Shira S.

Responsibilities:

• Research: Explored and compared LLMs such as LLaMA and Gemma, focusing on their capabilities and key differences.

• Developed a full GUI application that receives text input from the user, sends it to an LLM, returns the model’s output to the client - and containerized the entire system with Docker.

• Research: Explored image-to-text models such as PromptCap and Florence, comparing their outputs and understanding how they generate labels from images.

• Research: Trained Textual Inversion on the Stable Diffusion SDXL model to enable it to learn and generate the desired image style.

• Trained a CycleGAN model on SEM and SEG images to enable translation between the two domains.

• Performed metric evaluation on the results (as FID, PSNR, SSIM, LPIPS, and more), and compared the performance with other models trained on the same images.

• Trained a classification model on real and generated images to evaluate how well it distinguishes between them, expecting a score around 0.5, but the results were poor - approximately 0.9.

• If more time were available, the next steps would include adding additional training, exploring explainability methods to understand how the model makes its distinctions, and training a model using the generated images to test how well it later performs on real images. If the model performs well on real data after being trained on synthetic data, it would indicate that the core goal of the project was successfully achieved.