ABOUT

Hi! I'm Supasorn or Aek. I'm now a lecturer at VISTEC, a new research institute in Rayong, Thailand with a really nice, foresty campus. I'm looking for students, RAs/interns, postdocs, and all sorts of collaborators. Shoot me an email if you're interested. Also for PhD, please apply directly here.

News / Talks

Two papers [1] [2] with one Oral @ CVPR 2024
One paper [1] @ ICLR 2024!
Three papers [1] [2] [3] @ ICCV 2023
Two papers [1] [2] @ CVPR 2023
One paper [1] @ ICLR 2023
Area Chair for ICCV 2023
Outstanding reviewer CVPR 2023
Best Oral Presentation Award (AI Session) from PMU-B Brainpower Congress 2022.
Our lab was awarded Google Research Gift!
Our lab won ThaiSC HPC sponsorship!
One Oral @ CVPR 2022 (Diffusion Autoencoders)
Our lab won "AI for All" frontier research grant 2021! (Research and Innovation Thailand PMU)
Our lab received an overseas Adobe Research Gift!
Area Chair for CVPR 2022
Two oral papers @ CVPR 2021; One best paper candidate!
Our lab won Frontier Research Grant from Research and Innovation Thailand
One paper accepted to ICDE 2021
Outstanding reviewer awards: CVPR 2019, CVPR 2021
KeyPointNet Oral @ NeurIPS 2018, Montreal
Keynote Speaker @ O'Reilly AI Conference 2018, London
Digital Thailand Big Bang 2018, Bangkok
TED 2018, Vancouver

Services

Area Chair: ICCV 2025, CVPR 2025, ICCV 2023, CVPR 2022
Reviewer: CVPR (outstanding reviewer x 3), ICCV, ECCV, SIGGRAPH + Asia, TPAMI, IJCV, TOG
DEI Chair: ACCV 2024
Program Chair: MLRS 2023

Bio

Before settling down in Rayong, I was a research resident at Google Brain working on geometric understanding in deep learning and image synthesis. I finished my Ph.D. from UW working with Prof. Steve Seitz and Prof. Ira Kemelmacher in Graphics-Vision group GRAIL. My goal is to bring computer vision out of the lab into the real world and make it really work in the wild. I went to Cornell for undergrad, and had a great pleasure working with Prof. John Hopcroft on social graph algorithms, and later got inspired by Prof. Noah Snavely with his computer vision class. I love hacking, coding, tackling hard problems, and I try very hard to make my solutions as simple as possible.

Contact

My first name at gmail. If I don't reply, feel free to ping me. It's likely I lost it in the pile of 30K+ mails.

RESEARCH

DiffusionLight: Light Probes for Free by Painting a Chrome Ball

P. Phongthawee, W. Chinchuthakun, N. Sinsunthithet, A. Raj, V. Jampani, P. Khungurn, S. Suwajanakorn CVPR 2024 (Oral)
A novel lighting estimation technique that inpaints a chrome ball into the scene using a pre-trained diffusion model. We propose an iterative inpaiting algorithm and fine-tune Stable Diffusion XL for exposure bracketing. Our method marks the first work that achieves good generalization to diverse, real-world images.
Paper Page Code

DNO: Optimizing Diffusion Noise Can Serve As Universal Motion Priors

K. Karunratanakul, K. Preechakul, E. Aksan, T. Beeler, S. Suwajanakorn, S. Tang CVPR 2024
A new method that effectively leverages pre-trained text-to-motion diffusion models to optimize diffusion noise for various motion-related tasks, without requiring retraining for each task. DNO allows for efficient motion editing and control, surpassing previous methods in achieving objectives while preserving motion content.
Paper Page Code

Diffusion Sampling with Momentum for Mitigating Divergence Artifacts

S. Wizadwongsa, W. Chinchuthakun, P. Khungurn, A. Raj, S. Suwajanakorn ICLR 2024
Our study reveals divergence and artifact issues when using high-order methods for low-step diffusion sampling. To tackle this, we introduce two approaches: integrating Heavy Ball momentum to enhance stability and a new method, Generalized Heavy Ball (GHVB), to balance accuracy and artifact suppression.
Paper Page Code

