Core Concepts
提案されたDDMIは、高品質な暗黙のニューラル表現を生成するためのドメインに依存しない潜在拡散モデルです。
Abstract
ABSTRACT
INRs provide flexibility and expressivity.
Existing methods limit generative model expressive power.
DDMI proposes adaptive positional embeddings.
Extensive experiments demonstrate superior performance.
INTRODUCTION
INR is a popular approach for representing arbitrary signals.
Recent research focuses on INR generative models using Normalizing Flows, GANs, and Diffusion Models.
DDMI aims to address limitations in existing methods by generating adaptive positional embeddings.
RELATED WORKS
Various works explore the use of INR in generative modeling.
Recent attention has been on domain-agnostic architectures for INR generations.
METHODOLOGY
DDMI utilizes a Discrete-to-continuous space VAE to connect discrete data and continuous functions.
HDBFs and CFC are introduced to enhance expressive power.
EXPERIMENTS
2D IMAGES
Evaluation on AFHQv2 Cat and Dog datasets at different resolutions.
DDMI outperforms baselines in FID scores and precision-recall metrics.
3D SHAPES
Evaluation on ShapeNet dataset with single-class and multi-class settings.
DDMI achieves best MMD and COV scores compared to baselines.
VIDEOS
Evaluation on SkyTimelapse dataset for video generation.
DDMI shows competitive performance compared to state-of-the-art models.
ANALYSIS
Decomposition of HDBFs demonstrates the effectiveness of capturing signals at different scales.
Ablation study shows the impact of components on enhancing fidelity and realism in DDMI.
Stats
DDMIは高解像度画像生成においてFIDスコアで優れた性能を示しました。
DDMIは3D形状生成において最も低いMMDスコアと最高のCOVスコアを達成しました。