Search:
Match:
2 results

Analysis

This paper addresses the limitations of 2D Gaussian Splatting (2DGS) for image compression, particularly at low bitrates. It introduces a structure-guided allocation principle that improves rate-distortion (RD) efficiency by coupling image structure with representation capacity and quantization precision. The proposed methods include structure-guided initialization, adaptive bitwidth quantization, and geometry-consistent regularization, all aimed at enhancing the performance of 2DGS while maintaining fast decoding speeds.
Reference

The approach substantially improves both the representational power and the RD performance of 2DGS while maintaining over 1000 FPS decoding. Compared with the baseline GSImage, we reduce BD-rate by 43.44% on Kodak and 29.91% on DIV2K.

Analysis

This paper addresses the challenges of generating realistic Human-Object Interaction (HOI) videos, a crucial area for applications like digital humans and robotics. The key contributions are the RCM-cache mechanism for maintaining object geometry consistency and a progressive curriculum learning approach to handle data scarcity and reduce reliance on detailed hand annotations. The focus on geometric consistency and simplified human conditioning is a significant step towards more practical and robust HOI video generation.
Reference

The paper introduces ByteLoom, a Diffusion Transformer (DiT)-based framework that generates realistic HOI videos with geometrically consistent object illustration, using simplified human conditioning and 3D object inputs.