DiFaReli: Diffusion Face Relighting

P. Ponglertnapakorn, N. Tritrong, S. Suwajanakorn ICCV 2023
Our state-of-the-art face relighting technique solves highly challenging scenarios such as cast shadows, strong highlights, unusual makeups, and facial accessories. It only needs 2D images to train---no need for light stage data, multiview images, relit pairs, 3D or lighting ground truth!
Paper Page Code

GMD: Guided Motion Diffusion for Controllable Human Motion Synthesis

K. Karunratanakul, K. Preechakul, E. Aksan, T. Beeler, S. Suwajanakorn, S. Tang ICCV 2023
GMD solves conditional human motion generation based on text prompts, reference trajectories, key locations, etc., without retraining diffusion models for each of these tasks. Our contributions include an effective feature projection scheme, a new imputation formulation, and a dense guidance approach. GMD achieves significant improvements over state-of-the-art methods.
Paper Page Code

Zero-guidance Segmentation Using Zero Segment Labels

P. Rewatbowornwong, N. Chatthee, E. Chuangsuwanich, S. Suwajanakorn ICCV 2023
We propose a novel problem zero-guidance segmentation---segment everything without "guiding" text queries or prompting. Our first solution to this problem leverages two pre-trained generalist models, DINO and CLIP, without any fine-tuning or segmentation datasets.
Paper Page Code

Learning Geometric-aware Properties in 2D Representation Using Lightweight CAD Models, or Zero Real 3D Pairs

P. Arsomngern, S. Nutanong, S. Suwajanakorn CVPR 2023
This study leverages CAD models to improve 2D scene understanding, overcoming the scalability issues of traditional 2D-3D datasets. By using geometrically aligned 3D spaces and pseudo pairs, it achieves comparable results in various tasks while using much less data.
Paper Page

StyleGAN Salon: Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer

S. Khwanmuang, P. Phongthawee, P. Sangkloy, S. Suwajanakorn CVPR 2023
A multi-view GAN optimization framework for virtual hair try-on. Our technique supports various highly challenging scenarios, such as transforming a long hairstyle with bangs to a pixie cut. Our method outperforms the state of the art according to the most comprehensive user study.
Paper Page

Accelerating Guided Diffusion Sampling with Splitting Numerical Methods

S. Wizadwongsa, S. Suwajanakorn ICLR 2023
This paper identifies the inefficacy of existing high-order numerical methods for guided sampling and proposes a solution using operator splitting methods. Our method has the same quality as 250-step DDIM while using 32-58% less sampling time on ImageNet256 and works across various tasks.
Paper Code

Toward Ant-Sized Moving Object Localization Using Deep Learning in FMCW Radar: A Pilot Study

N. Kumchaiseemak, I. Chatnuntawech, S. Teerapittayanon, P. Kotchapansompote, T. Kaewlee, M. Piriyajitakonkij, T. Wilaiprasitporn, S. Suwajanakorn IEEE Geoscience and Remote Sensing 2022
A deep learning-based approach to localizing a small moving object with a single millimeter-wave frequency-modulated continuous-wave (FMCW) radar.
Paper

Diffusion Autoencoders: Toward a Meaningful and Decodable Representation

K. Preechakul, N. Chatthee, S. Wizadwongsa, S. Suwajanakorn CVPR 2022 (Oral)
Our diffusion-based autoencoders can simultaneously infer compact semantic code and stochastic code from an input image. These codes are useful for downstream tasks and allow near-exact reconstruction. Diff-AE can solve tasks like real-image manipulation without GANs and their error-prone inversion and produces more discriminative latent space than StyleGAN-W.
Paper Page Code

