I am currently a PhD student at MIT CSAIL in the Computer Graphics group, advised by Prof. Fredo Durand. I also collaborate closely with Prof. Tzu-Mao Li and Prof. Jonathan Ragan-Kelley.
Currently, I focus on differentiable simulation methods, more specifically, on differentiable rendering and its applications to computer vision and perception. Other topics I'm actively pursuing include compilers for differentiable programming, real-time path tracing, neural rendering, and physically-based modelling.
My previous experience includes a 6-month research staff position at CMU (advised by Prof. Ioannis Gkioulekas and Prof. Anat Levin) and two consecutive summer internships at INRIA (advised by George Drettakis).
Throughout my PhD, I have been fortunate to be partially supported by a 2022 NVIDIA Graduate Fellowship and a 2019 MIT Edgerton Fellowship.
In my time-off, I like to travel, take pictures, and design websites. The end of this webpage has a few of my favourite pictures.
You can find a copy of my CV here.
If you have any questions, contact me at sbangaru@mit.edu
Updates:
Joining NVIDIA's real-time rendering research lab as full-time research scientist to push the limits of ML in Graphics!
Our paper on warped-area reparameteization of differential path integrals received a best paper award at SIGGRAPH Asia 2023!
Conference proceedings and Journal publications
Exploits common structure across a variety of BRDF models to derive efficient importance sampling methods for their derivatives, through a mix of positivization and product/mixture decomposition
HLSL-like shading language that supports first-class automatic differentiation and object-oriented programming for large-scale differentiable rendering frameworks and inter-operation with tensor ML frameworks
Received a Best Paper Award.
This paper shows how to combine the warped-area reparameterization with the material-form parameterization to present an efficient way to estimate derivatives w.r.t geometry in the presence of complex multi-bounce illumination.
This work shows how to differentiate a neural SDF renderer without approximations using reparameterization, with applications to 3D reconstruction on real and synthetic multi-view datasets.
This work systematically address the challenges of differentiating parametric discontinuities under integration, by developing a programming language, proving correctness of its semantics, and demonstrating applications including shader design and contact simulation.
First paper to derive and implement an unbiased estimator in the area-measure for the differentiable rendering equation.
Applies a path-space differentiable rendering algorithm to improve photometric BSDF recovery using concave objects.
A solution to the problems with shape-recovery methods. Uses a mitsuba shape-differentiable path tracer and a TensorFlow optimizer to obtain accurate reconstructions in the presence of interreflections and non-lambertian surfaces.
Similar to Action-Conditional Video Prediction, which predicts the next frame of an ATARI game given the user input, ACPNNs are designed to predict 2D and 3D environments where the output is based on perspective vision.
Exploration in multi-task reinforcement learning is critical in training agents to deduce the underlying MDP. We present a novel method to facilitate exploration in multi-task reinforcement learning using deep generative models.
Pictures from around the world!