Alleviating Exposure Bias in Diffusion Models through Sampling with Shifted Time Steps
Authors: Mingxiao Li, Tingyu Qu, Ruicong Yao, Wei Sun, Marie-Francine Moens
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically and theoretically show that, during inference, for each backward time step t and corresponding state ˆxt, there might exist another time step ts which exhibits superior coupling with ˆxt. Based on this finding, we introduce a sampling method named Time-Shift Sampler. Our framework can be seamlessly integrated to existing sampling algorithms, such as DDPM, DDIM and other high-order solvers, inducing merely minimal additional computations. Experimental results show our method brings significant and consistent improvements in FID scores on different datasets and sampling methods. |
| Researcher Affiliation | Academia | 1 Department of Computer Science, KU Leuven 2 Department of Mathematics, KU Leuven {mingxiao.li,tingyu.qu,ruicong.yao,sun.wei,sien.moens}@kuleuven.be |
| Pseudocode | Yes | We present the training and sampling algorithms of the original Denoising Diffusion Probabilistic Models (DDPM) (Ho et al., 2020) in Algorithm 1 and 2, respectively. |
| Open Source Code | Yes | Our code is available at https://github.com/Mingxiao-Li/TS-DPM. |
| Open Datasets | Yes | We report main results using pre-trained DDPM on CIFAR-10 (Krizhevsky, 2009) and Celeb A 64 64 (Liu et al., 2015). Moreover, based on DDIM sampler, a comparison to ADM-IP (Ning et al., 2023) is made, which uses ADM (Dhariwal & Nichol, 2021) as the backbone model. |
| Dataset Splits | No | The paper mentions using pre-trained models on standard datasets like CIFAR-10 and Celeb A, but does not explicitly specify the training, validation, or test dataset splits used for its experiments or refer to standard splits for these datasets within the context of their experimental setup. |
| Hardware Specification | Yes | Test conducted on an AMD EPYC7502 CPU and a RTX 3090 GPU using Pytorch time estimation API. |
| Software Dependencies | No | The paper mentions "Pytorch time estimation API" but does not provide specific version numbers for PyTorch or any other key software libraries used in the experiments. |
| Experiment Setup | No | The paper discusses the selection of window sizes and cutoff values for its proposed Time-Shift Sampler, but it does not provide specific hyperparameters for model training (e.g., learning rate, batch size, optimizer settings) as it primarily uses pre-trained models. |