Alias-Free Generative Adversarial Networks

Tero Karras; Miika Aittala; Samuli Laine; Erik Härkönen; Janne Hellsten; Jaakko Lehtinen; Timo Aila

Source-linked AI summary

Alias-Free Generative Adversarial Networks

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

arXiv:2106.12423v4Jun 23, 2021cs.CVcs.AIcs.LGcs.NEstat.ML

TL;DR

Typical GANs can attach fine detail to fixed pixel coordinates rather than object surfaces. This paper redesigns generator signal processing to suppress aliasing, matching StyleGAN2’s FID while producing more natural motion and different internal representations.

Problem
Typical GANs do not synthesize images through natural hierarchical refinement because fine detail is often fixed in pixel coordinates rather than precisely positioned by coarse features.
Method
The paper interprets network signals continuously and redesigns signal processing to suppress aliasing and enforce subpixel translation and optional rotation equivariance.
Results
The StyleGAN3 generator matches StyleGAN2 in FID, solves texture sticking, and produces more natural motion.
Takeaways & Limitations
The resulting coordinate systems attach detail to object surfaces and may benefit image and video editing.
Takeaways & Limitations
The architecture assumes training data are not badly aliased, and violating this assumption may cause training difficulties.

Abstract

from arXiv · show

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.

1 Introduction

The introduction identifies aliasing as a key source of positional side information that causes details to stick to image coordinates rather than transform hierarchically with objects. It proposes continuous-domain, bandlimited signal processing to suppress aliasing, yielding StyleGAN3 with altered internal representations and StyleGAN2-level FID.

Problem: StyleGAN2 exhibits “texture sticking,” with details remaining sharp at fixed pixel coordinates instead of moving coherently with the generated content.Averaging nearby latent samples should produce uniformly blurry images, but StyleGAN2 retains sharp details such as fur; latent interpolations likewise reveal coordinate-fixed detail.
Problem: Aliasing lets GAN generators access leaked pixel-grid information, providing positional references that bypass natural hierarchical image synthesis.The introduction identifies non-ideal upsampling and nonlinearities as two aliasing sources, alongside borders, per-pixel noise, and positional encodings.
Approach: The proposed solution treats network signals as continuous bandlimited functions represented on discrete grids and suppresses aliasing through signal-processing-based architectural changes.This framework is used to eliminate unwanted positional side information and restore more natural hierarchical refinement.
Results: StyleGAN3 matches StyleGAN2 in FID while producing coordinate systems that attach details to underlying surfaces and promise improved suitability for video and animation.The resulting generator is slightly heavier computationally, and its internal representations change drastically after aliasing is suppressed.

2 Equivariance via continuous signal interpretation

The section reinterprets discretely sampled feature maps as encodings of underlying bandlimited continuous signals, making spatial equivariance a continuous-domain commutation property. It derives frequency constraints and operation-specific conditions for preserving translation and rotation equivariance, with nonlinearities requiring explicit control of newly generated frequencies.

Continuous signal interpretation: Bandlimited continuous signals, rather than pixel-grid values, are treated as the actual feature maps, while discrete arrays serve only as their encodings.The representation uses ideal interpolation and sampling under the Nyquist–Shannon framework.
Equivariance conditions: An operation is equivariant when it commutes with the spatial transformation in the continuous domain, while its output must remain within the output bandlimit for faithful discrete representation.Rotation additionally requires a disc-shaped spectral limit of radius s/2 in both dimensions.
Convolution: Convolution introduces no new frequencies and is translation-equivariant; rotation equivariance additionally requires a radially symmetric discrete kernel, including viable 1×1 kernels.These properties follow from continuous convolution and the bandlimit-preserving interpretation.
Upsampling and downsampling: Ideal upsampling preserves the continuous signal while increasing sampling rate, whereas downsampling must low-pass filter before coarser discretization to preserve faithful representation.Translation equivariance follows automatically for both operations; rotation-equivariant downsampling requires a radially symmetric, disc-shaped frequency response.
Nonlinearity: Nonlinearity is the only primitive operation that generates novel frequencies, so a lower-cutoff reconstruction filter can precisely control the new information introduced by each generator layer.This filtering provides explicit control over the range of frequencies added before final discretization.

