---
kind: concept
slug: generative-adversarial-network
title: Generative adversarial network (GAN)
summary: >-
  A generator and a discriminator trained against each other, or a pretrained GAN reused without adversarial training. Twenty-nine method cards: translation, text, music, video, MRI, face editing and fitting, age progression, distillation, tabular oversampling, view synthesis. Thirteen use a GAN only as a component, benchmark, survey or metric. Numbers are author-reported.
tags: [gan, generative-models, adversarial-training, image-translation, text-generation, text-to-image, point-cloud, domain-adaptation, image-enhancement, tabular-data, benchmark, music-generation, mri-reconstruction, blind-super-resolution, source-free-adaptation, gan-inversion, image-inpainting, age-progression, distillation, view-synthesis, social-navigation, video-generation, 3d-face-reconstruction, image-editing, point-tracking, evaluation-metrics, graph-neural-networks, voice-conversion, data-imputation]
papers: ["1611.02200", "1612.05362", "1710.07346", "2102.04942", "1908.10962", "2008.02401", "2108.00946", "2204.08583", "1709.08624", "1609.05473", "1904.01480", "1904.12304", "1905.03198", "1912.01865", "1802.01886", "2111.02813", "2502.12048", "1611.06355", "2111.11843", "2607.19153", "1611.09904", "1709.00753", "1909.06581", "1911.11544", "2106.05744", "2304.02858", "2502.19316", "1702.08423", "2311.17042", "1812.01717", "1912.08804", "1803.01164", "2411.01814", "2103.04922", "2112.14683", "2302.02913", "1902.05978", "2305.10973", "2103.09396", "2101.10025", "2011.06170", "2105.14762"]
related: [domain-adaptation, auto-context-model, image-to-image-translation, medical-image-synthesis, text-guided-image-editing, latent-space-editing, gan-inversion, recurrent-neural-network, vision-language-pretraining, adversarial-text-generation, policy-gradient, contrastive-learning, conditional-batch-normalization, semantic-segmentation, spectral-normalization, actor-critic, audio-deepfake-detection, adaptive-instance-normalization, benchmark-validity, image-enhancement, vision-transformer, channel-attention, data-augmentation, hyperparameter-optimization, random-forest, image-super-resolution, data-consistency-layer, ensemble-learning, entropy-minimization, residual-learning, neural-style-transfer, knowledge-distillation, diffusion-models, generative-model-evaluation, novel-view-synthesis, differentiable-rendering, monocular-depth-estimation, deep-learning, convolutional-neural-network, motion-prediction, video-generation, implicit-neural-representation, variational-autoencoder, normalizing-flows, face-recognition, dense-correspondence, graph-neural-network, time-series-forecasting, frechet-inception-distance, algorithmic-fairness]
status: draft
updated_at: 2026-11-01
---

## Overview

A GAN trains two networks at once. The discriminator D learns to tell real
samples from generated ones, and the generator G learns to produce samples D
accepts as real. The baseline objective is a max over D of the expected log
score on real samples plus the log of one minus the score on generated ones,
with G minimising the same quantity [arxiv:1611.02200 §3, eq. 1]
[arxiv:1710.07346 §3, eq. 1].

Eighteen cards train a generator against a discriminator or critic: unpaired
translation, restyling of labelled aerial images, paired MR-to-CT synthesis,
two-stage fashion synthesis, multi-domain translation with style codes, stacked
text-to-image synthesis, motion generation, long text generation, sequence text
generation by policy gradient, a latent GAN for point-cloud completion, a
conditional GAN with trained encoders for face editing, a transformer generator
for underwater enhancement, one tabular GAN per failure type, a recurrent GAN
for continuous music events, a residual GAN for MRI reconstruction, a patch GAN
trained on one image to estimate its blur kernel, a class-conditional GAN
for source-free adaptation, and a refinement GAN that fills the holes left by
a point-cloud renderer for single-image view synthesis. Most of them condition G on an input or a label
rather than on noise alone. The fashion model
also takes noise
[arxiv:1710.07346 §3.1]. The music card's generator takes noise at every step
and no image or label input [arxiv:1611.09904 §2] [§3]. The kernel card's
generator takes the image itself, and the image is the only data it sees
[arxiv:1909.06581 §2] [§5.3]. The text generator in the long-text card has no input
beyond a start token and its own prefix [arxiv:1709.08624 §Methodology]. The
SeqGAN generator has the same kind of input, and its discriminator scores
whole sequences only [arxiv:1609.05473 §Method]. The
latent GAN in the shape card takes only a one-dimensional seed, which a separate
agent chooses from the scan [arxiv:1904.12304 §3.2] [§3.3]. Of the five
translation-style cards, four put the adversarial term beside reconstruction,
identity, cycle or style terms. The fourth is StarGAN v2
[arxiv:1912.01865 eq. 5]. The fashion card gives no extra term beyond the GAN objective
[arxiv:1710.07346 §3.4]. The motion card weights the adversarial term at 0.1
against its reconstruction losses [arxiv:2102.04942 §3.7.3]. The text-to-image
card adds a contrastive term to the adversarial losses it inherits from AttnGAN
[arxiv:1904.01480 §3.1] [§4.1]. The underwater card adds its adversarial term
with no coefficient, so weight one, beside LAB, LCH, RGB and perceptual terms
[arxiv:2111.11843 eq. 13]. The IcGAN card trains its encoders with squared
losses and no adversarial term [arxiv:1611.06355 eq. 3] [eq. 4]. The tabular
card trains each generator as a WGAN-GP and then filters its samples with a
separate classifier [arxiv:2607.19153 §3.3.1].

Three more cards also train a discriminator. The age-progression card puts
discriminators on both its latent code and its output image
[arxiv:1702.08423 eq. 3-4]. The distillation card trains a student against a
diffusion teacher, with a discriminator on frozen features of the student's
output [arxiv:2311.17042 §3.2]. The video card judges whole clips: its
discriminator reads k sampled frames of one clip, three by default, and scores
them jointly [arxiv:2112.14683 §3.2]. All three are described below.

Six more cards use GANs differently. The minimax transport card has a
two-player game, but its inner player is a convex potential, not a real-or-fake
classifier [arxiv:1908.10962 §3]. The StyleFlow and VQGAN+CLIP cards reuse a
pretrained GAN generator and train no discriminator
[arxiv:2008.02401 §6.2] [arxiv:2204.08583 §2]. The StyleGAN-NADA card also
trains no discriminator, but it does update a pretrained StyleGAN, with a
CLIP-based directional loss as its only training signal [arxiv:2108.00946 §4.1].
Two inversion cards also use a pretrained StyleGAN without a discriminator. The
Image2StyleGAN++ card keeps the generator frozen and optimises a W+ code and
then per-layer noise maps for one photo [arxiv:1911.11544 §5] [§4.2]. The PTI
card freezes the W code of one face, then fine-tunes the whole generator so
that it reproduces that face, with a locality penalty on nearby codes
[arxiv:2106.05744 §3.2] [§3.3]. The two cards therefore differ in whether the
generator's weights change.

Two more cards fit a pretrained GAN with no discriminator of their own. The
3D face card optimises a 3D face model and a texture latent against one photo,
through the GAN's output [arxiv:1902.05978 §3.1]. The point-editing card
optimises a StyleGAN2 latent so that handle points move to their targets
[arxiv:2305.10973 §3.2]. Both are described below.

Nine cards use a GAN without making it the method. Texygen is a benchmark
harness. It packages seven text generators, six of them GAN variants and one
maximum-likelihood baseline, with five metrics [arxiv:1802.01886 §2.1] [§2.2].
WaveFake builds a detection
dataset whose fake speech comes from several generators, some of them GAN-based
vocoders, and its detectors are statistical models and a CNN-GRU, not GANs
[arxiv:2111.02813 §2.2] [§3]. The EEG survey lists GAN decoders among its image
decoders, next to VAE and latent diffusion, and reports no new results of its
own [arxiv:2502.12048 §3] [§4.5]. The deep-learning survey has one section on
GANs, which gives the minimax objective and the usual remark that the generator
loss can saturate early. Its printed non-saturating term omits the discriminator,
so take the exact form from the original GAN paper [arxiv:1803.01164 §VII.A].
The generative-model survey covers energy-based, variational, adversarial,
autoregressive and flow models, and it compares them in one star-rated table
on CIFAR-10. Its FID and NLL values come from the cited works, and it lists
BigGAN at 14.73 and StyleGAN2 with ADA at 2.42 [arxiv:2103.04922 tab. 1]. The
table is not like-for-like, and its GAN loss rows mix losses to minimise with
losses to maximise, so take exact forms from the original papers
[arxiv:2103.04922 fig. 5a] [tab. 1].
The augmentation benchmark compares nine
augmenters, one of them CT-GAN, with nine ensemble models on binary and
multi-class imbalanced sets [arxiv:2304.02858 §4.1] [§5.1]. The CT-GAN row of
its table is one arm of that comparison, not a method the paper proposes
[arxiv:2304.02858 tab. 1]. The social navigation card uses a Social GAN
generator as a pedestrian forecaster inside a robot planner. Its own
contribution is extra cost terms for the planner, and the paper does not say
which data the forecaster was trained on [arxiv:2411.01814 §2.1] [eq. 5–8].
Two reviews name GANs among several generative models. The graph-neural-network
review describes graph GANs as one kind of graph generative network, with a
min-max loss over node pairs. It does not give their architectures, and it
quotes one synthetic-feeder comparison in which they reach a 27.2% connected
rate, against 8.3% for a random-matrix method. The dataset is not named
[arxiv:2101.10025 §II-E] [§III-D2] [tab. VI]. The GAN-evaluation review updates
the measures used for GANs, from FID variants to precision-recall extensions
and human judgement. It runs no experiments and quotes its numbers from other
papers [arxiv:2103.09396 §2] [§Results].

One card is an evaluation metric rather than a method. The Fréchet Video
Distance card does not train a GAN. It compares Gaussian fits of video features
for generators in general, and GANs are cited there as models that lack ground
truth for a direct comparison [arxiv:1812.01717 §2]. Identical sets have a
nonzero distance, so the card says to compare only at equal sample counts
[arxiv:1812.01717 §4.2].

Another card judges generative models for engineering design. It sorts its
metrics into similarity, novelty and diversity, constraint, performance and
conditioning groups [arxiv:2302.02913 §Method]. Its running example is one GAN
and one VAE on six-mode 2-D data [arxiv:2302.02913 §4.5]. There the GAN has the
lower maximum mean discrepancy, 0.013 against 0.045, and the higher constraint
satisfaction, 0.833 against 0.713. The VAE has the higher precision-recall F1,
0.475 against 0.386 [arxiv:2302.02913 tab. 2] [tab. 4]. The paper runs one
instance per model and tests no difference for significance [§Pitfalls]. The
same nearest-datapoint score is minimised as a realism measure and maximised as
a novelty measure, so the direction depends on the question [arxiv:2302.02913
tab. 1] [§5.1]. The released code covers the 2-D examples only [§Paper vs code].

