Splatent: Splatting Diffusion Latents for Novel View Synthesis

A new framework called Relightable and Dynamic Gaussian Avatar (RnD-Avatar) has been proposed to enhance the modeling of relightable and animatable human avatars from monocular video, addressing challenges in achieving photo-realistic results due to insufficient geometrical details related to body motion. This approach utilizes dynamic skinning weights for accurate pose-variant deformation and introduces a novel regularization technique for capturing fine geometric details.

AGORA has been introduced as a novel framework that enhances the generation of animatable 3D human avatars by extending 3D Gaussian Splatting (3DGS) within a generative adversarial network, addressing challenges in rendering speed and dynamic control. This framework utilizes a lightweight, FLAME-conditioned deformation branch for fine-grained expression control and real-time inference.

A novel online scene change detection (SCD) method has been introduced, which is pose-agnostic, label-free, and maintains multi-view consistency, achieving over 10 FPS and surpassing offline approaches in performance. This method utilizes a self-supervised fusion loss, fast pose estimation, and a change-guided update strategy for 3D Gaussian Splatting.

EmoDiffTalk has been introduced as an innovative solution for editable 3D Gaussian talking heads, addressing the limitations in emotional expression manipulation found in previous models. This new approach utilizes an Emotion-aware Gaussian Diffusion process, enabling fine-grained control over facial animations and dynamic emotional editing through text input.

A novel framework named CoherentGS has been introduced to enhance 3D Gaussian Splatting (3DGS) by addressing the challenges of sparse and motion-blurred input images, which often lead to poor reconstruction outcomes. This framework employs a dual-prior strategy, integrating a specialized deblurring network to restore sharp details and a generative model to improve the overall fidelity of 3D reconstruction.

TranSplat has been introduced as a novel method for fast and accurate object relighting within the 3D Gaussian Splatting framework, utilizing a theoretical radiance transfer identity that simplifies the process by leveraging spherical harmonic coefficients without needing explicit scene computations.

A novel framework named MoRel has been introduced, enhancing long-range motion modeling in dynamic videos through an Anchor Relay-based Bidirectional Blending mechanism. This approach addresses significant challenges in 4D Gaussian Splatting, including memory explosion and temporal flickering, by ensuring temporally consistent and memory-efficient modeling of dynamic scenes.

The introduction of ConsDreamer marks a significant advancement in zero-shot text-to-3D generation, addressing the multi-view inconsistencies that have plagued previous methods. By refining the score distillation process through a View Disentanglement Module, this new approach aims to eliminate viewpoint biases and enhance the quality of 3D content creation from textual descriptions.