3 Practical application to generator network

The StyleGAN2 generator is converted into a translation- and rotation-equivariant architecture through Fourier-feature inputs, signal-processing changes, and controlled layer modifications. These changes improve equivariance, although boundary and upsampling improvements incur a 16% FID compromise in config E.

3 Practical application to generator network: The practical goal is to make every synthesis layer equivariant to continuous translations and rotations, so finer details transform with coarser local features; the discriminator remains unchanged.The generator is evaluated step-by-step through configurations in Figure 3.
Input representation: Replacing StyleGAN2’s learned 4×4 constant with fixed Fourier features enables exact continuous input transformations and slightly improves FID.The frequencies are sampled uniformly within the circular band fc = 2 and define a spatially infinite map.
Input representation: Removing per-pixel noise supports a natural transformation hierarchy but is approximately FID-neutral and does not improve equivariance alone.The intended hierarchy requires each feature’s exact sub-pixel position to be inherited from underlying coarse features.
Normalization and simplification: Config D restores FID to original StyleGAN2 levels while slightly improving translation equivariance through normalization before each convolution and removal of several StyleGAN2 components.The changes also reduce mapping-network depth, disable mixing and path-length regularization, and eliminate output skip connections.
Boundaries and upsampling: Config E improves translation equivariance but compromises FID by 16% after adding a 10-pixel canvas margin and replacing bilinear upsampling with a windowed sinc filter.Smaller resampling filters hurt equivariance, whereas larger filters mainly increase training time.
Alias-free generator architecture: The final alias-free generator combines Fourier features, normalization, modulated convolutions, and filtered nonlinearities, with new layer specifications further improving translation equivariance.Config T eliminates remaining artifacts, while the number of layers can remain fixed across output resolutions; Figure 4 depicts the architecture.

4 Results

Across six datasets, StyleGAN3-T and StyleGAN3-R achieve FID results competitive with StyleGAN2 and perform equally well by FID, while accompanying videos show that the texture-sticking problem is solved. The results also show that alias-free networks develop phase-encoding feature maps and positional patterns relevant to synthesizing surface detail.

Results: StyleGAN3-T and StyleGAN3-R remain competitive with StyleGAN2 in FID across six datasets and perform equally well by that metric.The evaluation includes standard and unaligned FFHQ and METFACES, properly resampled AFHQ, and BEACHES; Figure 5 presents the dataset results.
Results: The accompanying videos show that StyleGAN3 solves texture sticking and produces more natural motion that better sustains the illusion of a coherent 3D scene.
Ablations and comparisons: Re-enabling noise inputs or using StyleGAN2’s original layer specifications significantly compromises equivariances, whereas mixing regularization can be re-enabled without ill effects.Styles can also be mixed reliably without explicit mixing regularization.
Ablations and comparisons: Lanczos filtering remains competitive in FID but compromises rotation equivariance, Gaussian filtering produces clearly worse FIDs, and doubling feature capacity improves FID at almost 4× training time.Halving capacity preserves high-level equivariances but causes considerable FID degradation.
Internal representations: In StyleGAN3, some internal feature maps encode phase rather than signal magnitude, with StyleGAN3-R showing more well-defined emergent positional encoding patterns for synthesizing surface detail.These representations provide the coordinate system needed to synthesize detail on object surfaces.

5 Limitations, discussion, and future work

The alias-free generator has limitations tied to discriminator equivariance and assumptions about training data, while several extensions remain open. The approach also raises societal risks because removing video artifacts may make generated content more convincing or deceptive.

Limitations: Discriminator non-equivariance may limit motion consistency, as FFHQ teeth fail to move correctly when the head turns, possibly because pixel locations are favored [59].The authors performed equivariant modifications only in the generator and suspect the discriminator causes this failure.
Limitations: Badly aliased training data can cause difficulties because the alias-free generator makes implicit assumptions about the nature of its training data.Point-sampled black-and-white cartoons produce mostly black-or-white pixels and jagged edges, a regime difficult for GANs generally.
Future work: Future work could reintroduce hierarchical-synthesis-consistent noise, improve path-length regularization, and extend equivariance to scaling, anisotropic scaling, or arbitrary homeomorphisms.The proposed path-length regularization should encourage neighboring features to move together rather than discourage their movement altogether.
Future work: Making attention layers equivariant may require temporarily increasing resolution, but attention complexity is challenging and tokenizing transformers such as VQGAN may conflict with equivariance [68] [18].Whether tokenizing-transformer GANs can be made equivariant remains an open question.
Societal impacts: Alias-free video generation can intensify societal harms by making disinformation more convincing or deceptive, including fake portraits and propaganda videos.Suggested mitigations include model watermarking and large-scale authenticity assessment on major social media sites [67].

