cv

Ph.D. in Computer Science

• Focus: Physically plausible and realistic video and image models.

M.Sc. in Computer Science

• Focus: Computer vision, machine learning, and advanced deep learning architectures.
• Master's thesis at the Max Planck Institute for Informatics (ReFree-S2V, NeurIPS 2026).

B.Sc. in Computer and Systems Engineering

• Bachelor thesis: Masked Controlled Image Generation (Supervisor: Prof. Dr. Ayman Khalafallah).

ReFree-S2V: Towards Realistic Co-Speech Video Generation via Reward-Free RL and Multilevel Speech Guidance

• Mohamed, S., Mughal, M. H., Dabral, R., Theobalt, C.
• A reward-free RL framework combined with multi-level speech representations to generate temporally consistent, high-fidelity talking-portrait videos. Project page · arXiv

Gaussian Windows for Multi-Focal Attention

• Mohamed, S., Mughal, M. H., Dabral, R., Theobalt, C.
• A Gaussian window mechanism that lets attention capture multi-focal structure in the data, beyond the locality imposed by positional encoding.

Diffusion-based Identity-Preserving Facial Privacy Protection

• Han, D.*, Mohamed, S.*, Li, Y., Denzler, J. (*equal contribution)
• A latent diffusion obfuscation pipeline that preserves facial expressions and fine details while preventing automated identity matching. Paper

Fusion is all you need: Face Fusion for Customized Identity-Preserving Image Synthesis

• Mohamed, S., Han, D., Li, Y.
• Identity-preserving conditioning via cross-attention feature injection for zero-shot personalized generation. arXiv

ShieldDiff: Suppressing Sexual Content Generation from Diffusion Models through Reinforcement Learning

• Han, D.*, Mohamed, S.*, Li, Y. (*equal contribution)
• RL-guided fine-tuning that unlearns harmful concepts without degrading benign generation. arXiv

TCIG: Two-Stage Controlled Image Generation with Quality Enhancement through Diffusion

• Mohamed, S.
• A two-stage generative model combining pose/layout conditioning with progressive latent diffusion refinement. arXiv

Master Thesis Researcher

• Advisors: Prof. Christian Theobalt, Dr. Rishabh Dabral (Jan 2025 – Feb 2026).
• Researched generative diffusion models for human video dynamics.
• Developed a reward-free reinforcement learning framework aligning video diffusion with multi-modal acoustic features.
• Scaled training across GPU clusters with PyTorch DDP and DeepSpeed.

Research Intern

• May 2024 – Oct 2024.
• Researched safety alignment for text-to-image models, leading to ShieldDiff (RL-based concept suppression).
• Developed an identity-preserving anonymization pipeline with conditional latent diffusion, published at ICASSP 2025.
• Designed zero-shot face fusion architectures for personalized image synthesis, surpassing baselines in identity fidelity.

Part-time Machine Learning Engineer

• Jul 2021 – Oct 2023.
• Fine-tuned open-source LLMs (LLaMA-2) with LoRA on custom financial instruction datasets.
• Built NLP sequence-classification pipelines to detect user life-stage events from transaction metadata.
• Designed high-concurrency caching infrastructure with Redis and RabbitMQ for sub-100 ms inference.

Software Engineer Intern

• Aug 2022 – Sep 2022.
• Implemented C++ reflection, dynamic JSON/binary serialization, and runtime GUI generation.

Machine Learning & Software Engineering Intern

• Jun 2021 – Sep 2021.
• Built real-time pose estimation and hand-gesture tracking with OpenCV and MediaPipe for automated workout logging.
• Optimized a C++ plotting engine for 1M+ point arrays, reducing latency from 60 s to under 1 s.

PyTorch D3PM (Discrete Diffusion Probabilistic Models)

• Clean-room PyTorch re-implementation of Google's D3PM with discrete Markov transition matrices for categorical latent spaces.

Attention-Based Skin Cancer Classification

• Multi-scale spatial attention models on HAM10000, reaching 94% classification accuracy.

Health Monitor: Distributed Big Data Log Engine

• Real-time log ingestion and batch analysis with Hadoop HDFS, MapReduce, and Apache Spark.

• 16th place out of 72 teams, ICPC Alexandria Collegiate Programming Contest