Issam Alzouby - AI Researcher

Generative ModelsDiffusion Models+2

Diffusion Transformers with Representation Autoencoders

Exploring the replacement of standard VAEs with pretrained representation encoders (DINO, SigLIP, MAE) to form Representation Autoencoders (RAEs). Analyzes challenges in high-dimensional latent spaces and achieves state-of-the-art ImageNet results.

Generative ModelsMotion Models+2

BAMM: Human Motion Modeling & Architecture

Deep dive into BAMM's architecture versus current T2M challenges. Details the Motion Tokenizer, Masked Self-Attention Transformer, Hybrid Attention Masking, and Cascaded Motion Decoding.

Generative ModelsMotion Models+3

MoMask: Generative Masked Modeling of 3D Human Motions

An in-depth look at MoMask, a framework for 3D human motion generation using masked modeling. Covers Architecture, Residual VQ-VAE, Masked Transformer training, and Inference.

Generative ModelsMotion Models+2

Training BAMM - Bidirectional Autoregressive Motion Model

A step-by-step walkthrough of the BAMM pipeline, covering motion data collection, training the VQ-VAE for discrete representation, and training the Bidirectional Autoregressive Transformer (BAR) for motion generation.

Generative ModelsMath Concepts+2

January 2026 30 min read

From Autoencoders to VQ-VAEs: A Mathematical Timeline

A comprehensive timeline and deep dive into the mathematical concepts evolving from standard Autoencoders to VQ-VAEs. Covers the progression of generative models, focusing on the shift from continuous to discrete latent spaces.