AI handwritten note in an image: write it, then read it back

To get a handwritten note inside an AI image, put the exact words in quotes in the prompt, ask for a handwriting style, and generate on a text-capable model such as openai/gpt-image-2.5 at quality: high. Then read the result back before you use it. Handwriting is drawn pixels, so a wrong letter needs a new edit, not a text change.

Prompt the words and the hand

Keep the note short. Quote the exact text, name the surface (lined paper, a sticky note, the back of a receipt), the pen (blue ballpoint, black marker) and the framing. Say that the writing must be legible and must not add other words. Short lines of five to ten words are easier to hold than a paragraph.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"openai/gpt-image-2.5","prompt":"Top-down photo of a yellow sticky note on a desk. Handwritten in blue ballpoint pen, legible, exactly: \"Back at 3. Coffee is on me.\" No other words.","quality":"high","aspect_ratio":"1:1"}'

Pick the quality on purpose

Sume's GPT Image 2.5 rows accept quality: auto|low|medium|high|xhigh|max and default to high when you omit it. Dense or small writing is where a higher setting earns its cost. See the GPT Image 2.5 API guide for the sizes and limits.

Read it back

Treat every note as unverified until a person or an OCR step has compared it with the string you asked for. Look for swapped letters, doubled letters and invented words, and check the last line first because that is where drift tends to show. If the note is for a printed product, compare it at the real print size.

Fix one word, keep the rest

When one word is wrong, use a masked edit rather than a full re-roll. GPT Image 2.5 takes mask_url and up to 16 references. Mask the line, quote the corrected text, and say to keep everything outside the mask identical. The pattern is in mask_url edits on GPT Image 2.5.

When the words matter more than the look

If the exact wording is legal, financial or medical, do not rely on a generated hand. Generate the paper and pen texture, then set the words with a handwriting font in your layout tool, where the text is data you can proofread.

Settings that matter for text

These are the fields on the GPT Image 2.5 rows that affect a text-heavy image.

Ask for more than one take

Set n to 4 and choose the best-written result. The endpoint price is per image, so four takes cost four times one take. For a short note this is usually cheaper than a long chain of corrections.