Movie2Game-Bench Introduces a Holistic Metric for Turning Films into Playable Games
A new benchmark, Movie2Game-Bench, evaluates film‑to‑game adaptations beyond basic playability.
Source: Grafy Research LabWhat is Movie2Game‑Bench?
Movie2Game‑Bench (M2G‑bench) is a newly released benchmark that measures how well an AI system can turn a feature‑film into an interactive game. The benchmark goes beyond simple playability checks and evaluates whether the resulting gameplay preserves key narrative elements, visual motifs, and character actions from the source film. According to the arXiv pre‑print by Jialing Xie, Yifan Zhou, Yiwen Zhang, and Junjun He, M2G‑bench combines three evaluation layers: a general quality assessment, a movie‑specific content preservation test, and playtesting with three GUI‑based model players.
Why does film‑to‑game adaptation need a new benchmark?
Traditional game‑generation benchmarks focus on mechanics, graphics, or level‑design quality, but they ignore the unique challenge of retaining a film’s story arc and emotional beats. The authors argue that a faithful adaptation must map cinematic scenes to player actions, translate dialogue into in‑game feedback, and respect the film’s pacing. Without a dedicated metric, developers risk producing games that are technically playable yet feel disconnected from their source material. M2G‑bench fills this gap by providing a structured protocol that links source content directly to gameplay steps and implementation evidence.
How was the benchmark constructed?
The benchmark covers 22 films spanning a range of genres, from classic adventure to modern sci‑fi. For each film, the researchers identified 658 source‑specific reference items – these include pivotal plot points, character relationships, visual motifs, and audio cues. Each reference item is paired with a concrete gameplay requirement, such as a specific player action or rule that should reflect the original scene. The evaluation protocol then proceeds in three phases:
- General Evaluation – measures overall playability, control responsiveness, and visual fidelity.
- Movie‑Specific Evaluation – checks whether the gameplay implements the reference items, using a scoring rubric that rates content preservation from 0 (none) to 5 (full).
- Playtesting – three autonomous GUI agents interact with the generated game, producing logs that are compared against the expected interaction sequences.
The authors released the full dataset, reference mappings, and evaluation scripts alongside the paper, enabling reproducibility.
What do the initial evaluations reveal?
Fourteen AI models were tested on M2G‑bench, ranging from large multimodal diffusion models to specialized text‑to‑code generators. The results show a wide performance gap:
- Playability – most models produced games that could be launched, but many suffered from buggy controls or missing assets.
- Content Preservation – the average movie‑specific score hovered around 2.1, indicating that only a minority of reference items were faithfully realized.
- Agent Playtesting – the GUI agents completed the expected interaction sequences in just 38% of the cases, highlighting the difficulty of aligning narrative intent with actionable gameplay.
These findings suggest that current generative pipelines excel at surface‑level creation (textures, geometry) but lack the reasoning needed to translate cinematic storytelling into interactive rules.
How can Grafy creators apply these insights?
Grafy’s Game Studio app already lets users turn a brief description into a playable 2D or 3D game. While Game Studio does not yet implement the full M2G‑bench protocol, creators can adopt several practices inspired by the benchmark:
- Map Narrative Beats Manually – Use the Graph app to build a branching prompt tree that links each major film scene to a specific gameplay mechanic. This mirrors the reference‑item mapping used in M2G‑bench and helps keep the generated logic grounded.
- Iterative Playtesting with Model Agents – Employ Chat to run simple GUI‑player scripts that simulate clicks and key presses. By feeding the agent’s logs back into the Graph, creators can spot gaps where the game diverges from the intended story.
- Leverage the Reader – Load the M2G‑bench paper into Reader to query specific sections (e.g., reference‑item definitions) while working in Game Studio. This keeps the source material at hand without leaving the workspace.
By incorporating these steps, developers can move beyond “looks good” prototypes toward games that truly echo their cinematic inspirations.
What are the next steps for research and tooling?
The authors outline several avenues for improvement. First, they plan to expand the benchmark to include interactive dialogue and dynamic music cues, which are currently under‑represented. Second, they suggest integrating reinforcement‑learning agents that can learn to satisfy the reference‑item constraints during generation, rather than relying on post‑hoc evaluation. Finally, they call for community contributions to broaden the film set and add multilingual content, making the benchmark a living resource for the AI‑generated game community.
The benchmark and accompanying resources are publicly available on arXiv (https://arxiv.org/abs/2611.24669).
Frequently asked questions
How does Movie2Game‑Bench differ from existing game‑generation benchmarks?
M2G‑bench adds a dedicated layer that evaluates narrative fidelity, linking specific film elements to gameplay actions. Traditional benchmarks focus mainly on graphics, level design, or control smoothness, leaving story preservation unmeasured.
Can I use Movie2Game‑Bench to evaluate my own game‑generation model?
Yes. The authors provide the full dataset, scoring scripts, and reference‑item mappings, allowing anyone to run the three‑stage evaluation on custom models.
Does Grafy’s Game Studio already support the full M2G‑bench workflow?
Not yet. Game Studio can generate playable prototypes from briefs, but the detailed content‑preservation scoring and autonomous GUI playtesting described in M2G‑bench are not built in. Creators can approximate the workflow using Graph, Chat, and Reader as described above.


