Movie2Game-Bench tests 14 models on film-to-game adaptation; playable sequences score just 37.9/100
Jialing Xie · hf · 2026-11-19
Researchers from Shanghai Jiao Tong University, Ant Group, and others introduce Movie2Game-Bench, evaluating film-to-game adaptation beyond playability.
- The protocol combines general evaluation, movie-specific evaluation, and playtesting by three GUI model players, linking source content to gameplay steps and implementation evidence
- The benchmark covers 22 films and 658 source-specific reference items; 14 models show complementary strengths in game quality, adaptation, and player experience
- Shared weakness: playable sequences have the lowest content-presence mean across all 14 models, averaging 37.9/100 vs 54.8 for iconic elements
- Among retained emotional-development content, narrative choices account for 60.0% vs 7.2% for scenario operation, pointing to two priorities: preserving source situations that support gameplay and connecting retained content to player actions
More from Multimodal
- NVIDIA open-sources MotionBricks, a real-time motion generator trained on 350K+ clips — Scobleizer · 2026-11-19
- Meshy releases UltraTex for physically plausible PBR texture generation — meshy-ai-team · 2026-11-19
- Redditor plans €300 first test of programmable AI video editing workflow — yeah280 · 2026-11-19
- Lag correction cuts SiT-XL/2 FID from 28.0 to 8.7 on ImageNet — cloneofsimo · 2026-11-19
- AI-generated 'valley where stars flow like water' short stuns Reddit — the-quiet-unreal · 2026-11-19
- Midjourney v8.2 photorealistic portrait prompt shared: candid morning bathroom scene — tisch_eins · 2026-11-19