Nex: Real-time View Synthesis with Neural Basis Expansion

S. Wizadwongsa, P. Phongthawee, J. Yenphraphai, S. Suwajanakorn CVPR 2021 (Oral) - Best paper candidate
We present NeX, a new approach to novel view synthesis based on enhancements of multiplane image (MPI) that can reproduce NeXt-level view-dependent effects---in real time. A 1000x speed up from SOTA.
Paper Page Code Data

Repurposing GANs for One-shot Semantic Part Segmentation

N. Tritrong, P. Rewatbowornwong, S. Suwajanakorn CVPR 2021 (Oral) / TPAMI 2022
We present a simple and powerful method that repurposes GANs for few-shot semantic part segmentation. Our approach achieves surprising and unprecedented performance and is competitive with fully-supervised baselines that require 10-50x more labeled examples.
Paper Extended-Paper Page

Self-Supervised Deep Metric Learning for Pointsets

P. Arsomngern, C. Long, S. Suwajanakorn, S. Nutanong ICDE 2021 / TPAMI 2021
We propose a self-supervised deep metric learning solution for pointsets, which enables effective representation learning on unlabeled datasets. Our key idea is the use of the Earth's Mover Distance (EMD) to generate pseudo labels. Our method produces results superior to other supervised competitors, including an NLP-specific approach based on BERT.
Paper Journal

Discovery of Latent 3D Keypoints via End-to-end Geometric Reasoning

S. Suwajanakorn, N. Snavely, J. Tompson, M. Norouzi NeurIPS 2018 (Oral)
We present KeypointNet, an end-to-end geometric reasoning framework to learn an optimal set of 3D keypoints, along with their detectors. Our model discovers semantically consistent keypoints across viewing angles and object instances and outperforms a fully supervised baseline on the task of pose estimation -- all without keypoint location supervision.
Paper Page Code

Synthesizing Obama: Learning Lip Sync from Audio

S. Suwajanakorn, S.M. Seitz, I. Kemelmacher-Shlizerman SIGGRAPH 2017 / TED 2018
Given audio of President Barack Obama, we synthesize photorealistic video of him speaking with accurate lip sync. Trained on many hours of just video footage from whitehouse.gov, our recurrent neural net approach synthesizes mouth shape and texture from audio, which are composited into a reference video.
Paper Page Code TedTalk Discussion

What Makes Tom Hanks Look Like Tom Hanks

S. Suwajanakorn, S.M. Seitz, I. Kemelmacher-Shlizerman ICCV 2015
Madrona Prize Winner / Innovation of the Year 2016
We reconstruct a controllable model of a person from a large photo collection that captures his or her persona, i.e., physical appearance and behavior. Our system is based on a novel combination of 3D face reconstruction, tracking, alignment, and multi-texture modeling, applied to the puppeteering problem.
Paper Page

Depth from Focus with Your Mobile Phone

S. Suwajanakorn, C. Hernández, S.M. Seitz CVPR 2015
We introduce the first depth from focus (DfF) method capable of handling images from mobile phones and other hand-held cameras. With this technique, we can automatically generate a depth map for every photo you take with your phone.
Paper Supplement Video Data Patent

Total Moving Face Reconstruction

S. Suwajanakorn, I. Kemelmacher-Shlizerman, S.M. Seitz ECCV 2014 - Madrona Prize Runner-Up
Our approach takes a single video of a person's face and reconstructs a high detail 3D shape for each video frame. We target videos taken under uncontrolled and uncalibrated imaging conditions.
Paper Page

Illumination-aware Age Progression

I. Kemelmacher-Shlizerman, S. Suwajanakorn, S.M. Seitz CVPR 2014
We present an approach that takes a single photograph of a child as input and automatically produces a series of age-progressed outputs between 1 and 80 years of age, accounting for pose, expression, and illumination.
Paper Page

Extracting the Core Structure of Social Networks Using (α, β)-Communities.

