Alias-Free Generative Adversarial Networks

By Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, Timo Aila. 2021

2,077 citations recorded by the source corpus.

We observe that despite their hierarchical convolutional nature, the synthesis process of typical generative adversarial networks depends on absolute pixel coordinates in an unhealthy manner. This manifests itself as, e.g., detail appearing to be glued to image coordinates instead of the surfaces of depicted objects. We trace the root cause to careless signal processing that causes aliasing in the generator network. Interpreting all signals in the network as continuous, we derive generally applicable, small architectural changes that guarantee that unwanted information cannot leak into the hierarchical synthesis process. The resulting networks match the FID of StyleGAN2 but differ dramatically in their internal representations, and they are fully equivariant to translation and rotation even at subpixel scales. Our results pave the way for generative models better suited for video and animation.

Subfields: Computer Vision and Pattern Recognition, Generative Models

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Generative Models

Citations in the corpus

Make-A-Scene: Scene-Based Text-to-Image Generation with Human Priors (2022, background): “Generative Adversarial Networks (GANs) [17] facilitate the generation of high ﬁdelity images [29, 3, 30, 56] in multiple domains by simultaneously training a generator network G and a discriminator network D , where G is trained to fool D , while D is trained to judge if a given image is real or…”
Generative Adversarial Networks (2021, neutral): “Recently a more advanced styleGAN architecture has been proposed [32, 18, 33].”
On Buggy Resizing Libraries and Surprising Subtleties in FID Calculation (2021, background): “With the proliferation of generative modeling techniques, such as Generative Adversarial Networks (GANs) [28], accurately discerning which methods are performing better has become a critical aspect of the field.”
GAN Inversion: A Survey (2021, methodology): “Recent advances in GANs, such as DCGAN [38], WGAN [44], PGGAN [11], BigGAN [12], StyleGAN [13], StyleGAN2 [14], StyleGAN2-Ada [70], and StyleGAN3 [71] have developed better architectures, losses, and training schemes.”
StyleGAN-XL: Scaling StyleGAN to Large Diverse Datasets (2022, methodology): “We first introduce the main building blocks of our system: the StyleGAN3 generator [Karras et al. 2021] and Projected GAN’s [Sauer”
Vocos: Closing the gap between time-domain and Fourier-based neural vocoders for high-quality audio synthesis (2023, background): “Traditional methods often use transposed convolutions that can introduce aliasing artifacts, necessitating additional measures to mitigate the issue (Karras et al., 2021; Lee et al., 2022).”
BigVGAN: A Universal Neural Vocoder with Large-Scale Training (2022, methodology): “This side effect can be suppressed by applying a low-pass filter [17].”
Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (2024, positive): “…transformers akin to GPT-2 and VQ-GAN [49, 19], with the only modification of substituting traditional layer normalization for adaptive normalization (AdaLN) – a choice motivated by its widespread adoption and proven effectiveness in visual generative models [33, 34, 32, 57, 56, 29, 46, 12].”
GET3D: A Generative Model of High Quality 3D Textured Shapes Learned from Images (2022, methodology): “Since many real-world datasets lack camera poses, we follow GANverse3D [34] and utilize pretrained 2D StyleGAN to generate a realistic car dataset.”
HyperInverter: Improving StyleGAN Inversion via Hypernetwork (2021, methodology): “Specially, our work can be further improved by applying iterative reﬁnement [3], multi-layer identity loss [49], or extending to StyleGAN3 [26].”
Creating, Using, Misusing, and Detecting Deep Fakes (2022, methodology): “The synthetic faces were generated using the state-of-the-art StyleGAN2 (Karras et al. 2020), ensuring diversity across gender, race, and apparent age.”
Blended Diffusion for Text-driven Editing of Natural Images (2021, background): “The most impressive results in text-driven image manipulation leverage the strong generative capabilities of modern GANs [6,20,26–28].”
Spherical Fourier Neural Operators: Learning Stable Dynamics on the Sphere (2023, background): “This is not the case in the discrete setting, however, as non-linear functions may introduce arbitrary frequency components, which makes them only approximately equivariant (Karras et al., 2021).”
Self-Consuming Generative Models Go MAD (2023, methodology): “, convolutional traces [40] or aliasing artifacts [41]) that may be reinforced by self-consuming generators.”
Bridging CLIP and StyleGAN through Latent Alignment for Image Editing (2022, neutral): “Besides, we have also integrated our method on Style-GAN3 [17] and StyleGAN-XL [34] for experiments on animal face dataset AFHQ [6, 7] and ImageNet [10].”