On the Frequency Bias of Generative Models
Authors: Katja Schwarz, Yiyi Liao, Andreas Geiger
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To achieve this, we first independently assess the architectures of both the generator and discriminator and investigate if they exhibit a frequency bias that makes learning the distribution of high-frequency content particularly problematic. Based on these experiments, we make the following four observations: 1) Different upsampling operations bias the generator towards different spectral properties. 2) Checkerboard artifacts introduced by upsampling cannot explain the spectral discrepancies alone as the generator is able to compensate for these artifacts. 3) The discriminator does not struggle with detecting high frequencies per se but rather struggles with frequencies of low magnitude. 4) The downsampling operations in the discriminator can impair the quality of the training signal it provides. |
| Researcher Affiliation | Academia | Katja Schwarz Yiyi Liao Andreas Geiger Autonomous Vision Group University of Tübingen and MPI for Intelligent Systems {firstname.lastname}@uni-tuebingen.de |
| Pseudocode | No | The paper does not contain any sections or figures explicitly labeled 'Pseudocode' or 'Algorithm'. |
| Open Source Code | Yes | We release our code and dataset at https://github.com/autonomousvision/frequency_bias. |
| Open Datasets | Yes | We further test our setting on natural images with a downsampled version of Celeb A [21]. [...] We train our version of PGAN on a downsampled version of FFHQ [18] at resolution 642 pixels and a downsampled version of 200k images from LSUN Cats [36] at resolution 1282 pixels. We finetune Style GAN2 on LSUN Cats (2562 pixels), AFHQ Dog [6] (5122 pixels) and FFHQ (10242 pixels). |
| Dataset Splits | No | The paper mentions using '90% and 10% for training and evaluation' for spectral classification, but does not explicitly specify a separate validation dataset split. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU models, CPU types, or memory specifications) used for running experiments. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | We choose the generator from PGAN [16]... We reduce the number of channels for faster training... As upsampling operations, we investigate bilinear and nearest neighbor interpolation, zero insertion, and reshaping... The generator is optimized to reconstruct the corresponding image with a pixel-wise L2-loss... We optimize the learnable tensors and discriminator weights in an alternating fashion... We train from scratch with R1-regularization and without progressive growing... until the discriminator has seen 15M images. We finetune pre-trained models... until the discriminator has seen 2.5M images. |