Luma uni-1: 9 references for images, 8 for edits, and Sume limits

Luma uni-1 accepts up to 9 image_ref images for text-to-image and up to 8 for image_edit, because the source image takes one of the 9 slots. On Sume there is no fixed number: each model carries its own reference ceiling in the catalog, and some models take none.

Luma facts are from its Models and Pricing page; Sume facts are from Image models, read 2026-10-01.

Where does the 9-versus-8 rule come from?

The Luma Models page says reference-guided generation accepts up to 9 reference images via image_ref for text-to-image, or up to 8 for image editing, "where the source image occupies its own slot". The pricing page repeats it: image_edit supports up to 8 references and image up to 9.

What are the reference limits on Sume?

They are per model. In the catalog code, edit-capable models default to a max_images of 4 and the Grok image entry sets 1. The docs say ChatGPT Image 2.5 supports up to 16 references, and a model whose input_references descriptor is {"min": 0, "max": 0} is text-to-image only and rejects references.

Reference URLs must be public HTTPS; localhost, private-network and non-HTTPS URLs are rejected before submission.

Does a Sume edit use up a reference slot?

The docs describe an edit as input_references on an image request and do not describe a separate source slot. Treat the descriptor's max as the total number of images you can send, and test one edit at that number before you rely on it.

How do I plan around the difference?

Keep your reference list sorted by importance and trim to the target model's max at call time. The Grok image reference post shows the same pattern for a lower ceiling.