Luma exposes Ray3.2 video-edit controls through its API

What changed

Luma announced Ray3.2 and API access on 9 June 2026, describing a set of controls for teams that want to revise footage and carry it into post-production. The company emphasized guided frames, performance handling and finishing formats. These were supplier claims at launch. The current API documentation gives a more useful starting point for a studio deciding how to test the service: it separates generating a new video, editing source footage and changing the frame shape into different request types.

For an edit, the developer supplies source video and a prompt. Guide images can anchor the desired look at selected positions in the source. Luma also documents controls for preserving aspects of the source, including pose and depth. A developer can begin with automatic controls or configure the manual controls, subject to the API's validation rules. The point for the creative team is practical: an engineer can expose the decisions that matter to a particular job inside a studio's own tool. That last use is our interpretation, not evidence of a deployed client integration.

Why it matters

A brand might want to explore a different treatment of a set while keeping a performer's blocking recognizable. Our suggested trial would give the reviewer a source frame, a target treatment and the moving result together. Ask the reviewer to distinguish an accepted creative change from an accidental alteration to the product or performance. A compelling sample frame alone cannot answer that question. It also cannot show whether the result survives a cut to an untouched shot elsewhere in the sequence.

What to do now

Read the specification for the interface you plan to use. The June announcement and the current API reference describe different keyframe limits. The API's edit guide now documents up to 64 guide images and an 18-second limit for source videos supplied by URL or inline data. Those are scoped API statements observed on 5 September, not universal promises about the consumer app.

Finishing also has its own settings. Luma's model reference describes HDR-encoded video and an EXR export option, with the EXR using ACES2065-1. That gives a colorist a documented format to inspect during a pipeline test. It does not demonstrate that a generated shot will match a camera original, preserve a textile color or need less correction. We recommend testing the actual export through the team's editing and grading applications before putting it into a delivery schedule.

Luma's editing guide rejects HDR below 720p, while its pricing table lists lower-resolution HDR edits; confirm the supported setting before budgeting a job.

Cost depends on the operation and output choices. The API pricing page distinguishes editing from generation and changes rates by resolution, duration and dynamic range. It also says that pay-as-you-go capacity is shared and has no latency service-level agreement. An agency quoting a time-sensitive revision should confirm the current rate and service terms for its account. The Human Input has not measured a queue time or calculated a client-job saving.

The first implementation need not become a studio-wide automation. Our recommendation is to choose a cleared test clip, restrict the operator's choices to the settings needed for that experiment, and keep a record of the input and the returned file. A person should inspect a completed output before anyone sends it to a client. A request accepted by the service is not the same thing as a finished video, and a finished video is not the same thing as an approved shot.

What to watch next

For a creative producer, the useful outcome would be a repeatable decision: which kinds of changes the team can accept, which details require a manual pass, and which requests should stay with the original production method. For the developer, the task is to preserve enough information to repeat the test. Neither decision requires assuming that a model's launch description predicts its result on an unreleased fashion campaign.