Projected GANs Converge Faster

By Axel Sauer, Kashyap Chitta, Jens Muller, Andreas Geiger. 2021

308 citations recorded by the source corpus.

Generative Adversarial Networks (GANs) produce high-quality images but are challenging to train. They need careful regularization, vast amounts of compute, and expensive hyper-parameter sweeps. We make significant headway on these issues by projecting generated and real samples into a fixed, pretrained feature space. Motivated by the finding that the discriminator cannot fully exploit features from deeper layers of the pretrained model, we propose a more effective strategy that mixes features across channels and resolutions. Our Projected GAN improves image quality, sample efficiency, and convergence speed. It is further compatible with resolutions of up to one Megapixel and advances the state-of-the-art Fr\'echet Inception Distance (FID) on twenty-two benchmark datasets. Importantly, Projected GANs match the previously lowest FIDs up to 40 times faster, cutting the wall-clock time from 5 days to less than 3 hours given the same computational resources.

Subfields: 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 Generative Models

Citations in the corpus

StyleGAN-XL: Scaling StyleGAN to Large Diverse Datasets (2022, methodology): “Lastly, [Sauer et al. 2021] apply differentiable data-augmentation [Zhao et al. 2020] before F which improves Projected GAN’s performance independent of the dataset size.”
Novel Class Discovery without Forgetting (2022, background): “Over the last decade, deep learning algorithms have achieved remarkable performances on multiple computer vision tasks [15,37,50,7,45], even outperforming humans on many of them.”
StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models (2023, background): “[ 56] shows that: where D ∗ SLM ( x ) is the optimal discriminator, T and G represent true and generated data distributions, while P T and P G are their respective densities.”
StyleGAN-T: Unlocking the Power of GANs for Fast Large-Scale Text-to-Image Synthesis (2023, methodology): “L G ] 23 J a n 2023 similarly behind diffusion models in ImageNet (Deng et al., 2009; Dhariwal & Nichol, 2021) synthesis until the discriminator architecture was redesigned in StyleGAN-XL (Sauer et al., 2021; 2022), which allowed GANs to close the gap.”
FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space (2025, unknown): “We tackle both challenges using latent adversarial diffusion distillation (LADD) [47, 48, 49], reducing the number of sampling steps while increasing the quality of the samples through adversarial training.”
Fast High-Resolution Image Synthesis with Latent Adversarial Diffusion Distillation (2024, unknown): “ADD leverages the Projected GAN paradigm [46], i.e., applying independent discriminators on features obtained from pretrained features network.”
Spacing Loss for Discovering Novel Categories (2022, background): “Availability of large amount of annotated data has fueled unprecedented success of deep learning in various machine learning tasks [5, 9, 18, 30, 36, 37].”
Improved Distribution Matching Distillation for Fast Image Synthesis (2024, methodology): “In ADD [23], the generator, initialized with weights from a diffusion model, is trained using a projected GAN objective with an image-space classifier [34].”
Scaling up GANs for Text-to-Image Synthesis (2023, unknown): “We further leverage off-the-shelf pretrained models as a loss function [36, 60, 64].”
Adversarial Diffusion Distillation (2023, methodology): “Following [57, 59], we use the hinge loss [32] as the adversarial objective function.”
Diffusion-GAN: Training GANs with Diffusion (2022, methodology): “As a result, a wide array of analyses and modifications has been proposed for GANs, including improving the network architectures (Karras et al., 2019; Radford et al., 2016; Sauer et al., 2021; Zhang et al., 2019), gaining theoretical understanding of GAN training (Arjovsky and Bottou, 2017; Heusel ”
Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (2024, unknown): “First, there have been several approaches in generative adversarial network (GAN; Goodfellow et al. 2014) that try to accelerate training with better convergence using pretrained visual encoders (Sauer et al., 2021; Kumari et al., 2022; Sauer et al., 2022; 2023a; Kang et al., 2023).”
Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (2024, methodology): “We calculate the FID on CLIP features as proposed by (Sauer et al., 2021).”
Improved Techniques for Training Consistency Models (2023, methodology): “Here we exclude methods based on FastGAN (Liu et al., 2020; Sauer et al., 2021) or StyleGAN-XL (Sauer et al., 2022) from our consideration, because both utilize ImageNet pre-trained feature extractors in their discriminators.”
Musika! Fast Infinite Waveform Music Generation (2022, unknown): “We use Cross Channel Mixing (CCM), ﬁrst introduced in [36], to randomly mix channels of the stereo stacked latent vectors before being fed to the discriminator.”
Multi-Concept Customization of Text-to-Image Diffusion (2022, methodology): “To adapt with just a few training images, efficient training techniques [28,33,61,68,83,84] are often useful.”
The Role of ImageNet Classes in Fréchet Inception Distance (2022, methodology): “[49] by training a Projected FastGAN [31, 49], which uses an ImageNet pre-trained EfficientNet [55] as a feature extractor of the discriminator, and compare it against StyleGAN2 in FFHQ.”