CLIP-MDGAN: Multi-Discriminator GAN Using CLIP Task Allocation
Shonosuke Gonda, Fumihiko Sakaue and Jun Sato
Nagoya Institute of Technology, Nagoya 466-8555, Japan
Keywords:
Image Synthesis, Image Distribution, GAN, Multi-Discriminator, Clip, Foundation Model, Multimodal.
Abstract:
In a Generative Adversarial Network (GAN), in which the generator and discriminator learn adversarially, the
performance of the generator can be improved by improving the discriminator’s discriminatory ability. Thus,
in this paper, we propose a method to improve the generator’s generative ability by adversarially training a
single generator with multiple discriminators, each with different expertise. By each discriminator having
different expertise, the overall discriminatory ability of the discriminator is improved, which improves the
generator’s performance. However, it is not easy to give multiple discriminators independent expertise. To
address this, we propose CLIP-MDGAN, which leverages CLIP, a large-scale learning model that has recently
attracted a lot of attention, to classify a dataset into multiple classes with different visual features. Based on
CLIP-based classification, each discriminator is assigned a specific subset of images to promote the devel-
opment of independent expertise. Furthermore, we introduce a method to gradually increase the number of
discriminators in adversarial training to reduce instability in training multiple discriminators and reduce train-
ing costs.
1 INTRODUCTION
In recent years, advances in research in VAE (Vari-
ational Auto Encoder) (Kingma and Welling, 2014),
GAN (Generative Adversarial Network) (Goodfel-
low et al., 2014), and diffusion models (Ramesh
et al., 2021; Rombach et al., 2022) have made it
possible to reproduce the image distribution within
datasets accurately and to generate high-quality im-
ages. Among these approaches, GANs have demon-
strated applications not only in high-fidelity image
generation (Brock et al., 2019) (Karras et al., 2020)
but also in diverse tasks such as domain transla-
tion (Karras et al., 2020), super-resolution (Ledig
et al., 2017), text-to-image generation (Zhang et al.,
2017), and anomaly detection (Zhang et al., 2017).
The mechanism of GAN, which improves capabili-
ties by competing among multiple networks, is very
important in advancing learning methodologies.
In GANs, training typically progresses by pairing
one Generator with one Discriminator, allowing them
to learn in competition. As the Discriminator’s abil-
ity to distinguish improves, so does the Generator’s
ability to produce realistic images. If we think of the
Generator as a student and the Discriminator as a su-
pervisor, students will be able to acquire a broader
and higher level of ability if they are taught by mul-
Figure 1: A method for automatically selecting the Discrim-
inator corresponding to each training image based on image
feature distribution.
tiple supervisors, each with different specialized ex-
pertise, rather than by a single supervisor. Thus, we
in this paper propose a method to enhance the Genera-
tor’s capabilities by adversarially training it with mul-
tiple Discriminators, each with different expertise, as
shown in Figure 1.
Several methods have previously been proposed to
enhance Generator performance using multiple Dis-
criminators (Dinh Nguyen et al., 2017; Durugkar
et al., 2017; Choi and Han, 2022). However, task
allocation to multiple Discriminators is not a trivial
problem and it often suffers from instability. Thus, in
this paper, we propose a method for allocating tasks to
multiple Discriminators by using CLIP (Contrastive
464
Gonda, S., Sakaue, F. and Sato, J.
CLIP-MDGAN: Multi-Discriminator GAN Using CLIP Task Allocation.
DOI: 10.5220/0013231900003912
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 2: VISAPP, pages
464-470
ISBN: 978-989-758-728-3; ISSN: 2184-4321
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

Language-Image Pre-training) (Radford et al., 2021),
a large-scale model pretrained on the relationship be-
tween images and texts using large amount of image
and text data.
In recent years, the advent of large-scale mod-
els pretrained on vast datasets has enabled zero-shot
performance across various downstream tasks by uti-
lizing well-trained Encoders within these large-scale
models. In light of this, we propose an efficient and
stable method for adversarial training between a Gen-
erator and multiple highly specialized Discriminators
by leveraging the Encoder of CLIP to streamline the
specialization of multiple Discriminators. We call it
CLIP-MDGAN.
