A generated in-between is authored by the model, not the editor

The shot

Two current video-generation products document a feature that takes a chosen start image and a chosen end image and generates the motion between them, rather than generating a clip from a single open-ended prompt. Luma Labs describes this for Dream Machine in its Keyframes guide: a start frame "defines the initial look of the video," and an end frame "sets the final point for the animation," with an optional text prompt to guide what happens between them. Google DeepMind's Veo 3.1 product page lists the same idea as "First and last frame," stating that it lets a user "create smooth, artful, and epic transitions between images provided for the first and last frame." Runway's own retirement notice for its Gen-3 Keyframes tool confirms the concept persists on that platform under a new name: the standalone feature was retired alongside Gen-3 Alpha, and its replacement is stated as "available now" as an Animate Frames app.

What the documents show

Luma's documentation is procedural rather than promotional: it names the two inputs, states Dream Machine's default clip length is four seconds, and describes an Extend function that adds an image keyframe to stretch a generation toward a new visual for up to nine more seconds. This is the vendor's own account of a shipping interface, not a demonstrated production result. Veo's page uses more evaluative language -- "smooth, artful, and epic" -- a marketing claim attributable to Google DeepMind, not an independently measured account. Neither page states the in-between motion is authored frame by frame; both describe it as generated from the two endpoints, leaving the interior to the model, not a human decision.

The workflow

Per Luma's guide, an editor uploads or generates a start frame, uploads or selects an end frame, adds an optional prompt describing the transition (Luma's example: "transition from sunset to starry night"), and generates the clip. The approach can be chained through Luma's Extend feature to build a longer sequence from successive keyframes. This is a documented alternative to a single open-ended prompt: the editor supplies the two fixed points an edit needs to land on, and treats the shot's interior as the model's problem to solve.

What the tool does not change

Neither vendor's documentation claims the feature replaces a human decision about which two frames should anchor a transition, or guarantees the generated interior motion will match the pacing, emotional beat, or physical logic an editor had in mind. Choosing the start and end images is still an editorial act; the software's stated contribution is confined to what happens in between.

• Would the transition still make narrative sense if the generated middle motion looked completely different from what was expected?
• Is the chosen end frame specific enough to constrain the model, or vague enough to invite drift?
• Has the vendor documented a duration that actually fits the gap this shot needs to cover?

A first-and-last-frame feature is a genuinely useful editorial constraint -- it turns generation into a fill-in-the-blank problem -- but both vendors' own pages describe the blank as filled by the model, not dictated by the editor.