Semi-Implicit Denoising Diffusion Models (SIDDMs)

Authors: yanwu xu, Mingming Gong, Shaoan Xie, Wei Wei, Matthias Grundmann, Kayhan Batmanghelich, Tingbo Hou

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our experiments, we evaluate our proposed method in the simulated Mixture of Gaussians and several popular public datasets, Cifar10-32 [33], Celeb A-HQ-256 [34] and the Image Net1000-64[35]. To study the effects of our model components, we also conduct the ablations to identify their sensitivity and effectiveness.
Researcher Affiliation Collaboration Yanwu Xu1,2*, Mingming Gong3, Shaoan Xie4, Wei Wei1, Matthias Grundmann1, Kayhan Batmanghelich2 , Tingbo Hou1 ... 2Electrical and Computer Engineering, Boston University, {yanwuxu,kayhan}@bu.edu 3School of Mathematics and Statistics, The University of Melbourne mingming.gong@unimelb.edu.au 4Carnegie Mellon University shaoan@cmu.edu 1*Work done as a student researcher of Google.
Pseudocode No The paper includes diagrams and equations but no structured pseudocode or algorithm blocks.
Open Source Code Yes The code is available at https://github.com/xuyanwu/SIDDMs.
Open Datasets Yes In our experiments, we evaluate our proposed method in the simulated Mixture of Gaussians and several popular public datasets, Cifar10-32 [33], Celeb A-HQ-256 [34] and the Image Net1000-64[35].
Dataset Splits Yes In our experiments, we evaluate our proposed method in the simulated Mixture of Gaussians and several popular public datasets, Cifar10-32 [33], Celeb A-HQ-256 [34] and the Image Net1000-64[35]. These datasets commonly have predefined train/validation/test splits, and the paper implicitly uses them for evaluation as shown in Tables 1-5 where FID/IS/Recall scores are reported, implying evaluation on a test/validation set. For example, Table 5 explicitly mentions FID scores on CIFAR10, which implies using its standard split.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., GPU models, CPU types, memory) used for running the experiments.
Software Dependencies No The paper mentions applying 'Unet [36]' for model architectures but does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes For the model architectures, we apply the Unet [36] as the ADM [2], and we follow the same efficient strategy as Imagen [4] to change the order of downsampling and upsampling layer in the Unet blocks. ... For the Image Net, we choose four denoising steps to get the best sample quality on the Image Net. ... We set the weights of AFD to be [0.0, 0.1, 0.5, 1.0, 5.0, ], where 0.0 represents only the adversarial term and denotes only the AFD term in our training. We apply the same sampling strategy as our full models and adopt four denoising steps.