Pedro Figueiredo

Generates realistic SVBRDF materials using diffusion models and reinforcement learning. We leverage Stable Diffusion XL and a realism reward function trained on real images to bridge the gap between synthetic training data and real-world material appearance.

Estimates spatially varying light selection distributions to improve importance sampling in Monte Carlo rendering. To efficiently manage hundreds or thousands of lights, we integrate our neural approach with light hierarchy techniques using a residual learning strategy.

Introduces a fast and expressive neural path guiding method that breaks down the 2D distribution over the directional domain into two 1D probability distribution functions (PDF) modeled by tiny MLPs. Integrates radiance caching to reduce variance of optimization.

Interpolates challenging near-duplicate photos leveraging bidirectional optical flows encoded via a hypernetwork. The proposed continuous representation generates superior intermediate frames particularly under challlenging illumination variations.

Introduces an enhanced automatic audio description system that leverages both movie scripts and video visual cues to deliver succinct, synchronized descriptions for blind and visually impaired users.

Explores the use of a ray tracer application to enhance student engagement in functional programming through a teaching experience with positive outcomes.

Proposes an efficient bidirectional conversion of triangle meshes to signed distance fields using a fast octree-based discretization, which enables advanced operations like offsetting and morphological transformations.

Reports on a Physically-Based Rendering course experience in a computer science program, emphasizing its role in boosting student motivation, reducing dropout rates, comparing its content with other courses, and offering suggestions for future iterations.

Finetuned pre-trained image diffusion models to enable the next generation of real-time generative gaming on mobile GPUs. Introduced automatic data resampling to improve scene-wide conditioning, driving more consistent video generation. Engineered multi-node distributed training to accelerate iteration speed while adhering to a strict inference VRAM budget.

Researched frame interpolation algorithms for DLSS, exploring neural optical flow estimation toward multi-frame generation. Designed architecture improvements to mitigate ghosting artifacts from disocclusions and to optimize rendering speed.

Built a containerized deep-learning resource forecasting pipeline to optimize infrastructure allocation. Developed an internal semantic search engine to accelerate documentation discovery for engineering teams.

Created OpenGL application using octrees to leverage signed distance functions as a way of performing fast affine transformations in 3D meshes. Novel mesh representation allows real-time computation of operations not feasible in real-time before. Research on signed distance functions and real-time cone tracing.

Developed prototypes for Ericsson’s edge computing platform leveraging quick response times from 5G hardware. Engineered and deployed 5G IoT applications for cloud computing hosted on AWS and Microsoft Azure. Research on containerization applied to 5G IoT applications.

Developed and optimized of software for image and video processing. Implemented application that leveraged object recognition via CNNs and text-to-speech tools for the visually impaired.

Taught rendering concepts to a public school teacher via interactive ray tracing assignments on WebGL to inspire graphics-related classroom activities in the following school year.

Lectured and developed assignments for practical classes of CSCE 121 Introduction to Program Design & Concepts (C++) - Spring 2021; and CSCE 221 Data Structures and Algorithms Spring 2022. Developed main project deliverable and performed grading for CSCE 310 Database Systems - Summer 2022. Hosted office hours and performed grading for CSCE 441 Computer Graphics - Fall 2021, Fall 2022; and CSCE 448 Computational Photography - Spring 2023.

Lectured practical functional programming undergraduate classes for Spring and Fall 2019. Research focused on ray tracing application on functional programming for motivating functional programming students.

Estimates surfaces of arbitrary watertight triangle meshes by approximating signed distance functions (SDFs). We use an efficient octree to approximate distance values of nearby triangles. Allows real-time visualization of operations best defined in the implicit SDF representation such as the offset, union, and intersection.

Logon Renderer is a project started on the Physically Based Rendering course, offered by Christian Pagot, at Universidade Federal da Paraíba. The application is a physically based path tracer written from scratch in C++ for educational purposes.