Liaoruo Wang, John Hopcroft, Jing He, Hongyu Liang, Supasorn Suwajanakorn Internet Mathematics, 2013
We present a heuristic algorithm that in practice finds a fundamental community structure and demonstrate that the core structure in social networks is due to underlying social structure rather than high-degree vertices or degree distribution.
Paper

Detecting the Structure of Social Networks Using (α,β)-Communities

Jing He, John Hopcroft, Hongyu Liang, Supasorn Suwajanakorn, Liaoruo Wang 8th Workshop on Algorithms and Models for the Web Graph (WAW) 2011
A talk I gave on my algorithm used in the paper (not in WAW).
Talk

CON$ULTING


San Francisco, CA	Mountain View, CA	Bangkok, Thailand	San Francisco, CA

WORK EXPERIENCE

AI Foundation

Since 2017 I'm serving as a member of the Global AI Council, AI Foundation, a San Francisco-based company focusing on applications and platforms enabled by personal AI.

Google Research Intern

Summer 2015

I spent a wonderful summer working with Michael Rubenstein, Ce Liu, image/video/motion analysis experts, at Google Cambridge/MIT on motion-based 3D reconstruction, and received tremendous help and input from the team with Bill Freeman, Dilip Krishnan, and Inbar Mosseri.

Google Software Engineering intern

Summer 2013

I lingered around Seattle in 2013 and worked with Carlos Hernandez, a 3D-vision expert, and Steve Seitz, my academic advisor, at Google Seattle in Fremont on uncalibrated depth from focus and eventually published a paper two years later.

In 2010 during undergrad, I interned with Harish Venkataramani as a software engineering intern at Google Mountain View on a project related to Gmail and Google+.

Cornell Student Web Programmer

2007 - 2008

I built a website for NY beginning farmer project under the College of Agriculture and Life Sciences and implemented a questionnaire app that helps guide farmers to the help they need.

SOURCE CODE

My work consists of multiple large components with ten of thousands lines of code, so this will take time and I may not be able to provide support. The "research-code" is neither polished nor properly commented, but I decided to release it now for research and educational purposes. No commercial use allowed.

Synthesizing Obama: Learning Lip Sync from Audio

Here's the network training code that takes as input processed MFCC coefficients and outputs mouth fiducial points represented as PCA coefficients.

Uncalibrated Photometric Stereo

Future release.

3D Optical Flow

Future release.

PERSONAL

My other interests include: photography, 3D printing, product design, graphics design, software dev, startup, sitting in a hammock, getting lost for fun. I play badminton and squash and sports that involve gliding and wheels such as skiing, snowboarding, skating. I flew off my bike a few times and enjoyed numerous wipe-outs from surfing, but still survived. I love hacking and building things and here are some of things I built.

Light Field Bot - 3D Photosphere

Supasorn Suwajanakorn

An automatic DSLR rig for capturing 2D/3D 360 photosphere and dense light field for VR. The rig is controlled with Wemos D1 Mini (ESP8266). It has a joystick and OLED screen for displaying menu. The structure is aluminum and 3D printed parts, designed in Fusion 360. Software and hardware are open-source.
Thingiverse Stitched OmniStereo 3D Pano

High-Res 360 Spherical Camera

Supasorn Suwajanakorn

A prototype camera that automatically captures a 360 x 180 photosphere similar to Google Street View with resolution up to 100M pixel. It's a low-cost version ~$70 of gigapan made from a Raspberry Pi + custom 3D printed case.
The camera rotates around the no-parallax point and is remotely controlled from a smartphone through a dedicated WiFi.
Github UW MakerSpace Tour My Boston APT

2D Pattern to 3D Origami Popup

Supasorn Suwajanakorn, Jonathan Hirschberg, Roopa Roa, Michael Tomaine Winner - The Faculty Award (Boom 2011) - Cornell University
An application that lets you design a 2D pattern that can be automatically turned into a 3D foldable Origami popup model.
Video

Low-noise Stills from Video

Supasorn Suwajanakorn, Andre Baixo Graphics Project - Graphics 2013 - UW
A mobile app which combines multiple noisy shots taken hand-held into a single low-noise photo. What makes it special is that you can throw your tripod away and take pictures with your shaky hand and it will work just fine.
Report