Two more cards use a GAN as one part of a larger system. The partial
multi-view card's core model, CPM-Nets, has no adversarial term. Its CPM-GAN
variant adds one discriminator per view, trained on observed values as
positives, to fill in missing entries [arxiv:2011.06170 §3] [alg. 2]. The
released code has no GAN part [arxiv:2011.06170 §Paper vs code]. The emotional
speech card is mainly a dataset paper. It benchmarks two of its own converters,
CycleGAN-EVC and VAWGAN-EVC, on one speaker [arxiv:2105.14762 §5.1]. The GAN
is a baseline there, not the contribution [arxiv:2105.14762 §2] [§3].

The numbers below are author-reported, on different data and with different
metrics. The page therefore compares design choices rather than one benchmark.

## Variants

### Unpaired translation with a frozen feature function (DTN)

The Domain Transfer Network learns a map from a source domain to a target
domain using unpaired samples from both. A fixed, pretrained function f
supplies the representation. G is built as a small trainable network g
applied after f, so only g decodes f's features [arxiv:1611.02200 §4]. The
discriminator has three classes: generated from source, generated from target,
and real target. The generator must push both generated classes towards the
real-target class [arxiv:1611.02200 eq. 3, eq. 4]. Two further terms keep the
output faithful. One asks f to give the same features before and after
translation for source samples. The other asks G to leave target samples
unchanged [arxiv:1611.02200 eq. 5, eq. 6]. A total-variation smoothness term is
also defined, but the digit setting sets its weight to zero [arxiv:1611.02200
§5.1]. Card: [Unsupervised Cross-Domain Image Generation](/papers/1611.02200).

On SVHN to MNIST, the full model reaches 90.66%. A plain binary GAN with
feature constancy reaches 13.71% in the same table. Removing the identity
term gives 88.40%, and removing the constancy term gives 74.55%
[arxiv:1611.02200 tab. 1]. On faces, translated emoji give a median rank of
16 for identity retrieval, against 16311 for hand-made emoji. The face
experiment uses a non-public face set and a non-public feature network
[arxiv:1611.02200 tab. 4, §5.2].

### Cycle-consistent restyling of labelled source images (aerial segmentation)

The aerial card adapts a segmenter trained on one labelled aerial dataset to
another, with no target labels. It trains a translator in both directions on
unpaired patches from the two datasets. Each direction has its own adversarial
loss, and a reconstruction term asks each round trip to return its input
[arxiv:1905.03198 §4.1] [eq. 3] [eq. 8]. The source labels are then reused for
the translated source images, and the source segmenter is fine-tuned on them
[arxiv:1905.03198 §4.2] [§4.3]. The segmenter is BiSeNet with a ResNet-101
feature extractor [arxiv:1905.03198 §5.2.1]. Card: [Unsupervised Domain
Adaptation using Generative Adversarial Networks for Semantic Segmentation of
Aerial Images](/papers/1905.03198).

On Potsdam to Vaihingen, average accuracy goes from 0.35 before adaptation to
0.52 after [arxiv:1905.03198 tab. 3]. Building accuracy goes from 0.23 to 0.71
and Tree from 0.06 to 0.51. Low vegetation falls from 0.38 to 0.27
[arxiv:1905.03198 tab. 4]. The discussion calls the other classes unchanged
[arxiv:1905.03198 §6], which the Low vegetation row does not support. The target
model is scored on the Vaihingen test set after every epoch, and the reported
gain is taken after eight epochs [arxiv:1905.03198 §5.2.4]. That choice uses
target labels, so the best-epoch numbers are optimistic. The authors'
repository has no license file [arxiv:1905.03198].

### Paired 3D conditional GAN with auto-context (MR to CT)

The second variant is supervised. A 3D fully convolutional generator maps a
32³ MRI patch to the matching 16³ CT patch, and a 3D CNN discriminator judges
real and generated CT patches [arxiv:1612.05362 §2.1, §4]. The generator loss
combines a weighted adversarial term, an L2 term against the true CT, and a
3D gradient-difference term. The weights are 0.5, 1 and 1 [arxiv:1612.05362
eq. 5, eq. 7] [arxiv:1612.05362 §4]. A stack of such GANs is trained in
sequence with the Auto-Context Model. Each stage also receives the previous
stage's CT estimate as an extra input channel. The paper uses two such stages
[arxiv:1612.05362 §2.2, §4.2]. Card: [Medical Image Synthesis with
Context-Aware Generative Adversarial Networks](/papers/1612.05362).

On 16 brain and 22 pelvis subjects, under leave-one-out evaluation, the
author-reported PSNR is 27.6 for brain and 34.1 for pelvis. The strongest
baseline, structured random forest with auto-context, reports 26.3 and 32.1
[arxiv:1612.05362 tab. 1, tab. 2]. The baseline numbers are copied from that
method's paper rather than re-run [arxiv:1612.05362 §4]. In an ablation
where the dataset is not stated, adding the adversarial term raises PSNR from
24.7 to 25.9 over a generator trained alone [arxiv:1612.05362 §4.1].

### Paired transformer generator with colour-space losses (underwater enhancement)

The U-shape Transformer keeps the paired image-to-image layout of pix2pix, an
encoder, a bottleneck and a decoder, and changes two parts. A channel-wise
attention block (CMSFFT) replaces the skip connections. It attends across
colour channels, using the four encoder scales as its inputs [arxiv:2111.11843
§3.2.3] [eq. 4] [eq. 7]. A four-layer spatial transformer (SGFMT) sits at the
bottleneck [arxiv:2111.11843 §3.2.2] [eq. 1] [eq. 3]. The generator loss adds
LAB and LCH colour losses, an RGB L2 term and a VGG perceptual term to a GAN
term [arxiv:2111.11843 eq. 13]. Its LSUI dataset has 4279 pairs of underwater
photos and reference images, and 3879 of them are used for training
[arxiv:2111.11843 §3.1] [§4.1]. Card: [U-shape Transformer for Underwater Image Enhancement](/papers/2111.11843).

Author-reported PSNR on Test-L400 is 24.16, against 22.91 for the next best
method, Ucolor, with SSIM 0.93 against 0.89 [arxiv:2111.11843 tab. 3]. On
Test-U90 the PSNR gap is 2.13 dB, and on Test-L400 it is 1.25 dB, so the
abstract's "more than 2 dB" holds only for one of the two sets [arxiv:2111.11843
tab. 3]. Adding the channel block to a pix2pix baseline raises Test-L400 PSNR
from 19.34 to 22.47, and the spatial block raises it to 21.78 [arxiv:2111.11843
tab. 6]. The ablation has no run without the adversarial term, so the paper
does not isolate what the GAN contributes [arxiv:2111.11843 tab. 6]. The
paper says the source photos were licensed for academic use only, so check
that before training on LSUI for a commercial product [arxiv:2111.11843 §3.1].

### One generator with style codes for many domains (StarGAN v2)

StarGAN v2 uses one generator G(x, s) for all domains, so it does not need one
translator per pair of domains. The style code s comes either from a mapping
network, which reads random noise and a target domain, or from a style encoder,
which reads a reference image and its domain. Both have one output head per
domain, so every domain has its own style space [arxiv:1912.01865 §2.1] [app.
E]. The generator injects the style into each decoder block through adaptive
instance normalization [arxiv:1912.01865 §2.1]. The discriminator has one
real-or-fake output per domain [arxiv:1912.01865 §2.1].

Training adds three terms to the adversarial loss [arxiv:1912.01865 eq. 5]. A
style-reconstruction term asks the encoder to recover the style that was used
[eq. 2]. A diversity term asks two different style codes to give different
images [eq. 3]. A cycle term translates the output back with the input's own
estimated style, so pose and layout survive [eq. 4]. The paper drops the
denominator from the original diversity-sensitive loss, which it says made
training unstable, and it decays the diversity weight linearly to zero
[arxiv:1912.01865 §2.2] [app. B]. Card: [StarGAN v2: Diverse Image Synthesis for
Multiple Domains](/papers/1912.01865).

On CelebA-HQ, the author-reported latent-guided FID is 13.7, against 33.1 for
MSGAN and 31.4 for MUNIT [arxiv:1912.01865 tab. 2]. On AFHQ it is 16.2, against
61.4 for MSGAN [tab. 2]. Reference-guided FIDs are 23.8 on CelebA-HQ and 19.8
on AFHQ, against 39.6 and 69.8 for MSGAN [arxiv:1912.01865 tab. 3]. In the
ablation, the style code and diversity term bring the CelebA-HQ FID from 98.4
for StarGAN to 13.7 [arxiv:1912.01865 tab. 1]. In a human study, 68.9% of
votes for quality went to StarGAN v2 on CelebA-HQ and 88.1% on AFHQ
[arxiv:1912.01865 tab. 4]. The study has 76 valid workers and no significance
test [arxiv:1912.01865 §3.2]. The paper's real-image FID on CelebA-HQ is 14.8,
and StarGAN v2's 13.7 is below it. The card warns against reading that as
beating real images, because the test set is small [arxiv:1912.01865 tab. 2].

The official code is CC BY-NC 4.0, which rules out commercial use. Its loss
uses a mean where the paper writes an L1 norm, so the λ values need the same
convention to keep their balance [arxiv:1912.01865 §Paper vs code]. Implement
from the paper, and check the AFHQ image licences yourself, because the paper
lists none per image [arxiv:1912.01865 §Pitfalls].

### Two-stage conditional GAN through a label map (text-guided fashion)

The fashion model redresses a photo of a person from a text description and
keeps the wearer's pose and body shape. It is two conditional GANs trained
in sequence. The first generates a 7-label segmentation map. Its input is a
coarse 8×8 map of the wearer in which all clothing is merged into one class,
plus a design code built from attributes and the text [arxiv:1710.07346 §3.1,
§3.2, eq. 2]. The second generates one image channel per label, and each pixel
takes its value from the channel of its label [arxiv:1710.07346 eq. 3, eq. 4,
§3.3]. The paper's motivation is that plain text-conditioned GANs lose the wearer's
structure and blur region boundaries [arxiv:1710.07346 §1]. The two stages are trained separately, because
the argmax between them is not differentiable. Both discriminators also receive
the conditions [arxiv:1710.07346 §3.4]. Training needs one photo per person,
with one description and its segmentation map. No pairs of the same person in
two outfits are used [arxiv:1710.07346 §3.1]. Card: [Be Your Own Prada: Fashion
Synthesis with Structural Coherence](/papers/1710.07346).

The author-reported attribute mAP is 82.6, against 71.8 to 74.7 for the
baselines and 86.8 on real images [arxiv:1710.07346 tab. 1]. A two-stage model
without the per-label mapping reaches 74.7 in the same table, which isolates the
mapping's contribution. In a 50-person user study, the full method has the best
mean rank [arxiv:1710.07346 tab. 2]. The paper reports no FID, and its user
study has no significance test [arxiv:1710.07346 §4.3].

### Conditional autoencoder with discriminators on code and image (age progression)

The age-progression model is a conditional autoencoder, not a plain GAN. An
encoder maps a face to a personality vector z. A generator takes z and a
10-bin age label and produces the image at that age. Two discriminators are trained alongside it.
One pushes the distribution of z towards a uniform prior, and the other judges
generated images conditioned on the age label [arxiv:1702.08423 eq. 3-4]. No
paired examples of the same person at two ages are needed
[arxiv:1702.08423 §1]. Card: [Age Progression/Regression by Conditional
Adversarial Autoencoder](/papers/1702.08423).

