U of Toronto CS PhD at DGP | Prev. @AdobeResearch @NVIDIA : )
Toronto, Canada
- Tweets 154
- Following 938
- Followers 1,079
- Likes 11,762
Pinned Tweet
Check out our latest work on image model merging for stylization!
Paper announcement📢 GimmBO: Interactive Generative Image Model Merging via Bayesian Optimization
Chenxi Liu, Selena Ling (@seleniumlzh), Alec Jacobson
🏆 SIGGRAPH 2026 Best Paper
Selena and I will be presenting in LA (July 19-23)!
🌐 Project: gimmbo-project.github.io/
[1/7]
6
727
Adam works amazingly well for both ML and geometry problems. But when parameters are geometric, Adam is not rotation-equivariant and leads to artifacts.
We propose a fix with "VectorAdam for Rotation Equivariant Geometry Optimization" w/ @nmwsharp @_AlecJacobson #NeurIPS22 (1/7)
12
156
1,155
Our #Siggraph25 work found a simple, nearly one-line change that greatly eases neural field optimization for a wide variety of existing representations.
“Stochastic Preconditioning for Neural Field Optimization” w/ @merlin_ND @_AlecJacobson @nmwsharp
8
45
314
57,039
Our #SGP25 work studies a simple and effective way to uniformly sample implicit surfaces by casting rays. (1/9)
“Uniform Sampling of Surfaces by Casting Rays” w/ @_abhishekmadan @nmwsharp @_AlecJacobson
1
15
138
12,592
With a common Laplacian regularization energy that is intrinsic to a mesh, Adam’s first optimization step results in a different output while VectorAdam preserves the optimization trajectory under rotation. (6/7)
1
4
58
We show many more experiments across geometric optimization and machine learning in our paper. We also provide a PyTorch implementation here github.com/iszihan/VectorAda…. Please check out our #NeurIPS22 paper here bit.ly/3DGcsyZ and reach out if you have any questions! (7/7)
1
6
54
We introduce VectorAdam as a solution by extending the per-scalar operation of Adam to vectors. (3/7)
1
28
We observe that Adam’s first and second moment estimation operates per-scalar, which results in uniform rescaling for each component of the gradient vector, snapping the gradient along the diagonal direction of the current coordinate system. (2/7)
1
28
This simple change preserves the gradient direction and ensures uniform rescaling wrt gradient norm. (4/7)
1
1
28
Most importantly, if we take a functional perspective for optimizers, considering it as a function that maps initial condition to optimized result, we prove that VectorAdam is rotation-equivariant. (5/7)
1
1
28
It’s as simple as perturbing query locations according to a normal distribution. This produces a stochastic estimate of the blurred neural field, with the level of blur proportional to a scale parameter 𝛼.
1
2
15
2,273
We show many more experiments across different neural field representations in our paper. Please check out our #Siggraph25 paper here research.nvidia.com/labs/tor… and reach out if you have any questions!
1
14
1,579
We argue that this is a quick and easy form of coarse-to-fine optimization, applicable to nearly any objective or field representation. It matches or outperforms custom designed polices and staged coarse-to-fine schemes.
1
1
12
1,306
Surprisingly, optimizing this blurred field to fit the objective greatly improves convergence, and in the end we anneal 𝛼 to 0 and are left with an ordinary un-blurred field.
2
1
9
1,481
Geometric initialization is a commonly-used technique to accelerate SDF field fitting, yet it often results in disastrous artifacts for non-object centric scenes. Stochastic preconditioning also helps to avoid floaters both with and without geometric initialization.
1
1
8
1,830
Neural field training can be sensitive to changes to hyperparameters. Stochastic preconditioning makes training more robust to hyperparameter choices, shown here in a histogram of PSNRs from fitting preconditioned and non-preconditioned fields across a range of hyperparameters.
1
2
6
1,254
Congrats to @anagh_malik !!
Delighted to share the first project of my PhD, "Transient Neural Radiance Fields for Lidar View Synthesis and 3D Reconstruction".
We show unprecedented capabilities of synthesizing novel lidar scans from as few as 2 input views!
🖥️anaghmalik.com/TransientNeRF
7
1,064
We show many more experiments across different implicit surface representations in our paper. Please check out our #SGP25 paper here arxiv.org/pdf/2506.05268 and reach out if you have any questions! Code coming soon! (9/9)
6
578
With uniformly sampled points, one can also easily perform importance sampling using curvature or other quantities like losses, and construct geometry-aware regularization terms to improve neural implicit optimization. (8/9)
1
5
628
Our method is more efficient than the common alternatives: rejection sampling, sampling on extracted meshes via Marching Cubes, and a principled sampling algorithm using Markov chain Monte Carlo (e.g., Hamiltonian Monte Carlo). (4/9)
1
5
416
Lovely piece written by @yyinuoshi! :)
When the filmmaker Azza Cohen asked her 82-year-old grandma what was on her bucket list, the answers surprised her. Cohen’s short documentary “FLOAT!” follows her bubbe as she learns to swim. nyer.cm/7Bidrc0
4
457
A uniformly sampled set of points on implicit surfaces enables many downstream applications:
One can take our white noise samples and easily subsample to blue noise samples. (6/9)
1
4
360
Our white noise samples are also essential for enabling neural implicit deformation as proposed in [Yang et al. 2021]. (7/9)
1
3
329
So proud!
Fresh off the press: our paper received the David Marr Best Paper award (2/8000+) at #ICCV2023
3
755
!!
How can we convert stereo 360 video to 6DoF for more comfortable VR? We use multi-sphere images like a Russian doll - MatryODShka! visual.cs.brown.edu/matryods… with @imarhombus @seleniumlzh @SkyLi0n @c_richardt. Talking now @ #eccv2020 oral session bitly.com/3hCLw5N
3
Our method exploits a classic mathematical relationship: to sample a point set, gather all intersections of randomly-cast rays against the surface — and intersecting rays with implicit surfaces is easy! (3/9)
1
1
3
454
Suppose you have an implicit surface, like a neural SDF or shadertoy-style analytic function, and you want to uniformly sample points on the surface 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 lossy mesh extraction. (2/9)
1
2
455
If I understand the problem correctly, unfortunately I don't think VectorAdam could solve that as VectorAdam does a similar uniform scaling for gradients that has a vector structure.
2
:000 Looks so good!
Super excited to share what me and the amazing @LumaLabsAI team has been working on!!! 🥰🥰🥰
Introducing Genie 🧞♂️- 3D models generated in 10 seconds, which can then be turned into HQ 3D assets in 15 mins?!? 🤯🤯🤯
Link to join: lumalabs.ai/genie
Reach out to work with us on building state-of-the-art 3D foundation models!
2
875
@Ivana_G7 oh there should be a rule that ap teachers are not allowed to give more work after the exam T T
1
1
2
🦌🦌🦌
The Canadian launch of 🦌THE DEER🦌 was a success!! Thanks so much to everyone for coming out and to @jeanmarcahsen and Claire at @typebooks for being such great hosts.
Type Bookstore is almost sold out of copies!! Grab yours now before they're all gone
2
We did not apply to Gaussian Splats but this paper tried something similar there! ubc-vision.github.io/3dgs-mc…
1
2
210
More specifically, sampling on extracted meshes from isosurfacing algorithms like Marching Cubes requires expensive evaluation to a grid and easily aliases thin structures, while our method is both efficient and accurate. (5/9)
1
2
377