image 1" and "image 2", then list the keep items one by one.
• Pick the aspect ratio from five 2K-class options (1:1, 16:9, 9:16, 4:3, 3:4); when you need strict layout matching the original text, turn off prompt_extend; each image costs $0.05, $10 gets you 200 images, pick a scene and try one image first →.

Scene	Aspect ratio	Reference images	Key phrasing
Furniture placement	16:9	2	One image for the subject, one for the scene; list the keep items one by one
Chinese poster	3:4	0	Spell out the three lines word by word; turn off prompt_extend for strict layout
Product rework	1:1	1	Write the material changes and the kept shapes separately
Line-drawing coloring	16:9	1	Silhouette goes to the reference image, materials to the prompt
Realistic portrait	3:4	0	Fix the light direction; specify real skin texture

FAQ

Does clicking "Generate with this prompt" fill in the prompt automatically?

Yes. The prompt is already filled in when the model page opens; pick the aspect ratio and click run. Editing scenes still require you to upload reference images yourself, up to 10.

Will the Chinese text come out wrong?

All three lines of text on this page's poster are correct. The trick is to put each line's content in quotation marks, give it word by word, and specify the size hierarchy and position; when you need strict layout matching the original text, set prompt_extend to false.

Can it produce transparent-background PNGs?

The model's parameter docs say it can output RGBA transparent PNGs, but both of this model's generations returned opaque results (alpha all 255). If you need a reliable transparent background, use a dedicated background-removal model and don't put this step into your production pipeline.

How many reference images at most? How does the order take effect?

1–10 images, numbered in array order, referred to in the prompt as "image 1" and "image 2". No reference images means text-to-image, with reference images means editing — billing is the same either way.

How do I exclude an element?

2.1's inputs have no negative_prompt, so rewrite exclusions as positive descriptions: what is in the frame, how much of it, and where. For example, if you don't want extra furniture, write "there is only one chair and one window in the room".

Pro or Turbo?

Same inputs, reference image limit and aspect ratios; Pro costs $0.05 per image, Turbo $0.02 per image. Use Turbo for direction exploration and bulk drafts, Pro for final renders and precise layout.

---

Copy the furniture placement prompt above, head to the model page and generate your first composite →

Qwen Image 2.1 Pro series

• Qwen Image 2.1 Pro model page: generate online, 200 credits on sign-up
• View live image model pricing