In an AI-generated adaptation of F1, evaluators found controllable driving and a sprint—but recorded missing repair and rescue components of the reference sequence. The example comes from Movie2Game-Bench, a preprint submitted on November 18, 2026. For film fans and game creators, it suggests a useful starting point: decide what the player should do in a memorable situation, not just which familiar details should appear.
The researchers evaluated 14 combinations of coding models and their supporting software on 22 films. Across those systems, situations that could support gameplay scored lower for being retained than iconic film elements: 37.9 out of 100 against 54.8. These are content-presence scores under the authors’ rubric, not percentages of scenes lost or games that worked. Most judgments came from models, not human audiences.
For a creator, the interesting gap is between recognising the movie and having a part in it. The design question is not simply how much of the film survives. It is what the player gets to do with what survives.
What did the benchmark actually measure?
The assignment was ambitious: turn a complete audiovisual film into a self-contained browser game, with the design left open and up to three hours for unattended generation. The 22-film collection included F1, 12 Angry Men, Searching and Whiplash. This was not simply a contest to build action games.
The comparison also included the software that managed each model’s tools and coding work, often called its harness. Results therefore belong to those model-and-software combinations, not to models tested in isolation. The controlled prototype task differs from learning an environment from gameplay recordings, the question explored in WHAM-RT’s training-data experiment.
To judge the adaptations, the researchers assembled 658 film-specific reference items spanning plot, characters and relationships, iconic elements, playable sequences and emotional development. Models extracted and merged those inventories, with timestamp and dialogue checks. They were fixed before scoring, but were not a human-authored consensus about what each adaptation ought to contain.
Playable sequences had the lowest presence score in every tested system. Unsuccessful generation counted as zero; retained content could receive partial credit. The gap identifies weaker coverage of source situations, not a direct count of broken controls. Nor does it isolate whether omissions came from understanding the film, choosing a design, writing code or exploring the finished game.
Recognising the scene is only one creative job
The F1 example is useful because it contains both success and omission. Driving was controllable. A sprint was present. The evaluators nevertheless recorded missing repair and rescue components. Calling the whole game unplayable would erase what worked; calling the sequence fully adapted would erase what was missing.
Movie2Game-Bench separates source presence from interaction. For its highest interaction label, evaluators must connect observed play to implementation evidence in the code. A generator’s plan to include a feature does not establish that the feature reached the player.
That suggests two different creative decisions. First, which situation is worth carrying over? Second, how does the player participate? A familiar setting answers neither on its own. But neither question demands a steering wheel, a chase or a combat system.
A dialogue decision can be meaningful play. Choosing whom to trust may suit an adaptation better than adding an action sequence. The benchmark deliberately favours source-related interaction, while its conclusion acknowledges artistic goals beyond source coverage. A transformative adaptation can omit a famous scene on purpose. Fidelity is a design choice, not a universal measure of quality.
Did the human players agree?
Only partly—and on a different question. Most evaluation was automated, including exploration by three model players using screenshots and keyboard or mouse actions. In the human comparison, ten participants played 100 distinct games, with one human questionnaire per game.
Human and model-player scores showed positive agreement in their ordering, but model scores averaged 11.1 points higher on the 100-point scale. Agreement about which game scores better is not agreement about how good it is.
The separate overall-enjoyment question was excluded from that scored total. The human exercise also did not independently validate the film-reference inventories or the content-preservation gap. It cannot tell a studio how much a wider audience would enjoy an adaptation. A higher benchmark score is not a forecast of a hit game.
Write the player’s part into the brief
A practical response is to specify one playable situation before requesting a whole film adaptation. This is proposed design guidance, not an experimentally proven prompting recipe. The benchmark’s reference inventories were separate from the generation instructions; the study did not test whether this briefing approach improves results.
Situation: identify the source conflict or moment that matters, and say which details can change.
Action: describe what the player can choose or perform, rather than only what the screen should show.
Consequence: explain what changes because of that action and how the player will notice.
As a hypothetical brief for a detective game: let the player pick a witness, make the answer depend on that choice, and use the information to open a different line of inquiry. The interaction can live in the conversation; it need not become an action minigame.
Then ask to see that action and consequence in the playable prototype. The older SWE-Game research makes a useful complementary distinction: it evaluates implemented behaviour separately from visual presentation, and checks whether demonstrated outcomes also occur when no input is supplied. A success screen alone cannot show that the player caused the success. This is a different benchmark, not a replication of the film-adaptation result.
Automating play introduces another design choice, covered in AI-written neural game players versus ordinary code. It remains separate from deciding what a human player should experience.
For a film adaptation, keep recognition, working interaction and audience enjoyment as three separate questions. Movie2Game-Bench supplies evidence about the first two and limited human evidence about playtesting quality. The useful next brief is not merely “make it look like the movie”. It is: choose the moment, give the player a part in it, and make that participation matter.
