The field of novel view synthesis and 3D generation is moving towards more efficient and effective methods for generating high-quality images and 3D models from single images or videos. Researchers are exploring new approaches to address challenges such as cross-view consistency, novel view quality, and dynamic scene synthesis. Notable advancements include the use of lightweight foundation models, meta-learning, and Gaussian splatting frameworks to improve reconstruction quality and reduce computational overhead. These innovations have the potential to enable more immersive and interactive 3D scene exploration experiences.

Some noteworthy papers in this area include: VEIGAR, which presents a computationally efficient framework for 3D object removal without relying on an initial reconstruction phase. MetaQAP, which proposes a novel no-reference IQA model that leverages quality-aware pre-training and meta-learning to achieve exceptional performance on benchmark datasets. HoliGS, which introduces a holistic Gaussian splatting framework for embodied view synthesis that achieves superior reconstruction quality and reduces training and rendering time. Active View Selector, which reframes active view selection as a 2D image quality assessment task and achieves substantial quantitative and qualitative improvements. WonderFree, which enables users to interactively generate 3D worlds with the freedom to explore from arbitrary angles and directions. DBMovi-GS, which proposes a method for dynamic view synthesis from blurry monocular videos via sparse-controlled Gaussian splatting. Geometry and Perception Guided Gaussians, which seamlessly integrates geometry and perception priors to reconstruct detailed 3D objects from a single image.