SGA: Plug&Play Geometric Verification for Educational Video Synthesis
SGA checks LLM-made Manim animations for spatial conflicts before they render.
The paper describes a plug-in module that partially executes generated animation code, builds symbolic scene graphs, and refines scenes when it detects occlusion or layout problems. It also introduces MVQS, a deterministic score meant to estimate spatial integrity without rendering the video. On MMMC-Code, SGA improved MVQS in 7 of 8 tested model-and-pipeline settings, with a reported peak score of 73.11 using Code2Video and GPT-5.1. ArXiv · AI/CL/LG's note
The paper describes a plug-in module that partially executes generated animation code, builds symbolic scene graphs, and refines scenes when it detects occlusion or layout problems. It also introduces MVQS, a deterministic score meant to estimate spatial integrity without rendering the video. On MMMC-Code, SGA improved MVQS in 7 of 8 tested model-and-pipeline settings, with a reported peak score of 73.11 using Code2Video and GPT-5.1. ArXiv · AI/CL/LG's note
score 4