Restore a scratched family photo in three staged edits

Restore a damaged photo one kind of damage at a time: creases first, scratches second, fading third, each as its own edit on the previous result. Morphic's page on Ideogram 4.5 lists this as a use case, "photo restoration in stages", and gives the reason: clear one defect per pass and the faces do not drift away from the original. Ideogram 4.5 is ideogram/ideogram-v4.5 in the Sume Image API, and a call with a reference image edits that image.

Why separate passes

A single prompt that lists every defect asks the model to decide what is damage and what is texture across the whole frame. Splitting it gives each pass a narrow job and gives you a checkpoint. If pass two damages a face, you discard that result and rerun pass two from the pass-one file, not from the scan.

Write down what each pass is allowed to touch before you start. Creases run across faces, so pass one is the riskiest; if a crease crosses a face, name it in the prompt ("the crease across the left cheek") rather than relying on a general instruction. Scratches and dust are small and scattered, which suits a low-quality pass. Fading is a global tone change, so it is the pass where a higher tier is most likely to pay off.

Keep every intermediate as PNG. The reason is in editing an image five times: keep PNG between passes: each lossy save adds its own artefacts, which look like the drift you are trying to avoid.

The three prompts

Send the scan as the only entry in input_references on the first pass, and the previous output URL on each later pass. Leave aspect_ratio out so each result keeps the shape of its source, as the Image API docs describe for Ideogram 4.5 edits.

• Pass 1: "Remove the fold lines and creases. Keep the faces, clothing and background unchanged."
• Pass 2: "Remove the white scratches and dust specks. Keep the faces and the film grain."
• Pass 3: "Restore the faded contrast and tone to a normal range. Keep the composition identical."

What it costs

Ideogram 4.5 has three qualities in the Sume docs: low, medium and high, with medium the default. The provider list is $0.03, $0.06 and $0.22 per image for all sizes. Sume sells at list times 1.25, so the per-image figures below are the ones to expect; confirm them on the endpoint's pricing line before a batch.

Checks before you hand it back

Compare faces first, since that is where a restoration goes wrong. Crop the face from the scan and from each pass, and set them side by side. A face that looks cleaner but different is a failed pass, however good the scratches look. For a quick number, the pixel diff check tells you how much of the frame each pass changed. Vendor claims about untouched pixels come from Ideogram's own description; they do not replace your own look at the faces.