Adobe's Doki turns a script into an AI film in one document

Doki, an experimental tool from Adobe Research, turns a written script into a finished AI film inside a single text document. You write the story in plain prose, name your characters, settings and props as reusable tags, and Doki generates the shots in place, between your paragraphs, with a small preview sitting inline in the text. Adobe Research published the work and presented it at UIST 2026, where the paper won a Best Paper Honorable Mention. It is a research prototype. Adobe has announced no release date and has not said whether Doki becomes a product.

The reason to care is the workflow Doki replaces. Most AI video tools generate one clip at a time from one prompt, and you carry the results into a separate editor to assemble a story. Doki asks a simpler question: if a film can be generated from text, why not make the text the whole workspace?

What is Doki?

Doki is a text editor with a video engine behind it. The paper, "A Text-Native Interface for Generative Video Authoring," was posted to arXiv on March 10, 2026 by Xingyu Bruce Liu, Mira Dontcheva and Dingzeyu Li, all of Adobe Research. Adobe describes Doki as an experimental tool it codenamed Project Doki, built around the idea that everyone learns to write a story in school but almost nobody learns to operate a video editor.

The interface is deliberately bare. A blank document, a slash menu for inserting things, and a handful of buttons. There is no timeline, no track stack, no bin of thumbnails. Writing is the primary interaction, and everything else lives in the text.

How does writing become video?

Doki maps the structure of a document onto the structure of a film. The whole document is the video. Each paragraph is a sequence. Each sentence is a shot.

To make a shot, you type a slash command and Doki drops a preview node into the sentence. The words that follow become the shot's prompt. The node runs a two-stage render: first a still image, then the video, with the still used as the first frame. By default the prototype generated previews with Flux Kontext Pro and video with Google's Veo 3 Fast, though the paper says Doki supports a range of models chosen in its settings.

That staged pipeline is an economic decision as much as a creative one. The paper puts a Veo 3 video clip at $3.20 and an Imagen 4 image at $0.04, so checking the visual direction on a still before committing to video saves the expensive step for shots you already like. A small status dot on each node tells you what is happening: breathing yellow while the image renders, green when the clip is ready, red when an edit upstream has made the shot stale.

Multiple shots in one paragraph chain together. Later shots pull the previous shot's image as a reference, so a close-up of the same character in the same scene inherits the look without you re-describing it. Cross a paragraph break and the inheritance resets.

Audio goes in square brackets inside the prose, tagged as speech, music or sound effects, and is passed to models that generate synchronized sound. Two AI agents sit alongside the writing: a sidebar agent that can rewrite whole stretches of the document and an inline agent that appears when you select text and can turn a phrase into a reusable definition.

How do the tags keep a character consistent?

This is the part that separates Doki from a stack of clip generators. Instead of re-describing a character in every prompt, you define it once and reference it.

Definitions come in two kinds. @Mentions name the story's nouns: a Character, a Scene, an Ingredient (a prop), or a Frame. #Hashtags name its qualities: Style, Camera, Composition, Ambience. A character might be @Pandi = a little red panda wearing a suit and sunglasses; a look might be #filmnoir or #claymation. You can attach a preview image to a definition, and Doki then uses that image as the reference for every shot that mentions it.

The definitions propagate. Rename @Pandi and every reference updates. Change its description and Doki flags each shot that uses it as outdated, so you know what to regenerate. A definition named all applies to the entire document; a definition placed on a heading scopes to that section. When a shot renders, Doki resolves every tag into a structured prompt, has a language model rewrite that into fluent prose for the image model, and assembles up to four visual references, including the definition's own image and the previous shot in the sequence. The tag system is how a five-second clip becomes a coherent scene rather than a lucky draw.

What did the week-long study find?

Adobe Research ran a week-long diary study with 10 participants, six recruited from outside the company and four from within Adobe. Their roles ranged from filmmaker and animator to graphic designer and program manager. They used Doki independently for five days and submitted 46 videos in total.

The videos averaged 67 seconds, with most falling between 30 and 90 seconds. Eight participants produced original narratives; the rest split across instructional video, advertising, music and experimental work. On the System Usability Scale, a standard 100-point questionnaire, the group rated Doki 81.2, which the paper calls "Excellent." Eight of the ten said Doki sped up the move from a rough idea to deliverable content. One participant, P5, put a number on it: "So on average it will be 15 minutes to create about one minute or longer video."

The praise centered on focus. P5 also said, "I think the interface is perfect…the whole interface is very simple and clean." Participants described themselves as directors, with P1 saying, "Because I get to decide what every item looks like, I feel like this is my creation."

The failures were specific and frequent. Nine of the ten said Doki's output often failed to match their prompts, forcing repeated regeneration. P1 described persistent artifacts: "A lot of videos include nonsensical text that I couldn't remove, even when I explicitly prompted it." P10, trying to adapt a personal storybook, summed up the ceiling: "Maybe just achieve 80% of your imagination."

How does Doki compare with what Adobe already ships?

Doki is a research prototype, and it sits apart from the generative video Adobe sells today. Adobe's Firefly video generator turns a text prompt or an image into a clip inside a multi-model workspace, pulling in the Firefly Video Model alongside partners like Google's Veo 3.1, Luma's Ray3, Kling 3.0 and Runway. In Premiere, Generative Extend, powered by the Firefly Video Model, adds up to two seconds of video or ten seconds of audio to a clip. Both are shipping products with prices and plans. Doki is neither.

It is also separate from Project Live Set, the Adobe research preview shown at the MAX 2026 Sneaks panel on November 11, which streams video as it is generated so a creator can redirect a scene while it plays. SLOP TV covered that one separately. Doki was presented a week earlier at UIST, an academic conference, not on a product stage. The two come from the same company and both push at generative video, but they are different projects with different authors.

What are the limits, and will Doki ship?

The study's own caveats matter as much as its results. Ten people is a small sample, and four of them worked at Adobe. The paper does not compare Doki against a conventional editor or a rival tool, because, as the authors write, "there is no established point of comparison for generative video authoring." The participants were testing a prototype, not choosing between products, and external participants were paid $30 an hour.

The design has structural limits too. A linear document is good at narrative and bad at timing. Participants struggled to specify overlapping audio, music that runs across scenes, and transitions that bridge shots, what the paper calls the "temporal expressivity" limits of a document. Precise framing was hard: P9 said, "Doki is not good at when you want specific frame compositions." The authors also note that the study left open how much control Doki gives over the fine detail a professional edit demands.

Whether any of it reaches a product is unsettled. Adobe has moved research projects into shipping tools before, and the paper's interaction ideas are independent of Doki itself. For now the paper is the deliverable.

Doki is a research prototype with no release date. The paper is free to read on arXiv, and Adobe Research has opened no waitlist or preview for it.