Prompt Tip

How to keep garment length, sleeves, and details accurate with fewer redos.

Accuracy comes from your reference images, not from prompt length. The prompt only decides what changes and what stays.

Four rules

  1. 1.

    Add a worn shot

    Length, sleeves, and fit only transfer from a photo of the garment being worn. From cutout or pack shots alone, the AI guesses. A phone photo is enough; crop the face out.

  2. 2.

    Length in words, not numbers

    Write "cropped above the ankle" or "sleeves down to the wrist." The AI can't follow "2cm longer." It can follow landmark words: ankle, knee, wrist.

  3. 3.

    Build in two steps

    Finish one accurate front-facing base shot, then change pose and background with Edit Image. Asking for everything at once is what warps the garment.

  4. 4.

    Generate a few, pick the best

    The same setup gives slightly different results every run. When a generation is close but not right, rerun it two or three times and pick one. Faster and cheaper than rewriting the prompt.

Reference images

1. Worn shot
Best for accuracy. Length, fit, and sleeves carry over without changes. Faces can be cropped out.
2. Mannequin shot
Good.
3. Cutout / pack shot
Details transfer, but the AI guesses length and fit.
4. Flat-lay on the floor
Not recommended. Distorts the most.
  • Use original files, not screenshots. Low-resolution inputs are the top cause of detail drift.
  • Keep inputs minimal. Two similar garments get mixed together, and any extra face in your references can change the model's face.
  • A worn shot can include other clothes. Point to what you want: "use only the shirt from image 1."
  • For a back view, you need a back-view image. The AI invents whatever it can't see.

Register your products

Register once, and they ride along in every mode.

  1. 1.

    Upload up to 5 photos per product

    Register a worn shot, pack shot, and detail close-up together. Generation passes only the first 3 registered images, so put the worn shot in slots 1 to 3.

  2. 2.

    Add length and fit to the description

    The AI-written description captures pockets and zippers but leaves out length. Add it yourself: "cropped straight denim that shows the ankle."

  3. 3.

    Call it with @

    In any prompt, type @ and pick the product. Its images and description are included in the generation.

The prompt formula

Write in this order: which image supplies what, the final look you want, then what must stay.

Example prompt

Wearing the shirt from image 1 and the denim pants from image 2. Shirt untucked, sleeves down to the wrist. Pants cropped to show the ankle, straight fit. Keep the product details identical to the reference images.

  • Point at images by number: "the top from image 1, the pants from image 2." Numbers follow attach order. Registered products can be called with @ instead.
  • Write sentences, not keyword piles. "Cropped denim that ends above the ankle" beats "denim, cropped, street."
  • One change at a time. Length, pose, and background in a single prompt interfere with each other.
  • Write a fresh prompt per mode, and match garment words to what you attached. A "skirt" prompt over pants references fights itself.
  • Drop phrases like "keep the model identical to the selected model" and model names like "b37." They tangle with internal settings and end up changing the face.
  • Write in Korean. You do not need to translate.

Length & fit vocabulary

Use these instead of numbers like "2 to 3cm longer."

Pants length
shorts / knee-length / capri / ankle-baring crop / full-length over the shoe / dragging on the floor
Top & outer length
cropped above the waist / waist-length / hip-covering / mid-thigh / knee-length
Sleeves
short / elbow-length / three-quarter / wrist-length / slightly covering the back of the hand
Skirts & dresses
mini / below-the-knee midi / ankle-length maxi
Shoes
loafers on bare feet / worn over socks / hem slightly covering the shoe
Fit
slim / regular / relaxed / wide / oversized / straight / tapered / bootcut / high-waisted
Styling
tucked in / untucked / sleeves rolled up / buttons open

The two-step workflow

Base shot first, variations after.

  1. Step 1: The base shot

    Make one shot where the outfit is right: front-facing, full body, neutral pose. Judge only garment accuracy here. Background and pose come in step 2.

  2. Step 2: Vary with Edit Image

    Open the result and use the Edit Image button to change pose, background, angle, or expression. It works from a finished image, so the garment distorts far less.

In Edit Image, state only the change

Change only the pants to full-length, slightly covering the shoe. Keep everything else the same.

In Edit Image, state only the change

Change only the background to an outdoor street. Keep the outfit and model the same.

  • Re-select the model face when editing. It keeps the face much more stable.
  • Editing an edit degrades quality (you'll see a warning). After two or three rounds, start again from the base shot.
  • Resubmitting the exact same request can repeat the same result, so vary the wording a little.

When something goes wrong

Pants or sleeve length comes out different

You used only cutout or pack shots, with no length info. Add a worn shot, name the length in words, correct it with Edit Image, then nudge the rest with the free Retouch brush.

Logos, embroidery, or prints break

The detail is too small in the frame. Add a detail close-up and spell out any text letter-for-letter in the prompt. Final cleanup is faster in Photoshop.

The outfit isn't applied at all

The prompt has no garment description, or the design is one the AI does not recognize. Describe the garment in the prompt, down to material and construction for unusual designs. For partial swaps, use Freeform mode.

The model's face keeps changing

The references include another face, or the prompt has a model-lock phrase. Crop faces out of garment references and delete model-lock phrases and model names. A real photo holds up better than an AI-generated model image.

The face comes out smudged

The source image is low-resolution or heavily filtered. Use high-resolution originals and skip filtered images.

Details appear that don't exist

The prompt mentioned it, or a similar-looking reference slipped in. Remove the mention and any confusable images.

Top and bottom won't both come out right

The request includes too many garments at once. Generate one garment at a time, or add them one by one onto the base shot with Edit Image.

The pose or angle isn't applied

Prompt-only instructions are weak for pose and camera changes. Describe the pose in a sentence or register it via pose upload. For angles, use the presets or multi-angle on a base shot, and attach garment images from that angle.

Colors or textures look off

The input is low-resolution, or lighting phrases are interfering. Use high-resolution originals and remove lighting and filter phrases. Fine color is faster in post.

Results differ every time

Results differ because the AI redraws the image every time. Save one good base shot and run every variation from it.

Still stuck?

Send us the exact images and prompt you used. We'll set it up for your product and send it back.

Sofi prompt tip: accurate garments with fewer redos | Sofi 소피