AI Researcher & Developer
Mehdi Dehghani Firoozabadi
Generative models, diffusion, and compact neural architectures
I work on generative models — especially diffusion and score-based methods — with an interest in compact, interpretable architectures. My M.Sc. thesis at Tarbiat Modares University develops an improved diffusion-based generative model. Alongside research, I build forecasting pipelines and teach Python, machine learning, and deep learning.
Focus
Diffusion & score-based models
Noise schedules, sigma-score estimation, and generative models that treat denoising as the central computation.
Compact, interpretable nets
Parameter-efficient architectures and custom layers — including bilateral matrix projections — that stay inspectable.
Forecasting & inverse problems
Multivariate time-series pipelines in practice, and inverse problems approached with score-based diffusion.
Selected work
All work- Sigma score estimator2026
PyTorch training of ResNets to estimate ∇_σ log p(x,σ) for adaptive noise schedules in diffusion models.
- Image generation evaluations2026
Unified toolkit for FID, Inception Score, sFID, Precision, Recall, and FD_DINOv2, with automatic reference downloads.
- Bi-matmul network2025
Custom bilateral matrix layers that squeeze MNIST, CIFAR, and Fashion-MNIST inputs toward target shapes.
- EDM gradient-ascent denoising2026
Iterative denoising of CIFAR-10 images by maximizing log-probability with pretrained EDM models.
- Ideal denoiser2025
Closed-form ideal denoiser (EDM equation 57) with CIFAR-10 experiments and visualizations.