The training set is about 10,670 images [arxiv:1702.08423 §5.1]. The
loss weights are λ = 100 and γ = 10 [arxiv:1702.08423 §5.2]. The author-reported
user preference is 52.77% for this method against 28.99% for prior work
[arxiv:1702.08423 §5.3], and the paper has no face-recognition measure of
identity. The paper's vote counts exceed the arithmetic maximum for its panel,
so read the percentages with care [arxiv:1702.08423 §5.3]. The paper gives
its reconstruction term as L2, but the released code uses L1
[arxiv:1702.08423 §5.2]. The total-variation formula, the adversarial weights and
the weight initialisation are not stated.

### Stacked text-to-image GAN with a Siamese contrastive term (SD-GAN)

SD-GAN is a text-to-image GAN with three generator and discriminator stages.
Each caption of a pair goes through its own branch. At every stage, a contrastive loss pulls
the two discriminator feature vectors together when the captions describe the
same image and pushes them apart when they describe different images
[arxiv:1904.01480 §3.1] [§4.3]. The pull stops at a floor distance and the push
at a margin [arxiv:1904.01480 eq. 2]. The generator's batch norm layers are
replaced by semantic-conditioned batch norm (SCBN). Its scale and shift are
learned values plus a term that comes either from the sentence vector or from
word-level attention [arxiv:1904.01480 §3.2] [eq. 4] [eq. 5] [eq. 6]. Card:
[Semantics Disentangling for Text-to-Image Generation](/papers/1904.01480).

The author-reported Inception score is 4.67 on CUB, against 4.36 for AttnGAN,
and 35.69 on MS-COCO, against 25.89 [arxiv:1904.01480 tab. 1]. Each part helps
on its own. Without the Siamese term the CUB score is 4.49, without SCBN it is
4.51, and without both it is 4.11 [arxiv:1904.01480 tab. 4]. In a 50-rater
ranking study, SD-GAN was placed first in 68.76% of CUB rankings and 75.78% of
COCO rankings, against 20.54% and 17.69% for AttnGAN [arxiv:1904.01480 tab. 2].
The paper tried four values of the floor and 0.1 scored best on both datasets
[arxiv:1904.01480 tab. 5]. Its own tables disagree on one setting: a no-Siamese,
SCBN-with-word-cue run is 4.49 on CUB in one table and 4.45 in another
[arxiv:1904.01480 tab. 4] [tab. 6]. No official code was found. The
optimiser, learning rate, batch size and contrastive weight are not stated, and
the adversarial losses are inherited from AttnGAN [arxiv:1904.01480 §4.1].

### Least-squares critics over motion sequences (in-betweening)

The motion model fills the frames between seed frames and a target keyframe.
An LSTM generator produces the transition. Two feed-forward critics score
10-frame windows for long-term structure and 2-frame windows for short-term
structure [arxiv:2102.04942 §3.5]. They are trained with a least-squares
objective [arxiv:2102.04942 §3.6.2]. The generator's adversarial term has weight
0.1 beside L1 reconstruction losses [arxiv:2102.04942 §3.7.3]. The conditioning
relies on two additive embeddings: a sinusoidal time-to-arrival code and a
target noise that is switched off within 5 frames of the target
[arxiv:2102.04942 §3.3, §3.4]. Card: [Robust Motion
In-betweening](/papers/2102.04942).

The adversarial term is the smallest of the parts. On Human3.6M walking, the
average L2Q goes from 0.84 to 0.82 when it is added, and NPSS goes from 0.46 to
0.40 [arxiv:2102.04942 tab. 2]. On LaFAN1, the reconstruction-only generator is
worse than interpolation on L2P and NPSS at 45 frames [arxiv:2102.04942 tab. 3].
Inference takes about 0.30 s per transition on a CPU, and the plugin uses 170 MB
[arxiv:2102.04942 tab. 4]. Training runs for 55,000 iterations on Human3.6M and
350,000 on LaFAN1 [arxiv:2102.04942 §4.2, §4.3].

### Sequence generator trained by policy gradient with roll-out rewards (SeqGAN)

SeqGAN treats a text generator as a reinforcement-learning policy. The
generator is an LSTM with a softmax output, and each next token is an action
taken given the prefix [arxiv:1609.05473 §Method]. The discriminator is a CNN
with several kernel widths, max-over-time pooling and one highway layer. It
outputs the probability that a finished sequence is real [arxiv:1609.05473
§Method]. The only reward comes at the end. For an earlier token, its value is
the average discriminator score over N roll-outs that keep the prefix and
finish the sequence with a roll-out policy [arxiv:1609.05473 eq. 3] [eq. 4].
The generator then takes a plain gradient-ascent step on the policy gradient
[arxiv:1609.05473 eq. 8]. The discriminator is trained by cross entropy between
real sequences and generated ones [arxiv:1609.05473 eq. 5]. Training starts
with maximum-likelihood pretraining of the generator. Each iteration then takes
g generator steps and d discriminator steps, and the discriminator's negatives
are regenerated at every step [arxiv:1609.05473 alg. 1]. Card: [SeqGAN:
Sequence Generative Adversarial Nets with Policy Gradient](/papers/1609.05473).

The author-reported oracle NLL on synthetic data is 8.736, against 9.038 for MLE
[arxiv:1609.05473 tab. 1]. On real text, SeqGAN is above MLE on each metric
the paper reports. On Chinese poems the human score is 0.5356 against 0.4165,
and BLEU-2 is 0.7389 against 0.6670 [arxiv:1609.05473 tab. 2]. On Obama
speeches, BLEU-4 is 0.427 against 0.416 [arxiv:1609.05473 tab. 3]. On folk
tunes, the BLEU-4 is 0.9406 against 0.9210 [arxiv:1609.05473 tab. 4]. On
the music task the text says SeqGAN is better on both metrics, so lower MSE is
read as better: 20.62 against 22.38 [arxiv:1609.05473 tab. 4].
MLE pretraining needs about 150 epochs to converge. A 20-epoch pretraining
run makes adversarial training slow and unstable [arxiv:1609.05473 §Results]
[app. More Ablation Study] [fig. 4].

The reward has no baseline subtracted. The paper does not discuss the variance
that results [arxiv:1609.05473 §Pitfalls]. Each generator step at length 20
with 16 roll-outs runs 304 partial roll-outs and 320 discriminator passes per
batch of 64 sequences, in the official code [arxiv:1609.05473 §Pitfalls]. The
code is not licensed, and its update order and roll-out copy differ from the
paper's Algorithm 1 [arxiv:1609.05473 §Paper vs code]. The card cites it by
path only.

### Hierarchical text generator guided by discriminator features (LeakGAN)

LeakGAN targets long text, where one reward at the end of a sentence gives too
little signal. The discriminator is a CNN that reads the sentence so far and
exposes its intermediate features. A Manager LSTM maps those features to a
unit-length goal at each step. A Worker LSTM picks each next word under a goal
embedding built from the last c goals [arxiv:1709.08624 §Methodology] [eq. 2]
[eq. 4]. The paper sets c to 4 and the goal size to 16 [arxiv:1709.08624
§Experiment] [app. Experiment Settings]. The Manager learns from the change in
the features over the next c steps, weighted by a Monte Carlo estimate of the
expected reward [arxiv:1709.08624 eq. 7]. The Worker learns from an intrinsic
reward, the mean cosine between recent feature changes and the goals
[arxiv:1709.08624 eq. 8] [eq. 9]. Before adversarial training, the paper
pretrains the discriminator and generator by maximum likelihood [arxiv:1709.08624
§Training of G] [alg. 1]. Card: [Long Text Generation via Adversarial Training
with Leaked Information](/papers/1709.08624). The authors' code differs from the
paper in several settings, so the card lists the differences.

On synthetic data, the reported negative log-likelihood at length 20 is 7.038,
against 8.736 for SeqGAN and 8.247 for RankGAN. At length 40 it is 7.191,
against 10.310 and 9.958 [arxiv:1709.08624 tab. 1]. On WMT News, BLEU-2 is
0.956, against 0.859 for SeqGAN [arxiv:1709.08624 tab. 2]. On COCO captions,
BLEU-5 is 0.686, against 0.544 for RankGAN [arxiv:1709.08624 tab. 3]. In a
Turing-style test on WMT News, 55.4% of LeakGAN sentences were judged real,
against 23.6% for SeqGAN and 65.1% for real text [arxiv:1709.08624 tab. 5]. The
paper reports p < 10⁻⁶ against the second-best method in its main tables
[arxiv:1709.08624 §Evaluation Metrics].

### Minimax convex potentials for transport maps

The optimal-transport card learns the quadratic-cost transport map from
samples. Two input convex networks, f and g, play a minimax game, and the map
is the gradient of g [arxiv:1908.10962 §3, alg. 1]. The page includes it
because the paper is tagged as a GAN and compares against a GAN-based
transport method, W2GAN [arxiv:1908.10962 §4.1]. Its inner player scores
transported points as a convex potential, not as real or fake
[arxiv:1908.10962 §3]. Card: [Optimal transport mapping via input convex neural
networks](/papers/1908.10962).

The relative error of the transported mean, scaled by 100 and squared, is 0.02,
0.07 and 0.04 for shifts of 1, 5 and 10 in 784 dimensions [arxiv:1908.10962
tab. 1]. On 2-D toy data, the paper argues that its map does
not leave points between separated target modes, which continuous GAN-style
generators do [arxiv:1908.10962 §4.1, §4.3, fig. 3]. The bound on the map error
needs f to be strongly convex, and the paper does not say how to enforce that
[arxiv:1908.10962 §3.2]. Training differentiates through the gradient of g,
which needs second-order autodiff [arxiv:1908.10962 app. C.1].

### Conditional GAN with trained encoders for attribute editing (IcGAN)

IcGAN first trains a conditional GAN, then fits two encoders that invert it
[arxiv:1611.06355 §4]. The GAN follows DCGAN and uses a matching-aware
discriminator, which sees image and label pairs [arxiv:1611.06355 §4.3]. After
training, E_z maps a generated image back to its noise vector, and E_y maps a
real image to its attribute vector [arxiv:1611.06355 eq. 3] [eq. 4]. An edit
keeps the noise vector, which the paper links to pose and lighting, and
replaces the attributes [arxiv:1611.06355 §5.4]. Of the three ways to couple
the two encoders, two independent networks (IND) gives the lowest
reconstruction loss [arxiv:1611.06355 §4.1] [§5.3]. Card: [Invertible
Conditional GANs for image editing](/papers/1611.06355).

Author-reported reconstruction loss L_e on generated CelebA images is 0.429 for
IND, 0.432 for the variant in which E_z also reads the attribute estimate, and
0.500 for a shared encoder [arxiv:1611.06355 §5.3]. Edit quality is reported
only qualitatively, apart from attribute-classifier accuracy of 74% to 90%
across GAN configurations [arxiv:1611.06355 tab. 3] [§5.2]. That classifier is
described only by its architecture. The authors placed the attribute input at
the generator input and at the first discriminator layer [arxiv:1611.06355
§5.2]. The paper says the encoders use the GAN's optimiser settings, but the
released code uses others, so reproduce from the card's listed values
[arxiv:1611.06355 §4.3].

