Adaptive Multi-Modal Fusion Rendering Model for 3D Scene Visualization Based on Visual Image Language

Abstract

This paper presents an Adaptive Multi-Modal Fusion Rendering Model (AMMFRM) designed to address the limitations of traditional 3D visualization scene rendering based on visual image language. The proposed model integrates visual feature perception, semantic reasoning, and adaptive rendering strategy generation to improve rendering efficiency and quality. To validate the model, experiments were conducted on two benchmark datasets—FilmScene and GameWorld—which include complex cinematic and interactive game scenes, respectively. Comparative evaluations were performed against four baseline models: classic ray tracing, physically based rendering (PBR), Neural Radiance Fields (NeRF), and differentiable volume rendering. Results demonstrate that AMMFRM achieves a rendering quality score of 8.5 on FilmScene and 8.2 on GameWorld (out of 10), with semantic restoration rates of 92% and 90%, respectively. Rendering time was reduced by over 40% compared to ray tracing. These findings confirm AMMFRM's superior performance in multi-style, multi-scale scene rendering, establishing its potential as a robust solution for film, television, and gaming applications.