Hello, if you have any need, please feel free to consult us, this is my wechat: wx91due
CDS521 Course Dissertation:
Deadline: [24th April, 2025]
Objective:
This course dissertation aims to introduce you to two powerful generative models in AI: Diffusion Models and Generative Adversarial Networks (GANs). You will explore their underlying principles, compare their strengths and weaknesses, and understand their applications in generating realistic data (e.g., images, audio, or text). This is designed to be inclusive for all students, regardless of programming experience. If you have any questions or need clarification, feel free to reach out!
Requirements:
1) Theoretical Understanding:
In this part, you need to consider including the following parts in your report. Introduce the background and research status in this area. Explain the key concepts of Diffusion Models and GANs in your own words. Describe the step-by-step process of how each model generates data. Compare and contrast the two models in terms of training stability, output quality, and computational requirements on different tasks.
2) Application-Based Analysis:
In this part, you may want to consider including the following parts in your report. Provide examples of real-world applications for each model (e.g., image synthesis, video generation, or art creation). You can discuss the ethical implications of using these models (e.g., deepfakes, misinformation). You can discuss the future directions with these models.
3) Programming or Non-Programming Task (for students with/without programming background):
Use online tools or platforms (e.g., Runway ML, Artbreeder) to experiment with pretrained Diffusion Models or GANs. Generate a sample output (e.g., an image or text) and write a short reflection on the process and results.
Optionally, you can play with the project (https://github.com/rupeshs/fastsdcpu). You can also implement a simple GAN or Diffusion Model using a high-level AI framework (e.g., TensorFlow, PyTorch). Train the model on a small dataset and generate sample outputs.
5) Preliminary Information:
Diffusion Models: A generative model that gradually adds noise to data and then learns to reverse the process to generate new data. Start with a real data sample, corrupt it with noise over multiple steps, and then train a model to denoise it.
Generative Adversarial Networks (GANs): A framework where two neural networks (a generator and a discriminator) compete against each other. The generator creates fake data, while the discriminator tries to distinguish between real and fake data.
( https://poloclub.github.io/ganlab/ , https://thisbeachdoesnotexist.com/ )
6) Key Concepts to Review:
Basic probability and statistics (e.g., distributions, noise). Neural networks (high-level understanding). Loss functions and optimization (for GANs).
7) Reading Materials:
"Denoising Diffusion Probabilistic Models" (2020) by Jonathan Ho et al.
"Generative Adversarial Networks" (2014) by Ian Goodfellow et al.
8) Submission Guidelines:
Submit a 4-page PDF document, with your written explanations, comparisons, and analyses. All references should be properly cited. You can include screenshots or outputs from your experiments (if applicable).