About
I am a second-year PhD student at the Computer Science and Artificial Intelligence Lab (CSAIL) at MIT, advised by Professors Fredo Durand and Jonathan Ragan-Kelley.
I graduated from UC San Diego in Spring 2025 with a B.S. in Computer Science (Honors with highest distinction) and a minor in Mathematics. There I did computer graphics research at the Center for Visual Computing, mentored by Professors Ravi Ramamoorthi and Tzu-Mao Li.
My research interests span neural graphics, differentiable rendering, and appearance modeling, with a focus on efficient algorithms for 3D graphics at the intersection of machine learning and computer graphics.
Publications
Venkataram Sivaram, Sai Praveen Bangaru, Ravi Ramamoorthi, Tzu-Mao Li, Jonathan Ragan-Kelley, Frédo Durand
A graphics program is many parts that must agree on how memory is shaped and shared, and today they agree only by convention. What if the compiler checked instead? Declaring each resource once, as a contract, catches mismatches at build time rather than as a corrupted frame.
Venkataram Sivaram, Ravi Ramamoorthi, Tzu-Mao Li
Renderers model how light bounces in great detail, but rarely how it is made. Can a neon tube be rendered from the physics that lights it? Simulating the discharge directly gives its uneven, characteristic glow, with a few knobs for different gases.
Kaiwen Jiang, Venkataram Sivaram, Cheng Peng, Ravi Ramamoorthi
Reconstructing 3D shape from photos with Gaussian blobs tends to smooth away fine detail. Is that loss inherent, or just an approximation in the math? Removing the approximation sharpens the recovered surface at no extra cost.
Venkataram Sivaram, Tzu-Mao Li, Ravi Ramamoorthi
Neural networks compress 3D shape well, but usually into formats no standard renderer can draw. Can a mesh go in and come back out a mesh? Storing a coarse cage plus a small network that adds the detail shrinks the file substantially and rebuilds the original.
Wesley Chang, Venkataram Sivaram, Derek Nowrouzezahrai, Toshiya Hachisuka, Ravi Ramamoorthi, Tzu-Mao Li
ReSTIR makes lighting cheap by recycling samples between nearby pixels and frames. Does the trick still work when you need derivatives, as inverse rendering does? Recycling in parameter space instead carries the speedup over.
Experience
3D Graphics Software Intern, Slang Compiler Team
Summer 2024 & 2025Worked at NVIDIA on the Slang shader compiler's code generation and surrounding tooling across two summers, largely where one language's semantics have to survive translation into a backend that does not share them.
Teaching Assistant, CSE 167
Fall 2024, Winter 2025Taught computer graphics twice at UC San Diego, from rasterization and shading through ray tracing and global illumination. Ran sections and office hours, and wrote and graded the programming assignments.
Reviewer
2024–2026Review submissions for ACM SIGGRAPH and ACM Transactions on Graphics, mostly in rendering, geometry processing, and appearance modeling.
Awards
EECS Great Educators Fellowship
2025Awarded by the MIT EECS department to incoming graduate students with a demonstrated commitment to teaching.
A national award from the Computing Research Association recognizing undergraduates across North America who show outstanding potential in computing research.
Departmental award from the UC San Diego CSE department for sustained undergraduate research, for work at the Center for Visual Computing.
Honorable mention in the same Computing Research Association competition the prior year.
Other Works
Morphing one shape into another from outlines alone usually needs derivatives, which outlines do not give up easily. Can it be done without them? Treating each outline as a pile of mass and moving one onto the other pushes the vertices directly.
A report on TokenFormer (Wang et al., 2024), which treats a transformer's parameters as attention tokens so the model can grow without retraining from scratch.
A global illumination method that keeps light probes in hash tables rather than a dense grid, buying higher probe density with fast updates and cheap neighbor lookups.
A ReSTIR variant that stores samples in world space rather than on screen, via a k-d tree or uniform grid, so they survive leaving the frame and can still be reused.
Blogs
Coming soon!