### Pretrained GAN reused for attribute editing (StyleFlow)

StyleFlow keeps a pretrained StyleGAN2 frozen. It trains a conditional
continuous normalizing flow between the StyleGAN W latent and a standard
normal, with 17 attribute values as the condition [arxiv:2008.02401 §5.1,
§5.2]. Sampling runs the flow in reverse with the target attributes
[arxiv:2008.02401 §6.1]. Editing runs the flow forward under the source
attributes, reverses it under the target attributes, and writes the new vector
into only the W+ rows that the edit covers. The other rows keep the source
values [arxiv:2008.02401 §6.2]. Training
maximises a likelihood, not an adversarial objective [arxiv:2008.02401 §5.3,
eq. 8]. Card: [StyleFlow](/papers/2008.02401).

On sequential edits of generated faces, the FID is 53.15, against 64.69 for
GANSpace and 67.08 for InterFaceGAN [arxiv:2008.02401 tab. 1]. On real images,
the paper says StyleRig scores better on the projection itself
[arxiv:2008.02401 §7.7]. Training uses 10,000 samples, 10 epochs and a batch
size of 5 [arxiv:2008.02401 §5.1, §5.3]. The attribute labels come from an
external face-attribute service, so reproduction needs equivalent labels
[arxiv:2008.02401 §5.1]. The authors' code is marked non-commercial, so the
card's notes apply to any use of it [arxiv:2008.02401].

### Text-only domain shift of a StyleGAN with a directional CLIP loss (StyleGAN-NADA)

StyleGAN-NADA moves a pretrained StyleGAN2 to a new domain from a source text
and a target text, with no images from the target domain [arxiv:2108.00946
§1, §4]. Two copies of the generator share one frozen mapping network. The
frozen copy gives a source image for each latent code, and the trained copy is
updated so that the CLIP-space change between the two images points along the
change from the source text to the target text [arxiv:2108.00946 §4.1]. The
loss is one minus the cosine between those two directions [arxiv:2108.00946
eq. 2]. The authors reject a global CLIP loss on the trained generator, which
they report fails and leads to adversarial pixel patterns and mode collapse
[arxiv:2108.00946 §4.1] [app. B] [fig. 15]. For larger shape changes, each
iteration updates only the k layers that change most under a short latent
optimisation [arxiv:2108.00946 §4.2]. The paper gives k as all layers for
texture changes, two thirds of the layers for small shape changes and three
for animal changes [arxiv:2108.00946 app. I] [tab. 3]. A latent mapper is
added for large shape changes, where the adapted generator can miss the target
[arxiv:2108.00946 §4.3]. Card: [StyleGAN-NADA: CLIP-Guided Domain Adaptation of
Image Generators](/papers/2108.00946).

The FID rows in the card's table are for few-shot methods started from the
zero-shot generator, not for the zero-shot method alone. For dog to cat with 10
target images, starting from the zero-shot generator moves Ojha et al. from
45.13 to 41.22, MineGAN from 79.31 to 51.81 and TGAN from 87.11 to 54.49
[arxiv:2108.00946 tab. 2]. With 100 images for Ojha et al., the FID goes from
27.54 to 16.17 [arxiv:2108.00946 tab. 2]. A short discriminator-only catch-up
phase gives the most consistent gains across the three ways tried
[arxiv:2108.00946 tab. 1]. In a 2-alternative user study of 1,200 responses from
218 users, the zero-shot cat generator is reported as preferred over most 5-
and 10-shot baselines on quality. The text gives no percentages, so read the
per-method values from the figure [arxiv:2108.00946 §5.1] [fig. 12]. Texture
changes take about 300 iterations, roughly 3 minutes on one V100. Animal
changes take 2000 iterations plus a latent mapper, about 6 hours in total
[arxiv:2108.00946 app. I].

### Text-guided latent optimisation with a frozen VQGAN decoder (VQGAN+CLIP)

VQGAN+CLIP trains no network. It optimises the latent grid of a pretrained
VQGAN decoder with Adam, so that CLIP scores the decoded picture closer to the
prompt [arxiv:2204.08583 §2]. Each step averages a squared spherical distance
over many augmented crops [arxiv:2204.08583 §2, §2.3]. Editing differs from
generation only in the starting latent, which is an encoded input image
[arxiv:2204.08583 §2]. The paper's settings are a learning rate of 0.15 and
400 iterations [arxiv:2204.08583 §2]. Card: [VQGAN-CLIP](/papers/2204.08583).

Human raters gave a mean of 4.6 out of 5 on 12 prompts, against 3.3 for the best
GLIDE baseline. The baselines were best-of-five and this method was best-of-one
[arxiv:2204.08583 tab. 1, §3.2]. One image takes 188 s on a V100 and training
costs zero GPU-hours [arxiv:2204.08583 tab. 2]. The paper reports a Colab run
under three minutes that needs under 11 GB of GPU memory [arxiv:2204.08583
§6.1].

### Latent GAN with a learned seed for point-cloud completion (RL-GAN-Net)

RL-GAN-Net completes a partial 3-D point cloud in a latent space. An autoencoder
maps each 2048-point shape to 128 numbers. A latent GAN is trained on those
codes, not on raw points, with the WGAN-GP loss [arxiv:1904.12304 §3.1] [§3.2]
[app.]. The GAN's seed is one-dimensional. A DDPG agent reads the code of a
partial scan and picks the seed, and the generator's output is decoded into a
completed shape [arxiv:1904.12304 §3.3] [alg. 1]. The agent's reward is a
weighted sum of three negated losses: a Chamfer term, a term that matches the
scan's code, and a discriminator term, with weights 100, 10 and 0.01
[arxiv:1904.12304 eq. 5]. Each episode is a single action [arxiv:1904.12304
app.]. Card: [RL-GAN-Net: A Reinforcement Learning Agent Controlled GAN Network
for Real-Time Point Cloud Shape Completion](/papers/1904.12304).

The paper reports about 1 ms per completion. The optimisation baseline is
quoted at 324 s for a batch of 50 shapes [arxiv:1904.12304 §4.1] [§5]. That
figure is taken from the baseline's own paper and was not measured on this
hardware. With 70% of points missing, PointNet classification reaches 82.5%
on the vanilla completions and 83.8% on the hybrid, against 50.2% on the raw
partial input [arxiv:1904.12304 app.] [§5]. The hybrid chooses, per input,
between the GAN output and the autoencoder output using the discriminator. It is
described only in text [arxiv:1904.12304 §3.4]. The official code departs from
the paper in its reward and its Chamfer definition, and it stores every
transition as terminal, so the discount factor has no effect
[arxiv:1904.12304]. The paper's weights were chosen for its own loss scales.

### One tabular GAN per failure type, pooled into the training set

The method oversamples rare machine failures in tabular sensor data. The
failure rows are split by failure type, and one WGAN-GP generator is trained
per type, each with its own Optuna study [arxiv:2607.19153 §3.2.5] [§3.3.2].
The generators come from the tabgan library and use a gradient penalty,
mode-specific normalisation for continuous columns and Gumbel-Softmax for
categorical ones [arxiv:2607.19153 §3.3.1]. Each generator's synthetic failures
are drawn in proportion to its group size and pooled with the real rows. A
separate classifier, LightGBM by default, keeps only the synthetic rows it
labels as real with high confidence [arxiv:2607.19153 §3.2.5] [§3.3.1]. Card:
[Breaking the Homogeneity Assumption: Specialized Multi-Generator Adversarial
Learning for Rare Failure Detection in Predictive Maintenance](/papers/2607.19153).

On AI4I 2020 with a Random Forest, the author-reported PR-AUC is 0.8574. That
is above 0.835 for cost-sensitive learning and 0.8245 for one GAN over all
failures [arxiv:2607.19153 tab. 4] [§4.1]. Recall rises from 0.5882 to 0.8824,
at a precision of 0.5696 [tab. 4]. The F1 is 0.6923, below the baseline's 0.740
[tab. 4]. The SVM baseline's recall of 0.9412 is higher than either multi-GAN row
[tab. 4]. The numbers come from single runs, with no variance or significance
test [tab. 4]. Compute is the main cost: about 14 hours for the multi-GAN and
11 hours 20 minutes for the single GAN, against minutes for the classical
methods [arxiv:2607.19153 §5.1].

### Recurrent GAN with feature matching over continuous music events (C-RNN-GAN)

C-RNN-GAN models music as a sequence of continuous events instead of symbols
from a fixed menu. Each event has four real values: tone length, frequency,
intensity and the time since the previous tone [arxiv:1611.09904 §2.1]. The
generator is a two-layer LSTM with 350 units per layer, and it reads a uniform
random vector at every step [arxiv:1611.09904 §3]. The discriminator is a
two-layer bidirectional LSTM that gives one decision for the whole sequence
[arxiv:1611.09904 §3]. Before the adversarial phase, the generator is trained
for six epochs to predict the next event. In that phase, the generator's loss
is replaced by feature matching on the discriminator's last representation
before its output layer, and D is paused while its loss is below 70% of G's
[arxiv:1611.09904 §3]. Card: [C-RNN-GAN: Continuous recurrent neural networks
with adversarial training](/papers/1611.09904).

The paper has no numeric table. Its comparisons are curves over epochs, which
show more unique tones and a wider tone span than the next-event baseline, and
a higher polyphony score for the three-tones-per-cell variant [fig. 3a–c]. The
authors did no formal listening study [arxiv:1611.09904 §4.1]. The code differs
from the paper in several settings, including an extra rule that also pauses G
and a halved discriminator learning rate, and it has no license file
[arxiv:1611.09904 §Paper vs code].

### Residual GAN with a cyclic data-consistency loss for MRI (RefineGAN)

RefineGAN reconstructs an MRI image from under-sampled k-space in one forward
pass [arxiv:1709.00753 §I]. The generator is a residual convolutional
encoder-decoder. It predicts a correction to the zero-filled image, and the
discriminator reuses its encoder path [arxiv:1709.00753 §III-E]. A cyclic loss
sits beside the adversarial term. Under-sampling the generator's output again
should reproduce the measurements, and a fully sampled image that is
under-sampled and reconstructed should come back close to itself
[arxiv:1709.00753 §III-D] [eq. 6]. The adversarial term is weighted 1, the
measurement term 1 and the image term 10 [arxiv:1709.00753 §III-F] [eq. 7].
A second generator with the same architecture is chained after the first, and
the chain is trained jointly [arxiv:1709.00753 §III-G]. Card: [Compressed
Sensing MRI Reconstruction using a Generative Adversarial Network with a Cyclic
Loss](/papers/1709.00753).

The paper reports 0.106 s for brain and 0.112 s for chest on 256×256 images, and
8.57 s and 9.37 s for CSCMRI [arxiv:1709.00753 tab. 1]. The introduction says
under 100 ms, which its own RefineGAN row does not meet [§I] [tab. 1]. PSNR is
shown only as plots, and the text says it passes 42 dB at 40% sampling
[arxiv:1709.00753 §I] [fig. 8]. The paper states that the cyclic loss adds
about 1 dB over SingleGAN [arxiv:1709.00753 §IV-A]. Its code has no license
and is for research use only, so implement from the paper.

