Issam Alzouby - AI Researcher

Language ModelsDiffusion+2

March 2026 20 min read

Soft-Masked Diffusion Language Models

Continuous feedback MDLM utilizing soft masking to enrich mask tokens with weighted top-k prediction tokens, maintaining the benefits of masking while retaining more prediction information.

Generative ModelsRectified Flow+2

March 2026 20 min read

Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

An overview of Rectified Flow, learning a simpler, direct transport between distributions. Outperforms fast diffusion methods and conceptually simpler, achieving both high quality and fast inference.

February 2026 20 min read

Taming Teacher Forcing for Masked Autoregressive Video Generation

An exploration of Masked Autoregressive Video Generation, focusing on the MAGI framework. Covers the shift from Masked Teacher Forcing (MTF) to Complete Teacher Forcing (CTF) to resolve training-inference mismatch, along with dynamic training strategies.

Video GenerationMasked Modeling+2

QuantizationDeep Learning+2

January 2026 20 min read

Index Backpropagation Quantization

An analysis of resolving traditional codebook learning limitations using Index Backpropagation Quantization (IBQ). Covers the backpropagation trick and specific loss function differentiations compared to standard VQ.

Human MotionGenerative Models+2

January 2026 30 min read

Introduction to Human Motion Generation Models

A foundational overview of human motion modeling, covering 3D parametric body models (SMPL), motion datasets (AMASS, HumanML3D), and current approaches including absolute vs. global data representations.

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.