Appendices · A Additional results

Additional results show that StyleGAN3 matches StyleGAN2’s FID across the tested datasets while substantially improving motion-consistent detail and preserving comparable training dynamics. Supplementary videos and analyses also demonstrate practical equivariance quality, semantic style mixing, and a BEACHES-specific limitation plausibly linked to removing noise inputs [34].

A Additional results: StyleGAN2 and StyleGAN3-T/R achieve comparable FIDs on FFHQ-U, METFACES-U, AFHQV2, and BEACHES without truncation.Uncurated sample sets for all four datasets are provided in Figures 7–10.
A Additional results: BEACHES shows a possible StyleGAN3 limitation: the generators appear to produce fewer scene layouts properly, potentially because removing noise inputs forces capacity toward random-number generation [34].This explanation is presented as a suspicion, and the authors leave reintroducing noise without breaking equivariance as an open direction.
A Additional results: Interpolation videos show StyleGAN3 details moving with depicted objects, unlike StyleGAN2 details that often remain glued to pixel coordinates in METFACES and AFHQV2.The examples include brushstrokes, cracked paint, and fur motion.
A Additional results: StyleGAN3 produces visibly more coherent facial motion than StyleGAN2 in FFHQ cinemagraphs, where StyleGAN2 exhibits jarring sticking of facial hair and skin texture.The cinemagraphs mimic small-scale head movement and facial animation using latent walks and StyleCLIP global directions [25] [52] [46].
A Additional results: EQ-T videos indicate that PSNR values above ∼50 dB look high-quality, while EQ-R values around ∼40 dB look good.These videos illustrate the practical relevance of the equivariance PSNR measurements reported in Figures 3 and 5.
A Additional results: Style mixing regularization continues to work similarly in StyleGAN3, with coarse layers mainly controlling pose and shape and fine layers mainly controlling coloring and finer facial features.Hand-picked examples show imperfect but recognizable layer-specific inheritance; mixing regularization was not used for those examples [34].
A Additional results: StyleGAN3’s translation and rotation equivariance does not appear to significantly alter training dynamics, whose convergence curves remain similar to those of Karras et al. [34] [32].Figure 12 compares the main translation- and rotation-equivariant configurations T and R with prior results.
A Additional results: Supplementary analyses include windowed, whitened average power spectra to reduce border-induced FFT artifacts, plus an animated visualization of the generator nonlinearity.The power-spectrum procedure removes the sharp image-border step that creates an axis-aligned cross artifact.

B Datasets … C.4 Two-dimensional filters

The paper introduces unaligned and rebuilt datasets tailored to alias-free training, then specifies Kaiser-based one- and two-dimensional filters for generator resampling. These filters preserve sampling geometry, control attenuation, and use radial symmetry where rotation equivariance requires it.

