Seeing Through the Facade: Leveraging Computer Vision to Understand the Realism, Expressivity, and Limitations of Diffusion Models
By: Christopher Pondoc, Joseph O'Brien, and Joseph Guman
Recent research in natural language processing and computer vision has shown progress in image generation from text. Projects such as DALLE-2 and Stable Diffusion 2.0 are examples of diffusion models trained on large language datasets and can create photorealistic images, transfer styles from one image to another, and process complex queries. For some, generative models amplify human creativity by serving as a tool that supercharges the creative processes. However, many others feel that developments will only lead to more problems, such as troubling deepfakes, the replacement of human artists by artificial intelligence, and disinformation that can be more quickly created and disseminated.
In the context of these malicious applications, we explore the task of binary classification between real and AI-generated images. In practice, our work could be applied to help tag disinformation and deepfakes around the Internet. Our process aims to answer three questions:
- How can we achieve the highest test accuracy on classifying between real and fake images? What is our limit with a custom-built convolutional neural network (CNN), and how we can leverage state-of-the-art computer vision techniques to push the bound higher?
- When making classifications, what features distinguish real images from AI-generated images, and DALLE-2 and Stable Diffusion 2.0? What visualization and training techniques can we use to generate these insights?
- Finally, which prompts generate the most photorealistic results and successfully trick the highest performing classifier? Can this information connect back to the expressivity of each diffusion model?
With this approach, we hope to offer both (a.) insights into the limitations of existing diffusion models as well as (b.) potential areas for improvement in the task of photorealistic image generation.
We split up our code into several components.
The main folder contains all of the notebooks used for the main part of the experiment. Model Training.ipynb is where we train all of our networks. Generating Heatmaps.ipynb is where heatmaps are generated after training a network. Prompt Analysis.ipynb is where we use natural language processing (NLP) to analyze and determine the characteristics of a good prompt for both DALLE-2 and Stable Diffusion.
This folder contains code pertaining to repeatedly used functions. dataset.py defines the dataset class used for training. transforms.py defines the image transforms applied when preprocessing.
This folder contains code pertaining to our neural network architectures. convbasic.py contains code for our baseline CNN, while transferlearning.py contains code for our ResNet-18 architecture.
This folder contains code for all of the analysis done outside of model training. Bias vs. Variance.ipynb contains code for generating graphs to visualize the bias-variance tradeoff experiment we ran to determine the optimal data.
This folder contains all of our graphs and output logs pertaining to training and testing. ConvBasic contains all results for our baseline CNN, while TransferLearning contains all results for our ResNet-18 architecture. Each folder also has a subfolder pertaining to the specific task it was trained on: real vs. DALLE-2, Stable Diffusion, or pitting Stable Diffusion against DALLE-2. Finally, BiasVariance contains the graphs necessary for making a conclusion on the bias-variance tradeoff experiment, which we outlined in the Methods section of our paper.
Each folder contains graphs of training and test accuracy against the proportion of training data used. Full text logs of training and testing are also included.
This folder contains miscellaneous items: old scripts used when trying out old datasets (i.e. ImageNet for our milestone), scripts for scraping the DALLE-2 data using the OpenAI API, random AI-generated images, and graphs from old training runs.
We decided not to upload our dataset nor our weights to GitHub. If you are interested in gaining access, please contact either Chris ([email protected]), O'Brien ([email protected]), or Guman ([email protected]) for more information.