Mam3DGen: An efficient 3D texture generation model based on Mamba network

With the widespread application of virtual images in animation, gaming, and figurines, converting 2D images into 3D models has become a major challenge in the field of computer graphics. In recent years, Hunyuan3D 2.0 has made significant progress in large-scale 3D synthesis, but its approach based on Transformer and diffusion models still faces issues such as high computational resource consumption and long training times. To improve computational efficiency during the inference phase, this paper proposes the Mam3DGen network architecture. This architecture introduces the Mamba network and, through hierarchical teacher-forcing strategy, knowledge distillation, and cross-modal attention mechanisms, maintains texture consistency and realism from multiple perspectives, enhancing the stability and accuracy of 3D texture generation while optimizing computational resource usage during training. Experimental results show that on the Danbooru2020 and AnimeDL-2M datasets, Mam3DGen outperforms existing methods. The generated 3D textures not only have higher quality but also show significant improvements in computational efficiency, especially demonstrating excellent performance in high-resolution generation tasks.