Projects

(Ongoing) Implementation of ICtCp color profile awareness in text-2-video diffusion models.

(Ongoing) Implementation of training and inversion free video editing pipelines of MMDiT video diffusion models such as CogVideoX or Qwen Video Edit.

Implemented simple advertisement banner insertion on courts of indoor and outdoor sports events that remains temporally consistent and visually appealing, with client agnostic metadata that can deliver hyper-personalized advertisement to viewers with light metadata packet.

Alleviate transformer’s nature of quadratic memory and compute complexity using a Mamba state-space model and transformer hybrid that reduces compute time by up to 30% with similar levels of performance and enable high resolution 3D prediction on stereo and flow downstream tasks.

Implemented a residual network-based radiance field predictor that compresses 6 DoF images and videos, leading to 10% to 50% bitrate savings.

Implemented transport of 6 DoF images using HEVC.

The method was later successfully adapted in 3D Gaussian Splatting compression as well.

Implemented a 100x faster than SOTA pipeline for a transformer-based image super-resolution algorithm that draws texture information from a reference texture provider image.

Development of Camera2 API pipeline for NVIDIA Tegra devices.