Edit the words on a neon sign photo with Ideogram 4.5 on Sume

To change the words on a neon sign in a photo, send the photo as the first input_references entry to POST /v1/images with model: "ideogram/ideogram-v4.5" and a prompt that names the old text, the new text, and the sign's style. Sume's Image API docs say that with references this model edits the first image, so no mask is needed for a plain text swap.

Neon is a good test of an edit model because the letters are not flat ink. They are glowing tubes with a halo that falls on the wall around them. Ideogram's launch post (read 2026-10-05) calls 4.5 "the most precise edit model". A partner page, Morphic's Ideogram 4.5 page (a partner, not the vendor), says it edits stylized text in place and copies unchanged pixels from the source. Treat both as claims to test on your own photo.

The request

The call below is the whole workflow. Quote the exact old and new strings, and say what must stay: tube color, glow, the wall, the camera angle.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: neon-edit-001" \
  -d '{
    "model": "ideogram/ideogram-v4.5",
    "prompt": "Change the neon sign text from \"OPEN LATE\" to \"OPEN 24H\". Keep the pink tube color, the glow on the brick wall and everything else identical.",
    "quality": "medium",
    "input_references": [
      {"type": "image_url", "image_url": {"url": "https://example.com/neon.jpg"}}
    ]
  }'

Photo conditions that help

The input photo matters as much as the prompt. A model can only preserve what it can see, so give it a clear view of the sign.

Shoot or pick a frame where the sign is sharp and not blown out. Neon tubes photographed at a long exposure often clip to white at the center, and a white core makes the letter shapes hard to read, both for you and for the model. If you control the shoot, expose for the tube color and let the wall go darker. Avoid heavy motion blur on the lettering, and avoid extreme angles where the letters overlap in perspective.

The partner page above lists a 25 MB maximum file for the model on its side. Phone photos are well under that, but a raw export or a large scan might not be. Resize first if you are near the limit. Sume's own rule is simpler: the reference URL must be public HTTPS, and the docs say Sume rejects localhost, private-network and non-HTTPS URLs before it submits the job.

Expect to iterate on the prompt, not on the image. If the letters come back in a different tube style, add the style words to the prompt and run again from the original photo, not from the bad result. Starting again from the original keeps errors from piling up. If only one letter is off, name that letter's position ("the third letter") and ask for that fix in a fresh call.

Keep a note of each prompt and the job id it produced. The Idempotency-Key header in the example is also a handy label: use a different value per prompt version so a retry of one version cannot be mistaken for a new version.

What to put in the prompt

Short, literal prompts work better for text edits than descriptive ones. These are habits, not vendor guarantees:

• Quote both strings in straight quotes, with the case you want.
• Name the style you want kept ("same pink tube lettering", "same glow").
• Say what must not move: the wall, the frame, other signs in the shot.
• Keep the new text near the length of the old text. A sign with room for 9 characters may not hold 20.
• Leave aspect_ratio out. The docs say an edit without it keeps the shape of the source image.

Check the result before you use it

A model that redraws glowing letters can still misspell a word or alter a second sign in the frame. Open the result at full size and read each letter. Compare the corners and the wall against the original: if the halo or the bricks changed, run the edit again with the unchanged parts listed in the prompt.

Start at quality: low to see whether the wording lands, then raise the tier for the keeper. The docs list Fal list prices of $0.03, $0.06 and $0.22 per image for low, medium and high, and say Sume's pricing lines already include its margin. Read the live line from GET /v1/images/models/ideogram/ideogram-v4.5/endpoints before you budget a batch.

If a request returns 200, the image URLs are in data[].url. If it returns 202, poll the job as described in Jobs and results.

Limits worth knowing

The docs list mask_url for ChatGPT Image 2.5 edits, not for Ideogram 4.5. The partner page above says Ideogram supports masks on its own side, but on Sume you steer the region with words. For a sign that fills only part of a busy photo, name where it is ("the sign above the door, left of the menu board").