In the image, an exhibit hall comes alive with a variety of posters and displays showcasing medical advances. A lively crowd has assembled to engage with the information presented. Among them, a man wearing a black jacket stands in front of a large poster, gazing at the detailed diagrams and data laid out on the display. His posture suggests curiosity and eagerness to learn more about the groundbreaking research being presented. In the foreground, a professional photographer captures the moment, capturing the spirit of innovation and progress in healthcare. The overall scene conveys a sense of excitement and wonder as people gather to explore the cutting-edge advancements being showcased at this convention.
Text transcribed from the image:
Highlight
CVPR
SEATTLE, WA JUNE 17-21, 2024
B
香港中文大學
The Chinese University of Hong Kong
HumanGaussian: Text-Driven 3D Human Generation with Gaussian Splatting
Xian Liu, Xiaohang Zhan2, Jiaxiang Tang³, Ying Shan
The Chinese University of Hong Kong 2Tencent Al Lab
Text-Driven 3D Human Generation
➤ Task Definition: Given an input text prompt that describes
a human, we generate high-quality 3D human with fine-
grained geometry and realistic appearance.
➤ Key Observations: 1) Existing works 3D representations
like mesh or neural fields via score distillation sampling
(SDS), which suffer from inadequate fine details or
excessive training time. 2) While 3DGS enables real-time
scene modeling, the sparse optimization gradient requires
structural guidance, especially in the human domain.
3) The naive SDS necessitates a large classifier-free
guidance (CFG) scale for image-text alignment, over-
saturated patterns and stochastic SDS loss make the
original 3DGS gradient-based density control unstable,
incurring blurry results with floating artifacts.
Our solution: Incorporate explicit structural guidance and
gradient regularization to 3DGS optimization!
Framework
Overview
Structure-Aware SDS
3Peking Unive
Our
How to initialize 3DGS in t
Initialize 3DGS center pos
How to optimize with both
Instead of distilling from a u
model capturing the joint dis
Texture-Structure Joint Model
Test Prompt: "A woman wearing
skirt, cropped top, and y
>How to a
Guide
Ve SDS=2
Annealed Nega
Regularize
CSD to ind
Snc T11
➤Size-cond
to remove
threshold
a Lin, Xihui Liu, Ziwei Liu5
Nanyang Technological University
Qualitative Visual
shape.
e?
ucture.
V (d; p.y) - v
convergence?
extural aspects.
p.y)-a)
adj
FSD and
uality:
p. 0)]
ly phase
in
ifacts.
Visual Compariso
G
Pow Shop Kandard ROA Red De
(A) baseline
Prund
6월 19일 수요일
6:22
ents
Aware SDS:
human appearance
try with explicit
uidance. The score
m both RGB and
is used to distill the
on and pruning.
Annealed Negative Guidance:
Decompose SDS into a noisier
generative score and a cleaner
classifier score. Eliminate the
floating artifacts based on 3DGS
size in a prune-only phase for
the realistic renderings.
Ablation Stu
HUSH
(a) HumanGanian (Dars)
Conclusion with C
We propose an
HumanGaussia
Aware SDS and
rompt Guidan
generat
and r