This is the code implementation for the NeurIPS paper "A Closer Look at Model Collapse: From a Generalization-to-Memorization Perspective."1
The project website: https://shilianghe007.github.io//model-collapse/index.html
Our code is based on the Hugging Face Diffusers codebase 2. To run the code, you first need to download the Diffusers codebase. And then place the file "train_unconditional.py" under the directory of "diffusers/examples/unconditional_image_generation". Please also prepare the needed training datasets.