### Blind kernel estimation with a linear GAN trained on one image (KernelGAN)

KernelGAN estimates the downscaling kernel of one low-resolution image. It
trains a small GAN on that image alone at test time, with no training data
[arxiv:1909.06581 §2]. The generator is a deep linear network: five hidden
64-channel layers with no activations, kernel sizes 7, 5, 3, 1 and 1, and a
1×1 output layer with stride 2 [arxiv:1909.06581 §4.1] [fig. 4]. The kernel is
the product of its filters [arxiv:1909.06581 §4.2]. The discriminator is a
patch network that scores each local patch [arxiv:1909.06581 §3]. A regularizer
keeps the kernel's sum near one, sparse, bounded and centred [eq. 3]. The
estimated kernel is then passed to a non-blind method, ZSSR or SRMD
[arxiv:1909.06581 §5.1]. Card: [Blind Super-Resolution Kernel Estimation using
an Internal-GAN](/papers/1909.06581).

Author-reported on 100 DIV2KRK images, KernelGAN with ZSSR reaches 30.363 dB and
0.8669 SSIM at ×2, against 29.368 dB for the earlier estimator with ZSSR, and
26.810 dB at ×4, against 26.080 dB [arxiv:1909.06581 tab. 1]. Each image takes
61 s on a V100 [arxiv:1909.06581 §5.3]. Equation 2 as printed gives D the
opposite objective to the one the text describes, and the code follows the
text [arxiv:1909.06581 §Pitfalls]. The code is for academic and non-commercial
use only, so implement from the paper.

### Class-conditional GAN for source-free domain adaptation (3C-GAN)

3C-GAN adapts a source-trained classifier to unlabeled target images without
using source data [arxiv:2502.19316 §1] [§3]. It has three networks. The
classifier C supplies the reference weights and is also the model being adapted. A class-conditional
generator G(y, z) makes samples for a chosen label. A discriminator D separates
real target images from generated ones, with no pairs [arxiv:2502.19316 §3.1]
[fig. 2]. G is also trained so that C labels its samples as the conditioning
label, through a cross-entropy term [arxiv:2502.19316 eq. 3] [eq. 4]. C is
trained on three terms: cross-entropy on generated samples, a squared-distance
anchor to the source weights [arxiv:2502.19316 eq. 5] [eq. 6], and a clustering
regularizer on target images [arxiv:2502.19316 §3.3] [eq. 9]. Card: [Model
Adaptation: Unsupervised Domain Adaptation without Source Data](/papers/2502.19316).

Author-reported, the paper gives 99.4% on SVHN to MNIST against 76.4% for the
source-only model [arxiv:2502.19316 tab. 1]. On the Office-31 average with
ResNet50 it gives 89.6% against 76.1% [tab. 2], and 81.6% on VisDA17 synthetic
to real with ResNet101 [tab. 3]. Most compared methods use source data during
adaptation [arxiv:2502.19316 §4.2]. The text says the gain on MNIST to MNIST-M is
around 40%, while the table gives 54.2% rising to 98.5%, about 44 points
[§4.2] [tab. 1]. Removing the classifier's cross-entropy term on generated
samples stops convergence [arxiv:2502.19316 §4.4] [tab. 4]. The semantic weight
λ_s is not stated, and no official code was found.

### Adversarial distillation of a diffusion teacher (ADD)

Adversarial Diffusion Distillation trains a few-step student from a pretrained
diffusion model. The student starts from the teacher's U-Net weights
[arxiv:2311.17042 §3]. A hinge-loss discriminator judges the student's samples
on frozen DINOv2 features through lightweight heads, with an R1 penalty of
1e-5 [arxiv:2311.17042 §3.2, eq. 2-3]. A squared-L2 score-distillation term,
weighted by a timestep function, keeps the student close to the teacher, and
the final model uses a variant of it [arxiv:2311.17042 §3.3, eq. 4]. The
adversarial weight is 2.5 [arxiv:2311.17042 §4]. The point of the method is
sampling in 1 to 4 steps without classifier-free guidance [arxiv:2311.17042 §1].
Card: [Adversarial Diffusion Distillation](/papers/2311.17042).

On COCO at 512 px, the ablation reports FID-5k and CLIP score. Adversarial loss
alone gives 20.8 and 0.315, the default gives 20.6 and 0.319, and
score-distillation without the adversarial term gives 22.3 and 0.325. Random
initialisation gives FID 293.6 [arxiv:2311.17042 tab. 1]. The larger 860M student (ADD-M)
reaches FID 19.7 and CLIP 0.326 in one step [arxiv:2311.17042 tab. 2]. The paper also
reports ELO-style preference from figure data, with the one-step large student
at 1009 and the four-step at 1107 [arxiv:2311.17042 §4.2]. The student is less
diverse than its teacher [arxiv:2311.17042 §4.3]. The paper does not state the
teacher's timestep distribution or the student's timesteps 2 and 3, and it
gives no optimiser, learning rate or batch size. The released code has no
training script, and the SDXL-Turbo weights it uses are non-commercial.

### Point-cloud rendering with a GAN refinement network (SynSin)

SynSin renders a new view of a scene from one RGB image and a relative camera
pose. A feature network gives each pixel a 64-number feature, and a depth
network gives each pixel a depth [arxiv:1912.08804 §3.1]. A renderer lifts
these into a point cloud, moves it to the new camera and splats each point as a
soft disc, keeping the K nearest points per pixel [arxiv:1912.08804 §3.2] [eq. 1]
[eq. 2]. Pixels that no point reaches are holes. A refinement network with
eight residual blocks fills them, and two discriminators at different scales
judge its output with a feature-matching term [arxiv:1912.08804 §3.3]. The
generator loss is the GAN term with weight 1, plus L1 with weight 1 and a
perceptual term with weight 10 [arxiv:1912.08804 §3.4]. Training uses image pairs
and their relative poses, with no depth or 3D labels [arxiv:1912.08804 §1]
[§3.4]. Card: [SynSin: End-to-end View Synthesis from a Single Image](/papers/1912.08804).

Author-reported PSNR on RealEstate10K is 22.31 with one input view. The Vox
baseline (its "Vox w/ ours") scores 21.88 in the same table, and 3DView scores
21.88 in the system comparison. StereoMag reaches 25.34, but it takes two input
views at test time [arxiv:1912.08804 tab. 1] [tab. 2] [§4.2]. In a five-rater
study on 100 animated trajectories, SynSin was preferred over Vox w/ ours in
68.7% of either-or votes [arxiv:1912.08804 tab. 4]. On Matterport3D the full
model scores 20.91 PSNR [arxiv:1912.08804 tab. 1]. The ablations in the card
change the depth supervision, the input features, the point count and the
fine-tuning setup, but none removes the GAN term. So the card cannot say how
much the adversarial loss adds [arxiv:1912.08804 tab. 1].

The paper's learning rates conflict with each other and with the code. The text
gives 0.0001 for the generator and 0.01 for the discriminator, and the appendix
gives 0.001 for its own model [arxiv:1912.08804 §3.4] [app. C]. The code's
default refinement network is a UNet, while the released scripts use the ResNet
variant the paper describes [arxiv:1912.08804 §Paper vs code]. The official code
mixes licences, including non-commercial CC BY-NC-SA parts, so implement from
the paper and check the components you reuse.

### Motion-conditioned video GAN on a StyleGAN2 backbone (StyleGAN-V)

StyleGAN-V keeps StyleGAN2 for image synthesis and adds time. A content code,
passed through the StyleGAN2 mapping network, gives each clip its look. A grid
of motion noise is passed through two padding-less one-dimensional
convolutions, and an acyclic sinusoidal embedding turns the result into a
motion code for any real-valued time [arxiv:2112.14683 §3.1] [app. B]. The
motion code is tiled into the constant input of the synthesis network, so
frames are made independently and none feeds the next [arxiv:2112.14683 §3.1].
The discriminator takes k frames of a clip, extracts features from each with a
StyleGAN2-style backbone, joins them at 16 by 16, and is conditioned on the
time gaps between frames [arxiv:2112.14683 §3.2] [fig. 6]. Training frames are
sampled with at most 32 frames of total span [arxiv:2112.14683 §3.3]. The
author-reported FVD16 at 256 squared is 47.41 on FaceForensics, 79.52 on
SkyTimelapse and 1431.0 on UCF101, the lowest in the authors' table on each of
those sets. VideoGPT is lowest on RainbowJelly [arxiv:2112.14683 tab. 1]. On
FaceForensics the authors report FID 9.44 against 8.42 for StyleGAN2
[arxiv:2112.14683 §4.3]. The official code has no LICENSE file, and it differs
from the paper in several places, so the paper's formulas do not give its
behaviour [arxiv:2112.14683 §Paper vs code]. Card: [StyleGAN-V: A Continuous Video Generator with the Price, Image Quality and Perks of StyleGAN2](/papers/2112.14683).

