AI Innovation Institute

Abstract: Visual generation is a fundamental problem in computer vision and graphics, with applications ranging from 3D capture to image and video synthesis. Despite rapid progress in neural rendering and generative models, efficiency remains a key bottleneck: high-quality 3D reconstruction often relies on dense multi-view supervision; high-fidelity 3D synthesis requires costly optimization, training, and rendering; and modern image and video generators require substantial computation as the number of tokens grows rapidly for high-resolution generation. This dissertation focuses on efficient visual generation by improving sample efficiency in 3D reconstruction, representation efficiency in 3D generation, and computational efficiency in image and video synthesis. First, we improve the sample efficiency of neural implicit surface reconstruction. We integrate multi-view stereo probability volumes as a geometric regularizer, enabling high-quality sparse-view reconstruction from as few as three input images. Next, we introduce an explicit 3D representation for 3D generation, built from multi-view depth and RGB images with 3D Gaussian features. This design allows the model to use 2D generative priors while enforcing multi-view consistency through epipolar attention. We then address the computational bottleneck in image and video synthesis with importance-based token merging. Our method uses importance signals available during generation to preserve critical information while merging redundant tokens. Finally, we enable efficient mixed-resolution diffusion transformers via phase-aligned attention. This approach stabilizes attention under mixed-resolution token grids and unlocks high-fidelity image and video generation at reduced cost. Taken together, these contributions reduce the data requirements, representational overhead, and computational demands, thereby providing a foundation for high-quality, scalable, and efficient visual generation.

Speaker: Haoyu Wu

Location: NCS 120