Rovio & Juliet - Autonomous Indoor Navigation using SURF

Jae Yong Sung, Supasorn Suwajanakorn, Jong Hwi Lee Best Project Award - Robot Learning 2011
Cornell University
Our goal is to make Rovio a totally autonomous robot which can follow waypoints only based on image and reach the goal position while learning and avoiding obstacles on its way.
Video Report

"Mech Tournament"

Supasorn Suwajanakorn, Natachai Laohachai, Poom Pechavanish Best Game Award - 7th Thailand National Software Contest 2005 & Asia Pacific ICT Merit Awards
A 2D multiplayer shooting game based on DirectX 8 and DirectPlay. You get to choose your own robot and enter a battle against other players over a LAN network or against computers with challenging AI.

Classroom Controller

Supasorn Suwajanakorn, Pochara Arayakarnkul Best Application Award - 6th Thailand National Software Contest 2004
An application for controlling and monitoring a computer classroom. The teacher will have the ability to inspect students' screens or any running applications, limit internet or application access, issue online quizzes or polls on any machines in the classroom.

Remote Fish Feeder

Supasorn Suwajanakorn, Naiyarit Sanpol, Thanapol Tanprayoon Winner - National Science Project Contest 2003
In grade 9, I built a remote fish feeding device that is connected to a home telephone line. The device can then be activated by calling home from anywhere and press a secret passcode. Fish will thank you while you're sunbathing on a beach.

Google Cardboard Positional Tracking Hack

Supasorn Suwajanakorn

I'm trying to add positional tracking to Cardboard App (Full 6DOF). Proper calibration is needed. Tracking is done on computer and position values are sent through USB to the phone (TCP via ADB). Latency is not superb, but should be better than using phone's camera and phone's processing power.
Github YouTube

3D Augmented-Reality Graphing Calculator

Supasorn Suwajanakorn, Yu Cheng, Cooper Findley Showcase at Boom 2011 - Cornell Unversity
Our fun app that can plot a 3D graph on a piece of paper. During the contest, some number of kids had a lot of fun hand-drawing 2D patterns e.g. hearts, their names, random arts and visualizing 3D surface in real-time.

jitouch

Supasorn Suwajanakorn, Sukolsak Sakshuwong Runner-Up - Ars Technica Design Award Student Mac App NYTimes / macworld
An awarding-winning Mac app that expands the set of multi-touch gestures for trackpad, Magic Mouse for frequent tasks such as changing tabs in web browser, closing windows. It also recognizes handwritten alphabets as input gestures. Please check it out!
Website

Obox iPhone

Supasorn Suwajanakorn

A fast-paced puzzle game on iPhone that I made over one summer during an internship (not related). Unfortunately, it's no longer available and I did not have time to update for newer iOS.
Video

jiswitch

Supasorn Suwajanakorn

jiswitch is a free Mac application that introduces a new way to switch applications. In a nutshell, it allows users to assign any window a name, and later bring that window to the top whenever the same name is typed. This tool is meant for power-users or coders who have many windows opened.
Website

Designs

Boom 2010 Logo

Supasorn Suwajanakorn Winner - Science Showcase Logo Contest - Cornell University
A logo I designed for Cornell's "annual showcase of student research and creativity in digital technology and applications." I used Sunflow for global illumination rendering and Structure Synth to synthesize the model. -- Earn me an iPod touch

Cornell Class of 2011 Logo

Supasorn Suwajanakorn Winner - Cornell University

A logo selected to represent Cornell class of 2011. -- Earn me an iPod nano

Cornell Minds Matter Logo

Supasorn Suwajanakorn Winner - Cornell University

A logo selected to represent Cornell Minds Matter, an organization that promotes the overall mental and emotional health of all Cornell students. -- Earn me an iPod touch

Thai Festival 2010

Supasorn Suwajanakorn

A poster for Cornell Thai Festival 2010. It contains my hand-drawn Thai ancient art called "Lie Kranok" as seen above the text.

Cornell Thai Association

Supasorn Suwajanakorn

A logo for Cornell Thai Association. It's the clock tower with a new twist on the roof.