Remove dust, lint and wrinkles from a product photo with an AI edit

To remove dust, lint or wrinkles from a product photo with an API, send the photo as the only input_references entry to POST /v1/images with model: "ideogram/ideogram-v4.5" and a prompt that names what to clean and what to keep. On Sume the call is priced per image by quality: $0.0375 at low, $0.075 at medium and $0.275 at high, the same at every size. Ideogram describes 4.5 as its most precise edit model and says it reduces drift across multi-turn edits, which is the property a cleanup pass needs.

This post explains how to word the prompt, which quality to pick, and how to check that the label and the product edges did not change.

How Ideogram 4.5 edits on Sume

Without references ideogram/ideogram-v4.5 generates from text. With references it edits: the first image is the one that is edited and up to four more are references, so five in total. An edit without aspect_ratio keeps the shape of the source image. There is no mask_url for this model, and output_format is chosen by the provider and not accepted in the request, so send neither.

That makes this a prompt-only edit. Name the defects, name the parts that must not move, and ask for nothing else.

A prompt that stays inside the lines

Cleanup prompts fail when they invite redesign. Avoid words like improve or enhance. Say what you see: three lint fibres on the left sleeve, creases across the chest, dust specks on the bottle shoulder. Then add a preserve list: logo, stitching, label text, colour, framing and shadows stay exactly as they are.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ideogram/ideogram-v4.5",
    "prompt": "Remove the lint fibres on the sleeve and smooth the creases across the chest. Keep the logo, stitching, colour, framing and shadows exactly as they are.",
    "input_references": [
      {"type": "image_url", "image_url": {"url": "https://example.com/shirt.jpg"}}
    ],
    "quality": "medium"
  }'

Cost for a catalogue

Sume prices Ideogram 4.5 at the provider list multiplied by 1.25, which gives these per-image amounts. The list prices ($0.03, $0.06, $0.22) are on the Sume Image API docs page.

Picking the quality

Dust and lint are small defects, so start at low on ten photos and look at them at full size. Move to medium if edges soften or text on the product label changes. Save high for hero images. Because completed jobs are billed even if you reject the result, a ten-photo trial at each tier costs $0.375, $0.75 and $2.75, which is cheaper than finding a problem on photo 150.

What cleanup can and cannot fix

Cleanup edits are good at small, isolated defects: lint on dark fabric, dust on glass, a crease across flat cloth, a fingerprint on a bottle. They are weaker at defects that cross something structural, such as a wrinkle that runs through printed text or a scratch on a metal edge that defines the silhouette. In those cases the model has to guess what the clean surface looked like, and the guess can change the product.

For apparel, steam or press the item before the shoot where you can; an edit should remove the last ten percent of the problem, not the first ninety. For glossy packaging, shoot with a polarising filter to cut reflections that the model might read as dirt. And never use an AI cleanup to hide a real defect on the item itself, such as a tear or a stain: the customer will receive the real product.

Verify before you publish

Download the source and the edit at the same size and diff them with Pillow or NumPy. Real cleanup shows as a handful of small blobs. If the whole label lights up in the diff, the model re-rendered the text, and for products where the label is legally meaningful you should keep the original pixels for that region and composite them back in your own code.