B.1 FFHQ-U and MetFaces-U: FFHQ-U and MetFaces-U are rebuilt unaligned variants that retain natural camera and head-tilt variation through axis-aligned, randomly shifted crop rectangles.The modified datasets use original raw images and facial-landmark metadata, preserve the original image counts and specifications, and include reproducibility code in the public release.
B.2 AFHQv2: AFHQv2 corrects pixel-level downsampling artifacts by rebuilding the dataset with Lanczos resampling, leaving 15803 images after minor technical exclusions.The original artifacts caused convergence problems because their sharp stair-step aliasing was difficult to reproduce without direct access to the pixel grid.
B.3 Beaches: BEACHES contains 20155 licensed 512×512 beach photographs, but neither the full training data nor pretrained models can be redistributed.The training images were provided by Getty Images, and the dataset is proprietary.
B.4 Licenses: The datasets carry source-specific licenses: FFHQ uses Creative Commons BY-NC-SA 4.0 alongside the listed image licenses, METFACES uses Creative Commons BY-NC 2.0 with CC0 images, and original AFHQ uses Creative Commons BY-NC 4.0.These licenses are attributed respectively to NVIDIA Corporation, the Metropolitan Museum of Art, and NAVER Corporation where specified.
C Filter details; C.1 Kaiser low-pass filters; C.2 Selecting window parameters: The filter-details section builds practical low-pass filters from ideal responses using Kaiser windows, whose spatial extent and transition-band choices trade off against stopband attenuation.After discretization, the taps are explicitly normalized to sum to one, reducing cumulative scaling errors across consecutive layers.
C.3 Upsampling and downsampling: For upsampling by factor m, the design increases the tap count proportionally, preserving input-relative spatial extent and compensating for reduced transition-band attenuation.Even tap counts intentionally shift sample locations by 1/(2s′), matching the interpretation of each output group as subdividing one input interval; downsampling follows by filtering and discarding m−1 samples per group.
C.4 Two-dimensional filters: The two-dimensional construction is separable for ordinary filtering but replaces the non-radially symmetric response with a jinc-based radial low-pass filter when rotation-equivariant behavior is required.The radial filter uses a jinc function while retaining a separable windowing scheme to preserve spectral properties.
C.4 Two-dimensional filters: The final filter design uses separable two-dimensional filters for efficient upsampling and most downsampling, while config R uses a radially symmetric alternative to preserve rotation-equivariant cutoff behavior.The separable filter has a square frequency response and is used for all upsampling filters and downsampling in configs A–T; config R instead uses the radial filter except in its final two critically sampled layers.

C.5 Alternative filters … D.2 Optimizations for common upsampling factors

The paper evaluates alternative reconstruction filters and implements filtered nonlinearities efficiently through a fused CUDA kernel, specialized gradients, and optimizations for common sampling factors. Kaiser-filter comparisons use matched cutoff assumptions, while the CUDA design reduces memory traffic and redundant computation.

C.5 Alternative filters: Kaiser filters are compared with Lanczos and Gaussian alternatives using an implicit cutoff frequency of 0.5 and impulse-response scaling for comparable evaluation.The prototypical kernels otherwise lack an explicit cutoff-frequency control.
C.5 Alternative filters: The comparison sets the Lanczos extent to a = 2 and Gaussian standard deviation to σ = 0.4, choices found to work reasonably well across tested values.Gaussian support is truncated at |x| > 8/s to limit computation, with negligible practical effect because the response is already near zero.
C.5 Alternative filters: The Gaussian filter is radially symmetric and therefore suitable for rotation equivariance, whereas the Lanczos alternative uses the same separable filter in rotation-equivariant configuration R.Lanczos lacks a widely accepted radially symmetric counterpart.
D Custom CUDA kernel for filtered nonlinearity: A custom CUDA kernel fuses input, upsampling, nonlinearity, and downsampling, keeping intermediate computations in on-chip shared memory and writing only the final output.The kernel tiles output feature maps across CUDA thread blocks; separable filters use sequential 1D convolutions, whereas non-separable filters use a single 2D convolution.
D.1 Gradient computation: Gradient computation exploits the operation’s near self-adjoint structure by reversing phases and swapping the upsampling and downsampling parameters and filters.The remaining challenge is the nonlinearity evaluated at upsampled resolution.
D.1 Gradient computation: Leaky ReLU gradients conserve memory by storing only input-sign information, while 16-bit execution additionally records clamping so clamped outputs receive zero gradients.This specialization avoids storing the full high-resolution nonlinear input needed by a naïve general solution.
D.2 Optimizations for common upsampling factors: Figure 14 benchmarks the optimized kernel against native PyTorch across sampling-factor combinations and separable or non-separable filters, including forward and gradient-related cases.The tested configurations include 2×/2×, 4×/2×, and 2×/4× upsampling/downsampling combinations on an NVIDIA Titan V GPU.
D.2 Optimizations for common upsampling factors: For 2× and 4× upsampling, computing multiple neighboring outputs together removes redundant shared-memory loads, reducing instruction counts by nearly 25% and up to 37.5%, respectively.These optimizations are applied to both separable and non-separable filters.