## When to use which

| Situation | Use | Why |
|---|---|---|
| Source and target domains with no paired examples, and a pretrained feature function you can keep frozen | Unpaired translation (DTN) [arxiv:1611.02200 §3, §4] | The generator sits on f, and the paper's digit and face results rely on that constraint [arxiv:1611.02200 tab. 1, tab. 4]. |
| Labelled source images, unlabelled target images, and a shift that comes mainly from the sensor | Cycle-consistent restyling of the source set, then fine-tuning the segmenter on it [arxiv:1905.03198 §4.2] [§4.3] | The paper reports average accuracy from 0.35 to 0.52 on Vaihingen [arxiv:1905.03198 tab. 3]. Target labels pick the reported epoch [arxiv:1905.03198 §5.2.4], so expect a smaller real gain. |
| Paired, voxel-aligned volumes, such as MRI and CT of the same subjects | Supervised conditional GAN with an L2 term [arxiv:1612.05362 §2.1] | The L2 term anchors the output to the true target. The paper reports the adversarial term adds PSNR on top of a generator alone [arxiv:1612.05362 §4.1]. |
| Squared error alone gives blurry output | Add the adversarial term and a gradient-difference term [arxiv:1612.05362 §1, §2.1] | The paper's motivation is that L2 tends to blur. Its GAN output shows fewer artifacts than the plain FCN, judged by eye [arxiv:1612.05362 fig. 4]. |
| Translating along a chain where each step needs context | Auto-Context stacking [arxiv:1612.05362 §2.2] | The paper gains from a second stage [arxiv:1612.05362 §4.2]. Each stage costs a full-volume prediction over the training set. |
| Face-style target domain with a fixed recognition network | DTN with f from that network [arxiv:1611.02200 §5.2] | The paper's face result depends on a DeepFace-style network that is not public [arxiv:1611.02200 §5.2]. Plan for a substitute and expect different numbers. |
| Text-conditioned edits that must keep region structure, with a segmentation map for each training photo | Two-stage GAN through a label map [arxiv:1710.07346 §3.1, §3.3] | Per-label channels score 82.6 mAP against 74.7 without them [arxiv:1710.07346 tab. 1]. Each training photo needs its segmentation [arxiv:1710.07346 §3.1]. |
| Text-to-image training from scratch, where each training image has several captions and a paraphrase should give the same picture | Stacked GAN with a Siamese contrastive term and SCBN [arxiv:1904.01480 §3.1] [§3.2] | The paper reports 4.67 on CUB, against 4.36 for AttnGAN [arxiv:1904.01480 tab. 1]. Each part helps alone [tab. 4]. Optimiser and learning rate are not stated [§4.1], so a search is needed. |
| Sparse conditioning over long gaps, with inference that must run on a CPU inside a tool | LSTM generator with least-squares critics and additive embeddings [arxiv:2102.04942 §3.3, §3.5, §4.4] | The embeddings do most of the work. The adversarial term adds a small gain on L2Q and a larger one on NPSS [arxiv:2102.04942 tab. 2]. |
| You need a transport map that can split a region into separate targets, and no real-or-fake classifier | Minimax convex potentials [arxiv:1908.10962 §3, §4.1] | The paper reports spurious mass avoided on 2-D data, and small transported-mean error in 784 dimensions [arxiv:1908.10962 tab. 1]. The map bound needs strong convexity [arxiv:1908.10962 §3.2]. |
| A pretrained StyleGAN and attribute edits, without retraining the generator | Conditional flow on the W latent [arxiv:2008.02401 §5.2, §6.2] | The paper reports lower FID than the baselines on sequential edits [arxiv:2008.02401 tab. 1]. Its code is non-commercial. |
| A target style or domain described only in text, with no target images, and a StyleGAN you can update | StyleGAN-NADA: directional CLIP loss with adaptive layer freezing [arxiv:2108.00946 §4.1, §4.2] | The paper reports preference over most 5- and 10-shot baselines [arxiv:2108.00946 §5.1]. It says the method works best for style and fine detail and may struggle with large geometry or scale changes [arxiv:2108.00946 §6]. The paper's own tables list StyleGAN2 under the NVIDIA Source Code License and FFHQ and AFHQ under non-commercial terms, so check each checkpoint and dataset before commercial use [arxiv:2108.00946 tab. 4] [tab. 5]. |
| Few target images (10 to 100) with a few-shot method you already run | Start it from a zero-shot generator, with a short discriminator-only catch-up phase [arxiv:2108.00946 tab. 1] | The FID drops for Ojha et al., MineGAN and TGAN in that setup [arxiv:2108.00946 tab. 2]. The catch-up length is given only as the first 50 steps, in a table caption [arxiv:2108.00946 tab. 1]. |
| Text-to-image or text-guided edits with no training budget, judged on few prompts | Latent optimisation against CLIP [arxiv:2204.08583 §2] | Each image takes minutes of GPU time [arxiv:2204.08583 tab. 2]. The human study is 12 prompts [arxiv:2204.08583 tab. 1]. |
| Long text (more than about 20 words) where one end-of-sequence reward is too sparse, and the discriminator's features can be read at each step | Hierarchical generator with feature-derived goals (LeakGAN) [arxiv:1709.08624 §Methodology] | The paper reports lower NLL and higher BLEU than SeqGAN and RankGAN at lengths 20 and 40 and on WMT News [arxiv:1709.08624 tab. 1] [tab. 2]. The appendix reports training instability when the hierarchy is replaced by a fully connected layer [arxiv:1709.08624 app. Discussions]. |
| A pretrained latent GAN and autoencoder, with a need for one forward pass per input rather than a search over seeds | Agent that picks the GAN seed from the scan's code (RL-GAN-Net) [arxiv:1904.12304 §3.3] [alg. 1] | The paper reports about 1 ms per completion, against a quoted 324 s for an optimisation baseline [arxiv:1904.12304 §4.1]. The code departs from the paper's reward and Chamfer definition [arxiv:1904.12304]. Plan to fix the reward yourself. |
| Discrete sequence output where the discriminator scores finished sequences and you accept a policy-gradient update | SeqGAN: roll-out estimates of token values, then policy gradient [arxiv:1609.05473 §Method] [alg. 1] | The paper reports lower oracle NLL than MLE on synthetic data [arxiv:1609.05473 tab. 1]. Roll-outs make each step costly [arxiv:1609.05473 §Pitfalls], and the reward has no baseline [§Pitfalls]. |
| Several domains with one model, and outputs that should vary across samples or match a reference image | StarGAN v2: style code from noise or a reference, with diversity and cycle terms [arxiv:1912.01865 §2.1] [§2.2] | Latent-guided FID on AFHQ is 16.2, against 61.4 for MSGAN, which is trained per domain pair [arxiv:1912.01865 tab. 2]. The code is non-commercial, so implement from the paper [arxiv:1912.01865 §Paper vs code]. |
| Comparing text generators, adversarial or not, on shared data and metrics | Texygen as a harness, not as a method [arxiv:1802.01886 §3] | It ships seven generators and five metrics on a synthetic oracle and on COCO captions [arxiv:1802.01886 §2.1] [§4.1]. Its numbers come from its own runs, and its code departs from the paper in several settings [arxiv:1802.01886 §Paper vs code]. |
| Editing real photos by changing attributes, with a conditional GAN you train yourself and no pretrained generator to reuse | Conditional GAN with encoders trained after it (IcGAN) [arxiv:1611.06355 §4] [§4.1] | The IND encoder has the lowest reconstruction loss, 0.429 against 0.500 for a shared encoder [arxiv:1611.06355 §5.3]. Attribute edits are judged only qualitatively [§5.4]. |
| Paired restoration where the reference images come from a process you trust, such as colour correction | U-shape Transformer with LAB, LCH, RGB and perceptual losses [arxiv:2111.11843 §3.2] [eq. 13] | The paper reports the best PSNR on both full-reference sets [arxiv:2111.11843 tab. 3]. The ablation has no run without the adversarial term [tab. 6]. The source photos are academic-use only [§3.1]. |
| Rare failure types in tabular data where the minority is a mix of several causes, and hours of GPU or CPU time are affordable | One WGAN-GP per failure type, pooled, with a classifier filter [arxiv:2607.19153 §3.2.5] [§3.3.1] | PR-AUC is 0.8574, against 0.835 for cost-sensitive learning, from single runs [arxiv:2607.19153 tab. 4]. The F1 is lower than the baseline's, and the run took about 14 hours [tab. 4] [§5.1]. No official code was found. |
| Continuous event sequences such as note length, pitch, loudness and timing, with no fixed symbol menu | C-RNN-GAN: continuous events, feature matching and a freezing rule [arxiv:1611.09904 §2] [§3] | The paper's evidence is curves in a figure, with no numeric table [arxiv:1611.09904 fig. 3]. Several training settings are in the code only, so the paper alone will not reproduce its runs [arxiv:1611.09904 §Paper vs code]. |
| Under-sampled MRI with a known sampling pattern, and a need for one forward pass per image | RefineGAN: residual generator with a cyclic loss and a chained second generator [arxiv:1709.00753 §III-D] [§III-G] | The paper reports about 0.1 s per 256×256 image, against 8.57 s for CSCMRI on brain data [arxiv:1709.00753 tab. 1]. The introduction's 100 ms claim does not hold for its own RefineGAN row [arxiv:1709.00753 §I]. Its code has no license. |
| The blur kernel of one image is unknown, no training data is available, and you have a non-blind SR method that takes a kernel | KernelGAN: linear generator trained on the test image [arxiv:1909.06581 §4.1] [§5.3] | The paper reports 30.363 dB at ×2 with ZSSR, against 29.368 dB for the earlier estimator [arxiv:1909.06581 tab. 1]. Each image takes 61 s on a V100 [§5.3]. The code is non-commercial. |
| Source-free adaptation: only the trained classifier and unlabeled target images are on hand | 3C-GAN: class-conditional generator with a weight anchor and a clustering regularizer [arxiv:2502.19316 §3.1] [§3.2] [§3.3] | The paper reports 81.6% on VisDA17 and 89.6% on the Office-31 average [arxiv:2502.19316 tab. 2] [tab. 3]. Most of its baselines use source data [§4.2]. Its semantic weight λ_s is not stated [eq. 4]. |
| Sampling from a large diffusion teacher in very few steps, with some loss of diversity acceptable | Adversarial distillation into a few-step student (ADD) [arxiv:2311.17042 §1] [§4.3] | The 860M student reaches FID 19.7 in one step on COCO [arxiv:2311.17042 tab. 2]. The student is less diverse than its teacher [arxiv:2311.17042 §4.3]. The main text gives no optimiser or batch size [arxiv:2311.17042 §4]. |
| Editing a real photo with a pretrained StyleGAN, where a W+ code loses too much detail | Image2StyleGAN++: optimise W+, then per-layer noise maps, with masks for local edits [arxiv:1911.11544 §4.2] [§7] | The paper reports 39 to 45 dB reconstruction PSNR, against about 20 dB for the previous best [arxiv:1911.11544 §1] [fig. 4]. Inpainting beats partial and gated convolution at 256, 512 and 1024 pixels [arxiv:1911.11544 tab. 1]. No official code was found. |
| Editing a real face with StyleGAN2, where each identity may get its own generator copy and about three minutes of tuning | PTI: invert to the W space, then tune the generator with a locality penalty [arxiv:2106.05744 §3.2] [§4.5] | The paper reports LPIPS 0.09 against 0.34 for the best baseline on CelebA-HQ [arxiv:2106.05744 tab. 1]. The locality term is off by default in the code, and the paper's regularised results need it [arxiv:2106.05744 §Paper vs code]. |
| Choosing an augmenter for a tabular imbalance problem, with a GAN as one candidate | Oversample with SMOTE or ROS and LightGBM, rather than CT-GAN [arxiv:2304.02858 §7] [tab. 8] | The conclusion names SMOTE + LightGBM and ROS + LightGBM as best. Its table lists CT-GAN among the best pairs on five of 23 binary sets [tab. 8]. The cost advantage it claims is not measured [arxiv:2304.02858 §Pitfalls]. |
| One input image, a new camera pose, and no depth or 3D labels for training | Point-cloud renderer with a GAN refinement network (SynSin) [arxiv:1912.08804 §3.2] [§3.3] | The paper reports 22.31 PSNR on RealEstate10K with one input view [arxiv:1912.08804 tab. 1]. No run in the card removes the GAN term [tab. 1]. The code mixes licences, including non-commercial parts [arxiv:1912.08804 §Paper vs code]. |
| Robot navigation among walking people, with a forecast per person from recent positions, and a planner whose cost you can extend | Forecast-augmented TEB: a Social GAN forecast feeds three extra cost terms [arxiv:2411.01814 §2.2] [eq. 5–8] | The paper reports the shortest total time in 3 of 4 simulated scenarios and the largest minimum distance in 3 of 4 [arxiv:2411.01814 tab. 1]. Its text claims wins in every scenario, which the table does not show [§3.2]. The planner code is not in the repository [arxiv:2411.01814 §Paper vs code]. |
| Video clips of arbitrary length or frame rate, from a StyleGAN2 image generator you can train, with clips that can be sampled at any time | StyleGAN-V: motion codes from a grid of noise, and a discriminator on k sampled frames [arxiv:2112.14683 §3.1] [§3.2] | The paper reports the lowest FVD16 on three of four 256 squared benchmarks, with about 8 GPU-days on FaceForensics [arxiv:2112.14683 tab. 1]. Its RainbowJelly benchmark cannot be released [app. E]. Removing its discriminator design raises FaceForensics FVD16 to 154.0 [tab. 2]. |
| Judging a GAN for a design task where realism alone is not the goal | Score several metric groups, and read the same metric in the direction the question needs [arxiv:2302.02913 §Method] | On six-mode 2-D data the GAN wins maximum mean discrepancy and the VAE wins F1 [arxiv:2302.02913 tab. 2]. The single runs are not tested for significance [§Pitfalls]. |
| One photo of a face, with a 3D shape model and the skin texture to recover, and tens of seconds of optimisation per image are affordable | GANFIT: a 3D morphable model fitted through a GAN texture prior and face-recognition losses [arxiv:1902.05978 §3.2] [§3.3] | The paper reports a mean shape error of 0.94 to 0.95 mm, against 1.48 to 1.50 for the best baseline [arxiv:1902.05978 tab. 1]. Fitting takes about 30 s per image [§4.1]. The released repository has no fitting code [§Paper vs code]. |
| A pretrained StyleGAN2 image that must change shape or pose by moving chosen points, with no 3D model and no mask-based tracker | DragGAN: motion supervision on the latent, then nearest-neighbour point tracking [arxiv:2305.10973 §3.2] [§3.3] | The paper reports a landmark distance of 2.44 with one point, against 11.64 for UserControllableLT [arxiv:2305.10973 tab. 1]. Its repository is for research use only [§Paper vs code]. |
| Scoring a GAN for a report, where a single number is not enough | Several measures, each read against its known blind spot [arxiv:2103.09396 §5] [§4.1] | The survey says IS and FID mix quality and diversity into one number, and that both can be raised by memorising the training data [arxiv:2103.09396 §1.1] [§4.1]. It runs no experiments, so check each quoted number against its source [§Pitfalls]. |
| Several views per sample, with some views missing for some samples and no complete pairs to train an imputer | Shared-code model first (CPM-Nets, no GAN); add per-view discriminators (CPM-GAN) only if the imputed values must look realistic [arxiv:2011.06170 §2.1] [§3] | The paper reports its imputation as best at every missing rate on six datasets, against average, SVD, CRA and CPM without the GAN [arxiv:2011.06170 §4.4.1] [fig. 7]. The released code has no GAN part, so write that part from the paper [arxiv:2011.06170 §Paper vs code]. |
| Emotional voice conversion from speech with no frame-aligned pairs, judged by mel-cepstral distortion and listening tests | CycleGAN-EVC as the GAN converter, with VAWGAN-EVC as the comparison [arxiv:2105.14762 §5.1] [fig. 3] | On the ESD benchmark, its MCD is lower than VAWGAN-EVC's on all four neutral-to-emotion pairs [arxiv:2105.14762 tab. 5]. Training took about 72 hours against about 7 hours for VAWGAN-EVC on the same GPU [§5.4]. The benchmark uses one male speaker [§5.1]. |
| Training budget, learning-rate schedule, or iteration count | Use the card's values, not a default | Motion uses 55,000 to 350,000 iterations [arxiv:2102.04942 §4.2, §4.3]. VQGAN+CLIP uses 400 [arxiv:2204.08583 §2]. KernelGAN runs 3,000 iterations [arxiv:1909.06581 §2]. PTI tunes for up to 350 steps [arxiv:2106.05744 §4.5]. RefineGAN trains for 500 epochs [arxiv:1709.00753 §III-F]. The DTN and fashion cards do not state iteration counts. The text-to-image card states no optimiser or learning rate [arxiv:1904.01480 §4.1]. The MR-to-CT card reports a learning rate of 1e-6 and an iteration count only in the code [arxiv:1612.05362 §4] [code: 3dganversion/main.py#L9]. |

