Sketch to image AI: turn a drawing into a finished picture

To turn a sketch into an image with AI, scan or photograph the drawing, send it to an image model as a reference, and describe the finished picture: what it shows, the materials, the colors, the light, and "keep the layout of the sketch". The model uses the drawing as a guide for a new image, so check how closely the result follows your lines. On Sume, the sketch's public HTTPS URL goes in input_references on POST /v1/images.

Request fields and limits come from Sume's Image API docs and the model catalog that GET /v1/images/models serves, read on 2026-09-27. Going the other way, from a photo to an artwork, is covered in photo to painting AI.

How do I turn a sketch into an image with the Sume API?

Send the drawing as a reference, describe the finished result, and ask for several versions at once. The request below asks Nano Banana Pro for four renders of a pencil sketch.

• Scan or photograph the sketch flat and evenly lit, then host it at a public HTTPS URL. Localhost, private-network, and non-HTTPS URLs are rejected before submission.
• On models that list it, the docs recommend aspect_ratio: "auto" so the result matches the drawing's shape. Omitting the field is not the same as sending auto.
• n asks for several versions in one call. Image models are metered per image, so each version adds to the cost.

curl -X POST "https://api.sume.com/v1/images" \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/nano-banana-pro",
    "prompt": "Turn this pencil sketch into a finished product render: a brushed aluminum desk lamp, soft studio light, plain white background. Keep the shapes and proportions of the drawing.",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://example.com/sketch.png" } }
    ],
    "aspect_ratio": "auto",
    "n": 4
  }'

How do I turn a sketch into a realistic photo?

Describe a photograph, not a drawing. The prompt is the only place to say how finished and how real the result should look:

• Say "a photo of" and name the setting, such as a kitchen counter or a city street at dusk.
• Name a real material for every part you drew: oak, brushed steel, matte ceramic.
• Describe the light and the camera angle the sketch implies, such as eye level or from above.
• Say what must not change: the layout, the proportions, the number of objects.

Can I control how closely the image follows my lines?

Not with a setting. Today Sume's image request has no line-following, control-strength, or sketch-mode field, so the prompt is the only lever: say which parts must stay where you drew them.

If a result drifts, save the version closest to your drawing, host it at your own public HTTPS URL, and send it back as the reference with a narrower prompt. Every call is a fresh generation: no model serves seed, and a request that sends one returns 400 unsupported_parameter.

Which models keep the sketch's shape?

Models that edit from references and list auto can match the frame of your drawing. Seedream 5.0 Lite and Seedream 4.5 edit too, but they do not list auto, so give them a ratio from their own lists; a model refuses a value its catalog does not list. Today Seedream 4.5 returns one image per call when you send a sketch, whatever n asks for, so send one call per version there.

What are the limits of sketch-to-image AI?

• It does not trace. The output is a new image rather than your lines filled in, so proportions and small details can shift. Do not rely on it where dimensions must be exact.
• It returns raster files: png, jpeg, or webp. svg is in the request schema, but no model lists it today.
• Text-to-image-only models reject the sketch. Their input_references range is {"min": 0, "max": 0}.
• Results come back as Sume-hosted, signed URLs in data[].url, so save the images you keep.