Adobe’s Doki Treats a Video Script Like an Editing Surface

Most video editors make the timeline the center of the work. Adobe Research’s Doki prototype asks whether a writer might move faster by keeping the story, scene definitions, previews, and revisions in one document. The idea is not simply “generate video from a prompt.” It is to make the script itself an editable control surface.

A document that compiles into shots

In Doki, a creator defines reusable characters, settings, and visual styles, then writes scene descriptions in a document. The system can turn paragraphs into video shots, show previews inline, and let the creator refine scenes through text or an assistant. Audio and shot organization live alongside the narrative rather than in separate tools. The paper frames writing as the main interaction model, with previews and controls embedded where the creator is already thinking about a scene. ^[2]

That structure aims at a real friction point. A conventional editor gives precise control, but it asks users to learn tracks, clips, keyframes, and specialized menus. A text document is familiar and makes story order visible. When a sentence changes, the user can inspect the corresponding shot in context.

The tradeoff is that text does not specify every visual decision. A phrase like “the room feels tense” leaves camera, pacing, actor movement, and continuity underdetermined. A creator still needs ways to constrain what the model may change, preserve important details across scenes, and compare revisions without losing a good take. The prototype’s value depends on whether it exposes those controls clearly enough to support intentional editing.

What the study says, and what it does not

The paper reports a week-long diary study with ten participants of varied video experience, who produced 46 videos. Participants described faster movement from idea to draft, better coherence through parameterized assets, and a clearer understanding of narrative flow. They also surfaced limits involving predictability, precise control, and temporal expressivity. These findings are early interaction evidence, not a controlled demonstration that text-led editing is faster or better than established editors for professional work. ^[2]

Adobe’s November 5 announcement says the work received a Best Paper Honorable Mention at UIST 2026 and describes Doki as experimental. That recognition signals peer interest in the interaction design; it does not mean the prototype is a shipping product or that generative video has solved continuity and revision. ^{[1] [3]}

A useful design test for creative AI

For any text-first creative tool, ask whether the document preserves authorship or hides consequential choices. Can a creator see which description produced a shot? Can they lock a character, camera, or asset while changing the timing? Can they reject an edit without regenerating unrelated scenes? Does the tool preserve alternate versions and identify generated material?

A practical evaluation should compare equivalent tasks, not just ask whether a prototype feels delightful. Give novice and experienced editors the same short brief; measure time to a usable draft, number of revisions, continuity errors, and how much control participants feel they retain. Include a task requiring a precise local change, because broad generation is easier than revising one element while holding everything else constant.

Doki is a thoughtful exploration of a different authoring metaphor. Its strongest promise is a tighter connection between narrative intent and visible output. The hard product question is whether that connection can coexist with the fine control, predictable changes, and temporal editing that creators need once they move past the first draft.

For another example of AI changing a creator workflow, see our coverage of shorter generative video workflows and their limits.