Critical windows: non-asymptotic theory for feature emergence in diffusion models
Authors: Marvin Li, Sitan Chen
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our bounds with experiments on synthetic data and show that critical windows may serve as a useful tool for diagnosing fairness and privacy violations in real-world diffusion models. |
| Researcher Affiliation | Academia | 1Harvard College, Cambridge, MA, USA 2John A. Paulson School of Engineering and Applied Sciences, Harvard University, Cambridge, MA, USA. |
| Pseudocode | No | The paper describes mathematical frameworks and proofs but does not include any pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide an explicit statement about open-sourcing the code for its methodology, nor does it provide a direct link to a code repository for this work. |
| Open Datasets | Yes | We applied our attack to a DDPM that was trained on CIFAR-10 in (Duan et al., 2023) |
| Dataset Splits | No | For the Membership Inference Attack, the paper mentions '1000 training data samples and 1000 CIFAR-10 held-out samples' but does not specify a separate validation split or explicit training/validation/test splits for any of its experiments. |
| Hardware Specification | No | The paper mentions models like Stable Diffusion v2.1 but does not specify the hardware (e.g., GPU, CPU models, memory) used for running its experiments. |
| Software Dependencies | No | The paper mentions using 'CLIP with the ViT-B/32 Transformer architecture' and refers to the 'DDPM scheduler', but it does not specify version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | We produced 250 images from SD2.1., using 500 time steps from the DDPM scheduler (Ho et al., 2020a) and the prompt Color splash wide photo of a car in the middle of empty street, detailed, highly realistic, brightly colored car, black and white background. |