E Equivariance metrics … E.3 Rotation

The paper defines PSNR-based equivariance metrics for integer and fractional translations and arbitrary rotations, exposing both their sampling procedures and limitations from subpixel and rotational resampling. The rotation metric addresses anisotropic discrete-image bandlimits by filtering before and after continuous rotation, with a practical finite-support approximation.

E.1 Integer translation: EQ-T measures PSNR agreement between a translated generator output and an output regenerated after translating its input Fourier features, averaging 50,000 random latent-code and integer-offset samples.Offsets are sampled uniformly from [-s_N/8, s_N/8] in each image dimension, and comparison is restricted to the mutually valid region.
E.2 Fractional translation: EQ-T can converge to ∞dB for a perfectly integer-translation-equivariant generator, but it ignores subpixel behavior and can therefore miss aliasing-related failures.This limitation applies even to generators that are perfectly equivariant under integer translations but are not properly bandlimited.
E.2 Fractional translation: EQ-Tfrac extends translation testing to continuous offsets, using Lanczos-resampled reference images and a reduced valid region that accounts for the filter’s spatial extent.The metric samples offsets uniformly over the same range and normalizes the discretized Lanczos filter to enforce partition of unity.
E.2 Fractional translation: EQ-T and EQ-Tfrac agree reasonably up to approximately 40 dB, after which EQ-Tfrac saturates below 50 dB because subpixel resampling trades off aliasing, ringing, and high-frequency retention.The authors conclude from these results that configurations G–R are essent…
E.3 Rotation: Arbitrary-rotation equivariance cannot be measured perfectly beyond approximately 40 dB because rotation resampling is ambiguous and discrete-image bandlimits are not radially symmetric.A 45° rotation changes the rectangular input bandlimit into a diamond-shaped spectrum, creating both excess frequencies and missing diagonal frequencies.
E.3 Rotation: EQ-R compares continuous Fourier-feature rotations with high-quality reference-image rotations and pseudo-rotations, using angles uniformly sampled from 0° to 360°.The pseudo-rotation modifies frequency content as if the image had been rotated without actually rotating it, while comparisons use a valid region constrained by both filter footprints.
E.3 Rotation: The ideal rotation operator reconstructs a continuous image, rotates it, applies output bandlimiting, and resamples it; its rotation filter is the polygonal intersection of input and rotated rectangular spectra.This filtering is required before and after rotation to eliminate aliasing, although the result still lacks the highest representable diagonal frequencies.
E.3 Rotation: The practical rotation approximation replaces the ideal filter with a finite-support Lanczos filter, upsamples by m = 4, and uses bilinear lookups to approximate rotated sampling.The pseudo-rotation is implemented as a convolution with a related discrete filter.

F Implementation details

The alias-free generator was implemented by modifying StyleGAN2-ADA while retaining most of its training and architectural details. Experiments used a specified DGX-1 setup and official Inception-based FID evaluation, with code and pretrained models released publicly.

F Implementation details: The implementation builds on the official PyTorch StyleGAN2-ADA codebase and preserves discriminator architecture [34], weight demodulation [34], equalized learning rates [31], minibatch standard deviation [31], exponential moving averages [31], and mixed-precision training [32].
F Implementation details: Experiments ran on an NVIDIA DGX-1 with 8 Tesla V100 GPUs using PyTorch 1.7.1, CUDA 11.0, and cuDNN 8.0.5.FID was computed between 50k generated images and all training images using the official pretrained Inception network.
F Implementation details: The implementation and pretrained models are publicly available in the official StyleGAN3 repository.

F.1 Generator architecture

The generator architecture stabilizes signal magnitudes, enables per-image rotations and translations of Fourier features, and flexibly schedules layer filter parameters across sampling resolutions. These changes analytically preserve the corresponding continuous geometric transformations and constrain frequency content through critically sampled layers.

