Change a photo's camera angle with AI: an orbit, tilt, lens prompt

To turn a photo to a new viewpoint, write the move as a photographer's brief: orbit degrees, tilt, lens. Sume's angle tool does this on Nano Banana Pro.

5 min readSume
All posts

To change the camera angle of an existing photo with AI, describe the move as a photographer would: orbit the camera a number of degrees around the subject, raise or lower it, and name a lens. Send that prompt with the photo as a reference. Sume's own moodboard angle tool builds exactly this prompt and sends it to Nano Banana Pro.

Why bare numbers fail

A prompt such as Yaw: -44 degrees is metadata, not direction. Sume's moodboard code notes that image-edit models ignore lines like that and return something close to the source. So the tool turns the slider values into sentences: a three-quarter view that shows the left side, a full side profile, a rear three-quarter view, or a view from directly behind, depending on how far the orbit goes.

The prompt structure

The angle prompt has the same parts every time. You can copy the shape and change the numbers.

  • Open with the intent: re-photograph the exact same subject from a new camera position.
  • Lock what must stay: subject, identity, materials, colors, styling and background mood.
  • State the orbit in words and degrees, then the tilt (high angle looking down, or low angle looking up).
  • Set the zoom (closer or farther, as a percentage) and the lens: about 24mm for wide with converging lines, about 50mm for natural perspective.
  • Say that the sides now visible must be rendered consistent with the visible design.
  • For a large move, add that the composition must clearly differ from the input and that the original viewpoint must not be returned.
  • Close with a ban: no new objects, text or location.

Send it through the Images API

Nano Banana Pro is google/nano-banana-pro in the catalog. It accepts input_references, and aspect_ratio: "auto" keeps the shape of the source. Reference URLs must be public HTTPS.

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":"Re-photograph the exact same subject from a new camera position. Keep the same subject, materials, colors and background mood. Orbit the camera 45 degrees to the left around the subject so the result is a three-quarter view that clearly shows the subject's left side. Raise the camera and tilt it down 20 degrees. Shoot on a normal lens (about 50mm). Do not add new objects, text, or a different location.","aspect_ratio":"auto","input_references":[{"type":"image_url","image_url":{"url":"https://example.com/product.png"}}]}'

Check the result against the source

Compare the two images side by side. If the viewpoint barely moved, raise the orbit number or add the sentence that the original viewpoint must not be returned. If the product changed (a logo, a strap, a stitch), lower the move and re-run from the original photo, not from the last output. A 30 to 45 degree orbit is the safest first try because the model has the most visible information to work from. For a full back view the model has to invent the hidden side, so treat that result as a concept, not a product record.

Related fixes

If only the light is wrong, use a relight pass instead of a camera move. See relight a photo with AI. To keep the output shape, read Nano Banana Pro aspect ratio auto.

What to try first

Change one thing per pass. Moves that mix a large orbit, a steep tilt and a lens swap give the model several reasons to redraw the product. Start with the orbit alone, check it, then add tilt.

Suggested first passes (a starting plan, not a guarantee)
PassOrbitTiltLens
130 to 45 degrees0normal, about 50mm
2same15 to 25 degrees downnormal
3samesamewide, about 24mm

Keep the source clean

Use a sharp, evenly lit original with the whole subject in frame. If part of the product is cropped, the model must invent it when the camera moves. Re-run every variation from the original photo so errors do not stack from one output to the next.

Related posts

More in Use cases

All Use cases posts

Written by Sume