## Implementation sketch

Common core of the adversarial variants, in the form the cards share.
The variants differ in the extra terms, in how D scores its input, and in the
loss form.

```pseudocode
# Adversarial training step, common core  [arxiv:1612.05362 §2.1]
# G is conditioned on an input x, not on noise
# Variant terms are passed in as `extra`

def gan_step(x, y, w_adv, extra):
    # Discriminator: real target -> 1, generated -> 0  [eq. 3, eq. 4]
    fake = G(x)
    update(D, bce(D(y), 1) + bce(D(detach(fake)), 0))

    # Generator: fool D, plus the variant's extra terms  [eq. 5, eq. 7]
    loss = w_adv * bce(D(fake), 1) + extra(x, fake, y)
    update(G, loss)
```

Variant notes. For DTN, D has three classes, G is `g(f(x))`, f stays frozen,
and no paired `y` is used. The extra terms are feature constancy on source
samples and identity on target samples [arxiv:1611.02200 eq. 3–6]. For the
MR-to-CT variant, `extra` is an L2 term on the paired CT plus a 3D
gradient-difference term [arxiv:1612.05362 eq. 5, eq. 6]. Its discriminator
is a single-output CNN with binary cross-entropy [arxiv:1612.05362 eq. 3, eq.
4]. For the motion variant, the critics use a least-squares form: real windows
are pushed to 1 and generated ones to 0, each with a factor of one half
[arxiv:2102.04942 §3.6.2]. The fashion variant runs this step once per stage,
with the conditions given to D as well [arxiv:1710.07346 §3.4]. Learning rate,
batch size and iteration counts are not stated in either DTN or fashion text,
so treat them as tuning parameters [arxiv:1611.02200 §4]
[arxiv:1710.07346 §3.4].

For the aerial translator, the step runs in both directions. The reconstruction
term is the cycle loss on each round trip, which the paper leaves unweighted and
the code weights at 10. The code also adds an identity term that the paper does
not describe [arxiv:1905.03198 eq. 8] [§4.3]. For SD-GAN, `extra` is the
contrastive term on the discriminator features of the two branches, added to
the AttnGAN losses with no weight given [arxiv:1904.01480 §4.1]. For RL-GAN-Net,
the GAN is trained first with a WGAN-GP critic on latent codes. The agent is
trained afterwards against a frozen generator, so its reward does not update G
[arxiv:1904.12304 §3.2] [§3.3].

For 3C-GAN, G takes a label as well as noise, and D scores real target images
against generated ones without pairs. C is trained on G's samples with their
conditioning labels, so it also acts as the semantic judge for G
[arxiv:2502.19316 §3.1] [alg. 1]. For RefineGAN, `extra` is the cyclic loss,
and D reads the output of the chained generator [arxiv:1709.00753 eq. 6]
[§III-G]. For KernelGAN, both networks see crops of one image. D scores patches
with an L1 loss in least-squares form, and the kernel regularizer is added to
G's loss [arxiv:1909.06581 §2] [§4.2]. For C-RNN-GAN, D reads a whole sequence
and gives one decision, and feature matching replaces G's adversarial loss in
the adversarial phase [arxiv:1611.09904 §3]. For StyleGAN-V, D scores a whole
clip, so its logit and its R1 penalty are per clip. The loss is the
non-saturating logistic form [arxiv:2112.14683 app. B]. Its generator is a
StyleGAN2 network with the motion code as its constant input [arxiv:2112.14683
§3.1].

LeakGAN does not fit this core. Its generator is a word-level policy, and its
Manager and Worker take policy-gradient updates from feature changes and an
intrinsic reward, not one adversarial term [arxiv:1709.08624 eq. 7] [eq. 8].
StyleGAN-NADA does not fit it either. It has no discriminator, and its update
is one minus the cosine between a CLIP image direction and a CLIP text
direction [arxiv:2108.00946 eq. 2]. GANFIT and DragGAN do not fit it, because
neither trains a discriminator. GANFIT optimises a latent through a face
recognition loss, and DragGAN through a feature-matching loss on handle points
[arxiv:1902.05978 §3.2] [arxiv:2305.10973 eq. 1].

## Contradictions

None known between these cards on the same claim. The StyleGAN-NADA card
reports that a global CLIP loss on the trained generator fails and collapses
[arxiv:2108.00946 §4.1] [fig. 15]. VQGAN+CLIP optimises a latent with the same
global CLIP loss and reports strong human ratings [arxiv:2204.08583 §2]
[arxiv:2204.08583 tab. 1]. The two optimise different things, generator weights in one and a latent of
a frozen decoder in the other, and the cards do not test the global loss on
the other's setting, so the results do not conflict. Four further points keep
the cards from being compared directly. The DTN card says a direct binary GAN with feature
constancy on raw pixels "does not produce desirable results" and scores 13.71%
on its digit task [arxiv:1611.02200 §3, tab. 1]. The MR-to-CT card uses a
paired generator with L2 and reports PSNR gains from the adversarial term, on a
different dataset and metric [arxiv:1612.05362 §4.1]. The motion card reports
that its adversarial term adds little to L2Q [arxiv:2102.04942 tab. 2]. The
aerial card's gain is a segmentation accuracy, so it does not bear on the PSNR
gain in the MR-to-CT card [arxiv:1905.03198 tab. 3] [arxiv:1612.05362 §4.1].
Two cards compare an adversarial text generator with maximum likelihood, on
different corpora. SeqGAN is above MLE on BLEU-4 for Obama speeches, 0.427
against 0.416 [arxiv:1609.05473 tab. 3]. Texygen's own runs on COCO captions
put SeqGAN below MLE, 0.294 against 0.305 [arxiv:1802.01886 tab. 2]. The corpora
and implementations differ, so the cards do not show which sign holds in
general. All of these numbers are author-reported. The LeakGAN text results come from the authors'
own runs against SeqGAN and RankGAN, and the wiki holds no independent
reproduction of them [arxiv:1709.08624 §Experiment]. The underwater card's
ablation has no run without the adversarial term, so it adds no evidence on
whether that term helps [arxiv:2111.11843 tab. 6]. The tabular card disagrees
with itself. Its text calls the multi-GAN F1 the highest of all conditions and
also says it is slightly below the baseline. Table 4 gives 0.6923 for the
multi-GAN and 0.740 for the baseline, so the table is the value to use
[arxiv:2607.19153 §4.1] [tab. 4]. Its baseline PR-AUC is 0.835 in the table and
0.836 in a figure [tab. 4] [fig. 11].

The two inversion cards report incompatible numbers. PTI gives LPIPS, MSE,
MS-SSIM and identity similarity on CelebA-HQ [arxiv:2106.05744 tab. 1].
Image2StyleGAN++ gives PSNR and, for inpainting, SSIM, MSE and PSNR on its own
images [arxiv:1911.11544 tab. 1]. Their reconstruction gains cannot be ranked
against each other. PTI also reports that a mean or random code used as a fixed
pivot gives weaker edits [arxiv:2106.05744 §4.4]. Image2StyleGAN++ starts its
optimisation from a mean face code [arxiv:1911.11544 §7.1]. The two uses differ,
so the cards do not conflict. The augmentation benchmark disagrees with itself.
Its conclusion says CT-GAN is not a top performer, but its table lists CT-GAN
among the best pairs on five of 23 binary sets [arxiv:2304.02858 §7] [tab. 8].

