Luma Photon image reframe: 10 MB source, 2432x1024 max vs Sume

Luma's Reframe API takes an image URL, a target aspect ratio and a model, and generates the missing area so the picture fits the new frame; for images the models are photon-1 and photon-flash-1, with a 10 MB source limit and a 2432x1024 maximum output. Sume has no reframe endpoint, so the closest route is an image edit that asks for the new ratio and describes what should fill the extra space.

The numbers below are from Luma's Reframe page; the Sume side comes from the Image API docs.

What does Luma's image reframe accept?

The page lists four reframe models: ray-2 (10 seconds, 100 MB) and ray-flash-2 (30 seconds, 100 MB) for video, and photon-1 and photon-flash-1 for images, each with a 10 MB limit. Supported ratios are 1:1, 4:3, 3:4, 16:9, 9:16, 21:9 and 9:21. Maximum output is 2048x1152 at 16:9 for video and 2432x1024 at 21:9 for images.

An optional prompt guides what goes into the newly generated areas. Advanced use takes grid-position parameters in pixels: x_start, x_end, y_start, y_end, grid_position_x and grid_position_y, which place the original inside the new canvas. Costs follow the underlying model's pricing, and the page gives no separate reframe fee.

What is Sume's closest path?

Sume's Image API takes a prompt and input_references, and an aspect_ratio from a normalized list that includes 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 4:5, 5:4, 9:21 and 21:9, or auto. Send the original as a reference, set the target ratio, and write what should appear in the new area. Models only accept the ratios their catalog entry lists, so check supported_parameters first.

Be clear about the difference. Luma's call is built to keep the original pixels and fill around them, with pixel-level placement. A Sume edit regenerates the image under the new ratio, so the original may change slightly, and there is no placement field. If the original must stay untouched, crop or pad outside the model and use an edit only for the fill.

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":"Extend this photo to a wide banner. Keep the product exactly as is; continue the wooden table and soft window light on both sides.","input_references":[{"type":"image_url","image_url":{"url":"https://example.com/product.jpg"}}],"aspect_ratio":"21:9"}'

How do you keep the original intact?

Use a mask where the model supports one. ChatGPT Image 2.5 on Sume accepts an optional mask_url alongside references, which lets you mark only the area that may change. Without a mask, assume the model can redraw anything, and compare the result against the original before you ship.

• Check the output size against the placement you need; custom sizes need both edges in the model's allowed multiples.
• For a vertical-to-horizontal change, ask for the fill to continue lighting and surfaces, and avoid asking for new objects.
• For exact pixel sizes such as 1080x1350, Sume documents a post-step through target_pixels on the job, not a native output size on every model.
• Keep the source and result job ids together for review.

When should you use which?

Use Luma's Reframe when the original has to stay pixel-faithful, the source is under 10 MB and you want pixel-level placement on a 21:9 or 9:21 canvas. Use a Sume image edit when reframing is one step in a larger flow, with other images, video and audio on the same wallet.

For video, Luma's reframe is limited to 10 seconds on ray-2 and 30 seconds on ray-flash-2. On Sume, change a clip's shape by generating at the target aspect ratio or by cropping on the Timeline. Read the Image API page for the descriptors and the reference rules before you pin a model.