Yehudit E.

Responsibilities:

• Research - on Textual Inversion: Explored Textual Inversion techniques, including embedding creation and integration into SDXL, to evaluate potential improvements in generation quality.

• Research - Feature Extraction for Dataset Labeling: Used OpenCV to extract morphological and structural features from SEM images for dataset labeling and analysis, including attempts to extract descriptive text from features.

• Dataset Creation using Meta’s SAM-2: Used Meta’s SAM-2 segmentation model to create accurate structural masks for real SEM images and build a paired dataset (mask → image) needed for training and evaluating generative models.

• Pix2Pix Training (Segmentation → Image): Prepared datasets, configured experiments, and trained a Pix2Pix model to generate SEM images from segmentation maps, including epoch tracking, result analysis, and parameter tuning.

• Model Evaluation Metrics Development: Designed evaluation metrics and analysis tools to compare real and generated images, assessing structural accuracy, texture fidelity, and overall realism.

• Investigating Differences Between Real and Generated Images: Analyzed noise patterns, frequency characteristics, and imaging artifacts to understand the root causes of discrepancies between generated and real SEM images.

• Research - Noise Modeling & Enhancement: Added a learnable noise layer to the model, designed to capture SEM-specific noise patterns during training and re-inject them during generation, in order to test whether data-driven noise could improve realism and enhance model performance.

• Documentation & Results Presentation: Produced research summaries, visual comparisons, and presented findings and insights to mentors and the project team.