Two individuals are engaged in discussion in front of a detailed poster titled "Interactive3D: Create What You Want by Drag." The poster, affiliated with the Hong Kong University of Science and Technology and The Chinese University of Hong Kong, outlines various aspects of an interactive 3D model generation system. The chart illustrates stages such as architectural flow, interaction with Gaussian splitting, rigid dragging, deformable dragging, and semantic editing, showcasing transformative examples like animated dinosaurs and character modifications. Visual aids include QR codes for project details and practical applications. The individuals appear to be exchanging insights or explaining components of the research, contributing to the bustling, intellectually stimulating atmosphere of the event.
Text transcribed from the image:
CHONG KONG
NIVERSITY OF SCIENCE
ND TECHNOLOGY
THE CHINESE
UNIVERSITY OF
ĐI HONG KONG
ction of Interactive3D
Interactive3D: Create What You Want by
Shaocong Dong*, Lihe Ding*, Zhanpeng Huang, Zibin W
Overall Architecture
generation methods cannot offer users precise control.
Interactive3D, granting users precise control over the
red fire dragon
sive process through extensive 3D interaction capabilities.
o-stage generation frame-
ork based on SDS, providing
rious tools for precise con-
trol and guidance of the gen-
eration process.
Interactive Hash Refinement
module for further add details
and optimize the geometry.
a) Deformable Drag (b) Stretch (c) Rigid Drag (d) Semantic Editing
to all wings to enlarge wings to swing wings for barning wing
L. First Stage: Gaussian Splatting
Stage-I Interaction with Gaussian Splatting
Add/Remove Part Geometric Transform
NeRF Distillation
"A Draven Knight
Converter
dragse knight
Gaussian
Splatting
Deformable Dring
Interactive Hash-R
Project Page
Video
Interactive 3D Generation
() Interactive
Hash Refinemest
II. Second Stage: InstantNGP
Rigid Dragging
Stage-II Refinement with Instant NGP
100
InstantNGP
Occupancy Grid
Select Interested Reg
Interactive3D contains two stages with distinct 3D representations: (i)
for flexible user interactions; (ii) the Gaussian blobs are converted to
NERF distillation and fine-tuned by our Interactive Hash Refinement Modu
*Scan QR code of video for more applications and
Rigid Drag
Combination (add part)
Hash Refinement
235
select
Deformable Drag
prompt
TT
Y
Semantic Editing