Assignment 05 (Due date: Nov. 16, 2025)
Neural Networks
Assignment 05
 
Due: Nov. 16, 2025
 
The goal of this assignment is to develop a clear and comprehensive understanding of the architecture, components, and implementation of a Stable Diffusion model. Through this exercise, you will strengthen both your conceptual knowledge and hands-on skills in working with diffusion-based generative models and their underlying neural network mechanisms.
 
Description
A ZIP file is provided containing all required materials for this assignment. The archive includes:
  • Source code implementing a Stable Diffusion model based on the U-Net architecture.
  • A pre-trained model trained for 100 epochs.
 
You may use the pre-trained model to generate samples or as a reference point for comparison with models trained by you using a smaller number of epochs.
 
Instructions
 
  • Download and extract the provided ZIP file.
  • Review and run the code to familiarize yourself with its structure and execution.
  • The code includes comment lines and embedded questions designed to guide your exploration of the model’s structure and functionality.
  • Answer all embedded questions by adding your responses as comment lines directly within the code.
  • If any of your responses require mathematical derivations, equations, or detailed explanations, include them in a separate document (Microsoft Word or PDF) and submit it alongside your updated code.
  • Ensure your comments are clear, concise, and properly formatted, and that any referenced sections of code are easy to identify.
 
Deliverables
 
  • The updated source code file with your answers and explanations inserted as comments.
  • (Optional) A supplementary Word or PDF document containing extended discussions, derivations, or figures if needed.
 
Learning Outcomes
Upon successful completion of this assignment, you should be able to:
  • Describe the core concepts of diffusion models, including the forward and reverse diffusion processes.
  • Explain how the U-Net architecture functions as a denoiser within the Stable Diffusion framework.
  • Understand and describe the role of noise scheduling and latent space representations.
  • Execute sampling using a pre-trained diffusion model and interpret the resulting image generation process.
  • Modify, train, and evaluate diffusion models using practical code implementations.
  • Critically analyze model behavior and training outcomes through empirical comparison.
Top