The video card reports FVD values on its own protocol, with every baseline
retrained by its authors. Its FVD numbers therefore cannot be set against FVD
values from other papers [arxiv:2112.14683 §4] [app. C]. The FVD card says
that identical sets have a nonzero distance and that comparisons need equal
sample counts [arxiv:1812.01717 §4.2].

### Fitting a 3D face through a GAN texture prior (GANFIT)

GANFIT recovers shape and skin of a 3D face from one photo by gradient descent.
The shape is a linear 3D morphable model, with identity and expression bases.
The skin is a generator trained on about 10,000 UV textures. The texture
equation optimises a 512-value input latent; whether the generator's own
weights are also updated is not stated [arxiv:1902.05978 §3.1] [§4.1] [eq. 2]
[eq. 5]. A differentiable renderer produces the image, and the loss mixes a cosine
distance on face-recognition embeddings, intermediate face-network features, a
pixel L1 term and a landmark term [arxiv:1902.05978 §3.2] [§3.3] [eq. 8–11].
Fitting runs in two stages: a landmark-only alignment, then a joint stage over
all parameters [arxiv:1902.05978 §3.4]. Card: [GANFIT: Generative Adversarial
Network Fitting for High Fidelity 3D Face Reconstruction](/papers/1902.05978).

On the MICC Florence benchmark, the author-reported mean shape error in each of
three settings is 0.94 to 0.95 mm, against 1.48 to 1.50 for the best baseline
in each setting [arxiv:1902.05978 tab. 1]. The text gives a 36% drop against
the other two methods, and the Genova et al. row matches that figure [§4.3]. The identity results on LFW are given only as cosine
similarities [app. A]. The ablation of the identity, content and pixel terms
is qualitative [§4.4]. The text says the identity latent is averaged across
frames, but the equation sums them [§3.4] [eq. 13]. The released repository has
the evaluation scripts only, not the fitting code [arxiv:1902.05978 §Paper vs
code], so the hyperparameters come from the paper alone.

### Point-based editing of a StyleGAN2 image by motion supervision (DragGAN)

DragGAN moves chosen handle points of a pretrained StyleGAN2 image to target
points. Each iteration takes one optimiser step on the latent code. The loss
pushes the features near each handle toward its target, using the features
after the sixth generator block [arxiv:2305.10973 §3.2] [eq. 1]. Each handle is
then re-located by nearest-neighbour search in the same features
[arxiv:2305.10973 §3.3] [eq. 2]. No optical-flow or tracking network is used.
The loop stops when every handle is within a tolerance of its target
[§3.4]. Card: [Drag Your GAN: Interactive Point-based Manipulation on the
Generative Image Manifold](/papers/2305.10973).

On a face-landmark benchmark with one point, the author-reported mean landmark
distance is 2.44, against 11.64 for UserControllableLT and 13.43 for the same
method with RAFT tracking [arxiv:2305.10973 tab. 1]. The edit takes about 2.0 s
per run [tab. 1]. The paper names an RTX 3090 in its introduction but does not
say that Table 1 was timed on it [§1]. For real images the paper first inverts
with PTI, and results depend on that inversion [arxiv:2305.10973 §4.1] [fig. 5].
The paper reports 30 to 200 iterations in typical use, but the step length is
not fixed, so that count is an observation and not a stopping rule [§3.1]. The
distance unit is not stated [tab. 1]. The repository's licence is the NVIDIA Source Code License
for StyleGAN3, which allows research use only, so implement from the paper
[arxiv:2305.10973 §Paper vs code].

## Papers

- [Unsupervised Cross-Domain Image Generation](/papers/1611.02200): DTN, unpaired translation through a frozen feature function, with a three-class discriminator and identity and constancy terms.
- [Medical Image Synthesis with Context-Aware Generative Adversarial Networks](/papers/1612.05362): paired 3D conditional GAN for MR-to-CT synthesis, with an L2 and gradient-difference loss and Auto-Context stacking.
- [Be Your Own Prada: Fashion Synthesis with Structural Coherence](/papers/1710.07346): two conditional GANs, one for a segmentation map and one for per-label textures, for text-guided outfit changes.
- [Robust Motion In-betweening](/papers/2102.04942): LSTM generator with two least-squares critics and additive time-to-arrival and target-noise embeddings for transition generation.
- [Optimal transport mapping via input convex neural networks](/papers/1908.10962): transport map as the gradient of a convex network, learned by a minimax game against a second convex network.
- [StyleFlow](/papers/2008.02401): conditional flow on the StyleGAN latent for attribute sampling and editing, with no adversarial training.
- [StyleGAN-NADA: CLIP-Guided Domain Adaptation of Image Generators](/papers/2108.00946): text-only domain shift of a StyleGAN2 generator with a directional CLIP loss and adaptive layer freezing.
- [VQGAN-CLIP](/papers/2204.08583): training-free text-guided generation and editing by optimising a frozen VQGAN latent against CLIP.
- [SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient](/papers/1609.05473): a text generator as a policy, with a discriminator that scores finished sequences and roll-out estimates of token values.
- [StarGAN v2: Diverse Image Synthesis for Multiple Domains](/papers/1912.01865): one generator for many domains, with style codes from noise or a reference image, injected by AdaIN, and diversity and cycle terms.
- [Texygen: A Benchmarking Platform for Text Generation Models](/papers/1802.01886): a benchmark harness with seven text generators and five metrics. It adds no method of its own.
- [Long Text Generation via Adversarial Training with Leaked Information](/papers/1709.08624): LeakGAN, a text GAN whose Manager reads discriminator features and steers a Worker's word choice.
- [Semantics Disentangling for Text-to-Image Generation](/papers/1904.01480): SD-GAN, a stacked text-to-image GAN with a Siamese contrastive term and semantic-conditioned batch norm.
- [RL-GAN-Net](/papers/1904.12304): a latent GAN on autoencoder codes, with a DDPG agent that picks the seed for point-cloud completion.
- [Unsupervised Domain Adaptation using Generative Adversarial Networks for Semantic Segmentation of Aerial Images](/papers/1905.03198): cycle-consistent restyling of labelled source images, then fine-tuning a segmenter without target labels.
- [WaveFake](/papers/2111.02813): a dataset of generated speech, some from GAN-based vocoders, for audio deepfake detection. The GAN is not the method.
- [A Survey on Bridging EEG Signals and Generative AI](/papers/2502.12048): a literature survey in which GANs are one image decoder among several. It reports no new results.
- [The History Began from AlexNet: A Comprehensive Survey on Deep Learning Approaches](/papers/1803.01164): a deep-learning survey with one section on GANs. It adds no method and no code.
- [Invertible Conditional GANs for image editing](/papers/1611.06355): IcGAN, a conditional GAN plus two encoders that map a real image to a noise vector and an attribute vector, for attribute edits.
- [U-shape Transformer for Underwater Image Enhancement](/papers/2111.11843): a GAN-trained U-shaped generator with channel-wise and spatial transformer blocks and LAB and LCH colour losses, with the LSUI dataset.
- [Breaking the Homogeneity Assumption: Specialized Multi-Generator Adversarial Learning for Rare Failure Detection in Predictive Maintenance](/papers/2607.19153): one WGAN-GP tabular generator per failure type, pooled into the training set to oversample rare failures.
- [C-RNN-GAN: Continuous recurrent neural networks with adversarial training](/papers/1611.09904): GAN over continuous music events, with an LSTM generator, a bidirectional LSTM discriminator and feature matching.
- [Compressed Sensing MRI Reconstruction using a Generative Adversarial Network with a Cyclic Loss](/papers/1709.00753): residual GAN with a cyclic data-consistency loss and a chained second generator, for MRI reconstruction.
- [Blind Super-Resolution Kernel Estimation using an Internal-GAN](/papers/1909.06581): KernelGAN, a linear generator trained on one image to estimate its blur kernel for super-resolution.
- [Image2StyleGAN++: How to Edit the Embedded Images?](/papers/1911.11544): embedding a photo into StyleGAN's W+ space and then into noise maps, with masks for local edits and inpainting.
- [Pivotal Tuning for Latent-based Editing of Real Images](/papers/2106.05744): PTI, an inversion to the W space followed by a short generator fine-tune with a locality penalty, for editing real faces.
- [Ensemble and augmentation benchmark for imbalanced data](/papers/2304.02858): a benchmark of nine augmenters, CT-GAN among them, combined with ensemble models on imbalanced data.
- [Model Adaptation: Unsupervised Domain Adaptation without Source Data](/papers/2502.19316): 3C-GAN, a class-conditional generator and classifier for source-free domain adaptation.
- [Age Progression/Regression by Conditional Adversarial Autoencoder](/papers/1702.08423): a conditional autoencoder with an age-conditioned image discriminator and a prior-matching discriminator on the code, for age progression and regression.
- [Adversarial Diffusion Distillation](/papers/2311.17042): a few-step student distilled from a diffusion teacher, with a hinge discriminator on frozen features.
- [SynSin: End-to-end View Synthesis from a Single Image](/papers/1912.08804): single-image view synthesis with a point-cloud renderer, a depth network, and a refinement GAN with two discriminators.
- [Social GAN forecasts for robot navigation among people](/papers/2411.01814): a Social GAN forecaster feeds extra social cost terms into the Timed Elastic Band local planner.
- [Fréchet Video Distance](/papers/1812.01717): an evaluation metric for video generators, computed on Kinetics-400 features. It trains no GAN.
- [Deep Generative Modelling: A Comparative Review of VAEs, GANs, Normalizing Flows, Energy-Based and Autoregressive Models](/papers/2103.04922): a survey that compares five generative families, GANs among them, on cited CIFAR-10 numbers. It adds no method.
- [StyleGAN-V: A Continuous Video Generator with the Price, Image Quality and Perks of StyleGAN2](/papers/2112.14683): a StyleGAN2-based video GAN with motion codes for any real-valued time and a discriminator on k sampled frames.
- [Beyond Statistical Similarity: Rethinking Metrics for Deep Generative Models in Engineering Design](/papers/2302.02913): a review of metric groups for generative models in engineering design, with a GAN and a VAE on 2-D data as the running example.
- [GANFIT: Generative Adversarial Network Fitting for High Fidelity 3D Face Reconstruction](/papers/1902.05978): a 3D morphable model fitted to one photo, with a GAN texture prior and face-recognition losses.
- [Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold](/papers/2305.10973): point-based editing of a pretrained StyleGAN2 image, with motion supervision on the latent and nearest-neighbour point tracking.
- [Pros and Cons of GAN Evaluation Measures: New Developments](/papers/2103.09396): a survey of GAN evaluation measures published after 2019, with no new experiments.
- [A Review of Graph Neural Networks and Their Applications in Power Systems](/papers/2101.10025): a review in which graph GANs are one of several graph generative models, with one quoted synthetic-feeder result.
- [Deep Partial Multi-View Learning](/papers/2011.06170): CPM-Nets, shared per-sample codes rebuilt into observed views, with an optional per-view GAN (CPM-GAN) for missing entries.
- [Emotional Voice Conversion: Theory, Databases and ESD](/papers/2105.14762): the ESD emotional speech database, benchmarked with CycleGAN-EVC and VAWGAN-EVC. The GAN is a baseline, not the contribution.