Lightroom Prompt to Edit runs on Nano Banana: batch edits by API

Adobe's Lightroom v9.6 adds Prompt to Edit, a prompt-driven editing feature that Adobe says is powered by Google's Nano Banana model. If you want the same kind of edit applied to hundreds of photos from a script, you can send each photo and a prompt to Sume's POST /v1/images with a Nano Banana model and input_references, and read back one hosted URL per edit.

Adobe's announcement is the Lightroom v9.6 post on the Adobe Community, read 2026-10-02. Google's model ids are from its image generation page, and Sume's request contract from the Images API docs.

What did Adobe actually announce?

The Adobe post describes Prompt to Edit as a way to describe a desired edit, or pick from preset examples, to do restorations, retouching and transformations without manual adjustment, and says it is driven by Nano Banana. It is part of Lightroom v9.6, updated through the Creative Cloud desktop app. The announcement page I fetched does not state which Nano Banana version it uses, limits, regeneration behaviour or pricing, so none of that is claimed here.

Other coverage says the whole photo is regenerated rather than masked, but that came from a news report, not Adobe's page, so check Adobe's own help article before relying on it.

When is an API better than a Lightroom panel?

Prompt to Edit is a good fit when a photographer is working through one catalogue interactively and wants the result next to their other adjustments. An API fits when the edit is repetitive and the input is a list of files: removing background clutter on 400 listing photos, recolouring a product into five colourways, or fixing the same scan defect across an archive.

Neither replaces the other. Lightroom keeps your raw workflow and non-destructive edits; an API returns a new image URL per call and knows nothing about your catalogue.

How do I batch edits on Sume?

Each call takes a public HTTPS image URL as an input_references entry plus a prompt. Use aspect_ratio: "auto" so each output keeps the shape of its source, since omitting the field is not the same as auto. The endpoint blocks up to 30 seconds and returns 200; if the work runs longer or you send mode: "async" it returns 202 with a job to poll, so check the status code, not the body shape.

For a long list, use async with an Idempotency-Key per file and a webhook for completion. Keep the file name in metadata, which Sume stores on the job and does not send to the provider.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: photo-0001-dust-removal" \
  -d '{
    "model": "nano-banana-2",
    "prompt": "Remove dust and scratches. Keep composition, faces and colour unchanged.",
    "aspect_ratio": "auto",
    "mode": "async",
    "metadata": {"file": "photo-0001.jpg"},
    "input_references": [
      {"type": "image_url", "image_url": {"url": "https://example.com/photo-0001.jpg"}}
    ]
  }'

What will the output be like?

Sume returns a hosted URL for each result, in PNG by default or jpeg and webp through output_format on models that list them. It does not return a RAW file and does not write back into Lightroom, so you re-import the output yourself. Exact pixel sizes are not served on Nano Banana: native aspect ratios are, and exact dimensions are a documented post-step.

Cost is per image on the endpoint's pricing line, and failed or cancelled generations are not billed. For the per-image rates by resolution tier, see our Nano Banana 2 pricing post.

What about content rules and watermarks?

Google's Nano Banana output carries SynthID, per Google's own docs for its image models, and Sume's earlier post on watermarks and C2PA covers what survives on images that come through Sume. If your photos feed a stock agency or a client contract, read their AI disclosure rules before batching.

Photographs of real people raise a second issue: regenerating a face can change it. For portrait work, keep Lightroom's local adjustments for faces and use a prompt edit on backgrounds or objects, then compare outputs side by side.

What should you test first?

Run ten representative photos through the prompt and compare against Lightroom's result on the same files. Whole-image regeneration can change skin texture and fine detail, so check faces and text at 100 percent before you commit a catalogue to the workflow.