I am a PhD student at the GrUVi Lab at Simon Fraser University and the University of Cyprus, advised by Ali Mahdavi-Amiri, Daniel Cohen-Or and Yiorgos Chrysanthou.
I work on image and video generation and editing, with a focus on control, and on using video models as world models and reasoners. I also work on 3D/4D scene representations.
I am a PhD student working on image and video generation and editing. I want diffusion and flow models to be controllable, and I want to understand what video models know about the world, so they can be used as world models and for reasoning.
My recent work is on joint audio-video diffusion models. Soundwich splits the generated audio into stems (speech, music, effects, ambience) that can be edited independently. Sound Sparks Motion edits motion at test time, without retraining the model.
Earlier I worked on 3D and 4D reconstruction at ETH Zurich and INSAIT, and on robot manipulation at the University of Toronto. I did my BSc at Sharif University of Technology, and worked at Tapsi on ETA and navigation models.
Precise control over image, video and audio-video diffusion models without retraining them: tuning the conditioning at test time, and splitting what a model generates into parts that can be edited separately.
Video diffusion models appear to learn physics, geometry and cause and effect. I want to understand how to use them as simulators of the world, and as models that can reason.
Dynamic scene representations: Gaussian Splatting and implicit fields that handle large topological change, and compression that keeps visual fidelity.
* equal contribution
Decomposes the audio of joint audio-video generative models into independently controllable stems (speech, music, sound effects and ambience) tied together by a shared scene representation. Each stem can be adjusted, muted, replaced or remixed while staying in sync with the video.
A test-time tuning framework for motion-focused video editing inside audio-visual generative models. Instead of retraining, we tune two lightweight variables, an audio latent and a text-conditioning residual, guided by vision-language model feedback. The learned controls transfer across videos.
Lossy 3DGS compression with learnable Beta-distribution confidence scores. Low-confidence splats are pruned while visual fidelity is preserved. The method is architecture-agnostic and introduces average confidence as a scene-quality metric.
A new architecture and temporal-consistency technique for 4D reconstruction under large topological changes, with a Gaussian Splatting rendering scheme and a geometry-appearance disentanglement learned directly from RGB images.
Uses Scene Representation Transformers to render novel views from RGB observations, then predicts end-effector poses and gripper states from language-conditioned keyframe demonstrations.
Mostly tools for 3D vision and generative models on Apple Silicon.
Selected for Anthropic's Claude for Open Source program. My main open-source projects are mlx3D and mlx_diffuser.
Differentiable 3D computer vision on Apple Silicon, built on MLX: meshes, NeRF and 3D Gaussian Splatting with Metal kernels.
MLX implementations of well-known generative models for efficient training, inference and fine-tuning on Apple Silicon.
A PyTorch-compatible library to design, run and visualize models for different 3D computer vision tasks.
Ten puzzles for learning GPU programming in Metal.
A research-assistant agent for many research tasks, powered by Gemma.
A game engine built from the ground up for AI agents.
Video generative models and 3D computer vision, with Ali Mahdavi-Amiri and Daniel Cohen-Or.
Machine learning and computer graphics research.
Dynamic 3D reconstruction with SDFs inside Gaussian Splatting; built a dataset of hard deformable scenes.
NeRFs for dynamic scenes; SDF morphing from RGB images and SDF-based regularizers.
Robot manipulation from NeRF-based scene encodings.
Led two projects from research to production on the MAP team: a GNN for spatio-temporal city modelling and ETA prediction, and an attention-based ETA post-processing model.
First member of Tapsi Lab R&D. Built the Driver-Arrived AR feature, running the full pipeline client-side with TensorFlow.js.
Implementations from scratch, mostly for learning.
A tiny NeRF in PyTorch, small enough to train on a CPU.
GitHub
Basic ray tracing with only NumPy and Matplotlib.
GitHub
A perceptual loss for comparing blurred images.
GitHub
2D terrain generation with noise generators.
GitHub
TF-IDF index, ML ranking and a Streamlit frontend.
GitHub
Scanner, parser and semantic analysis, compiling to Python.
GitHubTeaching assistant at Sharif University of Technology, Oct 2021 – Jul 2025.
Service: Vice President of the Students' Scientific Chapter (SSC) · Head of the Technical Beta Team, Winter Seminar Series · Founder, Byte Publication · RTI Program, Oxford · 3D Vision Reading Group, University of Toronto.
Email is the best way to reach me. I am glad to talk about video generation, 3D vision or collaborations. Outside research I play video games and read science fiction (Steam, Goodreads).