Reversible Feature Transfer for Zero-shot Photorealistic 3D Scene Stylization

Photorealistic 3D Scene Stylization aims to render photorealistic stylized novel view images with provided style images. Existing methods face challenges in 3D scene content preservation, generalization and diverse outcomes. To address these challenges, we introduce a zero-shot photorealistic 3D stylization framework which manipulates 3D scene via 2D feature maps. A reversible feature transfer module is designed to process content and style features simultaneously and reduce the entanglement between them, ensuring robust content preservation and high-quality style transfer. Moreover, by style feature injecting based on this module, our framework can generalize to arbitrary style inputs for zero-shot stylization. Furthermore, we introduce a specifically designed 2D Dominant Supervised Scheme which utilizes different 2D style transfer methods as supervisors to learn diverse style feature. Experimental results demonstrate the effectiveness of our framework in preserving content details and achieving diverse stylization effects, surpassing existing 3D stylization methods.