Normalization: The architecture tracks a long-term exponential moving average of input signal magnitudes to eliminate layerwise drift after removing StyleGAN2 [34] output skip connections.Without this correction, signal magnitudes can drift, increasing run-to-run variation and occasionally causing mixed-precision numerical issues.
Transformed Fourier features: An additional affine layer maps w to t = (r_c, r_s, t_x, t_y), enabling each image to apply a learned rotation and translation to its input Fourier features.The transformation is initialized as t = (1, 0, 0, 0), then its normalized components define rotation about the canvas center followed by translation.
Transformed Fourier features: The resulting geometric transformation is implemented analytically by modifying Fourier-feature phases and two-dimensional frequencies, equivalently transforming the continuous representation of z_0.Normalization by the first two components makes the transformation independent of w’s magnitude, analogous to weight modulation and demodulation [34].
Flexible layer specifications: Flexible configurations define cutoff and stopband frequencies by geometric progression until the first critically sampled layer, then derive sampling rates and transition bands from them.For N = 14 layers and N_crit = 2, sampling rates are rounded up to powers of two satisfying s[i] ≥ 2f_t[i] without exceeding output resolution.

F.2 Hyperparameters and training configurations · F.3 G-CNN comparison

Training used baseline-derived hyperparameters with dataset-specific augmentation, transfer learning, and capacity adjustments, while config R applied progressively removed initial blur. The G-CNN comparison extended config T with p4-symmetric convolutions, providing equivariance only to quarter-turn rotations rather than arbitrary angles.

F.2 Hyperparameters and training configurations: All runs used 8 GPUs and trained until the discriminator saw 25M real images from scratch or 5M with transfer learning.Baseline runs used StyleGAN2 config F [34] for high-resolution datasets and the ADA 256×256 baseline config [32] for ablations.
F.2 Hyperparameters and training configurations: Most hyperparameters remained inherited unchanged from baseline configurations, while configs C and D disabled noise, path-length, and mixing regularization.Config D also reduced mapping-network depth to 2 and set the minibatch standard-deviation group size to 4, following the StyleGAN2-ADA recommendation.
F.2 Hyperparameters and training configurations: All experiments enabled horizontal-flip augmentation; ADA used default settings except on FFHQ and FFHQ-U, while METFACES datasets used transfer learning from the best corresponding FFHQ snapshots.The augmentation choices reflected horizontal symmetry and aimed to reduce discriminator overfitting.
F.2 Hyperparameters and training configurations: Generator capacity followed StyleGAN2’s inverse-resolution feature-map rule, with baseline settings using Cmax = 512 and Cbase = 214 or 215 depending on output resolution.The capacity parameters controlled the generator’s overall width across resolutions.
F.2 Hyperparameters and training configurations: R1 regularization weight γ was selected through dataset-dependent grid searches for baseline and transfer configurations, whereas low-resolution ablations fixed γ = 1.Baseline config B searched γ ∈ {1, 2, 5, 10, 20}; configs T and R searched γ0 ∈ {0.0002, 0.0005, 0.0010, 0.0020, 0.0050} under γ = γ0 · N/M.
F.2 Hyperparameters and training configurations: Config R blurred discriminator inputs with Gaussian noise starting at σ = 10 pixels and ramping to zero over the first 200k images.The blur was applied before ADA to discourage early overemphasis on high frequencies and could delay the generator’s high-frequency learning.
F.3 G-CNN comparison: Config R was compared with config T augmented by p4-symmetric group convolutions, which provide equivariance to 0°, 90°, 180°, and 270° rotations but not arbitrary angles.The implementation added a size-4 group dimension to intermediate synthesis activations and introduced corresponding redundancy in convolution weights.

G Energy consumption

The project’s computational effort and electricity consumption were measured in GPU-years and megawatt hours, respectively. It consumed approximately 225 MWh, with about 70% used for exploratory runs.

G Energy consumption: The project’s computational effort and electricity consumption were tracked in a detailed breakdown using single-GPU years and electricity usage.The computation unit was GPU-years on a single NVIDIA V100 GPU, following Green500 power-measurement guidelines.
G Energy consumption: Approximately 92 years of execution on a single NVIDIA V100 GPU would have been required to reproduce the project’s computational effort.The project used varying numbers of NVIDIA DGX-1 systems across different stages.
G Energy consumption: 225 MWh of electricity was consumed by the entire project, approximately 70% during exploratory runs.Exploration included developing new configurations, intermediate StyleGAN2-to-generator configurations, and key parameter ablations.