SAGE Matches Pro Directors on Storyboards and Opens the PROSE Dataset

A storyboard system scored 77.8 on an expert validated rubric. Professional directors scored 77.1 on the same rubric, across the same 18 test episodes in three genres.

That result comes from SAGE, described in a paper posted to arXiv on August 18, 2026 by a team of eight researchers. The more useful part of the release is the dataset that came with it.

PROSE is the first public dataset pairing screenplays with storyboards written by professional directors. It holds 68 episodes across three productions, and it is released under CC BY 4.0.

The Numbers From a Live Production

SAGE was not tested only on a benchmark. The team deployed it for 14 days inside a working short drama pipeline the paper calls Virtual Film Studio.

Measure	Result
Rubric score, SAGE	77.8
Rubric score, professional directors	77.1
Test episodes	18 across three genres
Deployment period	14 days
Narrative group outputs produced	1,344
Accepted without substantive edits	87.2 percent
Reduction in authoring time per episode	Over 83 percent

The acceptance rate is the number that matters. 87.2 percent of 1,344 outputs went through with no substantive edit, judged by the production team rather than by the researchers.

Read the rubric scores carefully. 77.8 against 77.1 is parity rather than a win. One rubric across 18 episodes is a narrow test, and the paper frames the result as reaching expert level rather than passing it.

What PROSE Actually Contains

The dataset is small and specific. Three productions, 68 episodes, each one a screenplay in Markdown paired with a director authored storyboard in CSV.

Production	Episodes
beyond-the-wall	20
his-toyboy	23
my-cure	25
Total	68

Every row in a storyboard file is one shot. The columns are the interesting part, because they show what a working director actually records.

Column	What it holds
Episode ID	Which episode the shot belongs to
Scene	The scene or setting
Shot description	The visual content of the shot
Shot size	Close up, wide and the rest
Camera angle	Where the camera sits
Camera movement	How the camera moves
Characters	Who appears in the shot
Dialogue	The lines spoken in the shot

Eight fields. That is the whole vocabulary of a professional shot plan for this format, and it is now public for anyone who wants to build against it.

Two practical notes for anyone loading the files. The column headers in the source CSVs are in Chinese, and the files carry a byte order mark, so read them with UTF-8 BOM handling.

How SAGE Learns Directing Rules

The method is easier to follow than the title suggests. SAGE compares each training screenplay against the storyboard a director produced from it, and derives rules that do not depend on the content of any one episode.

During generation the model records which rules each narrative group used. When feedback arrives, the system can trace it to the specific rule behind the decision and update that rule alone. Most approaches update the whole prompt or the whole model, and lose the attribution.

Evolved rules are then grouped into scenario packages with a routing index. Each narrative group retrieves only the small set of rules that fit its situation, which keeps the context window usable without a human picking rules for every scene.

The paper names three problems this addresses. Knowledge acquisition, because the craft is implicit. Knowledge refinement, because authored rules are rarely tested against outcomes. Knowledge injection, because the full rule set does not fit in a usable context.

What Is Released and What Is Not

Be precise here. As of August 19, 2026 the GitHub repository holds a LICENSE file, a README and the data directory. It carries the dataset.

The SAGE framework code is not in that repository. The paper describes the method and releases PROSE alongside it, so anyone rebuilding SAGE is working from the paper rather than from source.

The dataset licence is CC BY 4.0. That permits sharing and adaptation, including commercial use, with attribution and a link to the licence. For a research dataset drawn from three real productions, that is a permissive choice.

Where This Sits Against Prior Work

Storyboarding and shot planning are separate problems, and this blog has covered both.

System	What it plans	When it runs
SAGE	Shot lists from a screenplay	Before the shoot
ShotPlan	Shot timing inside generated video	At generation time
HoloCine	Coherent multi shot narratives	At generation time

SAGE is the pre production entry in that list. It produces a document a human reads, and the shot list it writes could go to a camera crew or into a generation pipeline without changing anything.

That distinction matters for the numbers. A generation system is judged on the footage. SAGE is judged on whether a production team accepted the plan, which is why the 87.2 percent figure carries more weight than a benchmark score.

Why the Pre Production Number Is the One to Watch

McKinsey argued in January 2026 that AI reshapes pre production before it reaches the camera. The McKinsey pre production report put scheduling, breakdowns and shot planning ahead of generation in the adoption order.

SAGE is the first result with a live production acceptance rate attached to that claim. Over 83 percent less authoring time per episode is a schedule change rather than a quality claim, and schedule changes are what actually alter how a production is staffed.

Storyboarding with AI already has a high profile example. Martin Scorsese joined an image model company as an adviser, shown storyboarding a scene, and the Scorsese advisory role covered what that looked like in practice.

The difference is the direction of travel. An image model draws the frames. SAGE writes the plan that says which frames to draw.

What This Means for Filmmakers

Nothing here ships as a product you can open today. The value is the dataset and the finding.

If you build tools, PROSE gives you 68 aligned screenplay and storyboard pairs under a permissive licence, which did not exist in public before. The eight column schema is a ready made target format for any shot list feature.

If you direct, the finding is that a rule based system reached parity on a short drama rubric and cut authoring time by more than 83 percent. Short form drama is the format where that lands first, because the episode volume makes the saving compound.

The practical step is downstream. Once a shot list exists, the video workspace turns individual shots into footage, and the image workspace covers concept frames and environment plates while the plan is still being written.

---

Sources

arXiv: "SAGE: Self-Evolving Storyboard Skills via Attribution-Guided Rule Evolution" (Maolin Ran, Xiaoyang Lu, Jiaqi Liu, Jian Wang, Weiwen Liu, Jianghao Lin, Yong Yu, Weinan Zhang. August 18, 2026)

arXiv (PDF): Full paper, 11 pages, 9 figures, 4 tables

GitHub: creDreams/PROSE (the dataset, CC BY 4.0)

GitHub: PROSE data directory (screenplays and director storyboards by production)

Creative Commons: CC BY 4.0 licence

arXiv: "PersonaShot: Benchmarking Person-Centric Narrative Continuity in Multi-Shot Video Generation" (related work on measuring continuity across cuts)