My Projects
Dynamic 3D Relighting in Video Diffusion Models
This project introduces a novel technique to teach AI how to separate an object's 3D shape from its lighting in videos. Using specialized LoRA models, we disentangle pure geometry (LoRA-G) from diverse lighting effects (LoRA-L).
The core feature is enabling precise, dynamic control over video lighting. Users can easily manipulate and animate light sources, such as moving a spotlight or changing the time of day, with temporally consistent results.
It will be released as an open-source toolkit, including a "Lighting Palette" of pre-trained styles and tools for the community to train custom models for their own unique objects and lighting scenarios.
GitHub
Sim3D - Realtime Cloth Simulation and Physics Engine
A 3D physics simulation engine developed from scratch in C++ using OpenGL. It features a mass-spring system for textured cloth simulation, advanced lighting with Phong shading, and efficient batch rendering. The engine includes a particle system with uniform grid-based collision detection and an interactive scene editor built with ImGui. Object-Oriented principles were applied to ensure a modular and extensible design.
View Devlog GitHub
DreamLab - Latent Diffusion Transformer for Text2Video Generation
This project is a smaller-scale implementation of a text-to-video generation model, inspired by OpenAI's Sora. I designed and experimented with various autoencoder architectures, focusing on minimizing quality loss through a combination of perceptual and reconstruction losses. I implemented the denoising process using a Diffusion Transformer (DiT) architecture and am currently working on incorporating spatio-temporal attention and text conditioning with CLIP embeddings.
GitHub