As illustrated in Figure 1, our CLIP-MDGAN se-
lects the appropriate Discriminator for each image
based on its image features. By strategically employ-
ing CLIP’s Text Encoder and Image Encoder, we can
cluster image groups effectively for each Discrimina-
tor. Since CLIP Encoders are trained to capture cor-
relations between images and text, the feature vec-
tors from its Image Encoder serve as valuable sym-
bolic descriptors, facilitating meaningful clustering of
large datasets. Hence, by clustering images into mul-
tiple classes using these feature vectors, we assign
each Discriminator a specific classification task. This
assignment allows each Discriminator to focus on
identifying images with distinct characteristics, fos-
tering high specialization among multiple Discrim-
inators and enhancing the Generator’s performance
through adversarial training. Our approach avoids the
instability seen in the existing methods during Dis-
criminator selection, while enabling each Discrimina-
tor to develop high specialization efficiently.
This paper presents both a semi-automatic and a
fully automatic approach for assigning tasks to mul-
tiple Discriminators using CLIP. Additionally, to en-
sure fast and stable task allocation to multiple Dis-
criminators, we introduce a method that gradually in-
creases the number of Discriminators in stages.
While this study focuses on enhancing a sin-
gle Generator with multiple Discriminators, this ap-
proach can be extended to frameworks involving mul-
tiple Generators (Ghosh et al., 2018; Hoang et al.,
2018), potentially enabling multi-Generator systems
to benefit from adversarial training with specialized
Discriminators.
2 RELATED WORK
Several methods have previously been proposed
to enhance Generator performance through the
use of multiple Discriminators. For example,
Figure 2: Overview of CLIP-MDGAN.
D2GAN (Dinh Nguyen et al., 2017) demonstrated
that employing two Discriminators with opposing
loss functions could help avoid mode collapse in
GANs. Albuquerque (Albuquerque et al., 2019) intro-
duced multi-objective discriminators, each focused on
distinct tasks, thereby enhancing the Generator’s per-
formance. GMAN (Durugkar et al., 2017) proposed
a balanced approach by combining a hard Discrimi-
nator (strict teacher) and a soft Discriminator (lenient
teacher) to regulate the Generator’s learning.
There have also been attempts to instill specialized
skills within multiple Discriminators during train-
ing. MCL-GAN (Choi and Han, 2022) utilizes Multi-
ple Choice Learning (MCL) (Guzman-Rivera et al.,
2012) to select a limited number of Discriminators
from a larger pool based on each dataset image, facil-
itating adversarial learning with the Generator. How-
ever, this method requires multiple Discriminators for
the classification of a single data instance, resulting in
overlapping roles among these Discriminators. Addi-
tionally, during the early stages of training, task allo-
cation to each Discriminator can be unstable, making
it challenging to maintain operational stability.
Thus, we in this paper propose a method for al-
locating tasks to multiple Discriminators by using
CLIP (Radford et al., 2021), a large-scale model pre-
trained on the relationship between images and texts.
By using the well-trained text and image encoders of
CLIP, our method can avoid instability during Dis-
criminator selection, while enabling each Discrimina-
tor to develop specialization efficiently.
3 TRAINING GENERATORS
WITH MULTIPLE
DISCRIMINATORS
The overall view of our CLIP-MDGAN is as shown
in Figure 2.
CLIP-MDGAN: Multi-Discriminator GAN Using CLIP Task Allocation
465

Figure 3: Hand-made class assignment for multiple dis-
criminators.
First, we divide Dataset X into N classes X
i
(i =
1, ··· , N) by using CLIP Encoders. Next, the CLIP-
based Discriminator Selector identifies which of these
N classes the image G(z) generated by the Generator
from noise z belongs to. G(z) is then passed to the
Discriminator D
i
selected by the Discriminator Se-
lector, and D
i
distinguishes between the dataset X
i
of
Class i and G(z).
In this way, the Discriminator is responsible for
identifying specific classes, which improves the over-
all discriminatory ability of the Discriminators, and
the Generator’s image generation ability is further im-
proved by adversarial training with these Discrimina-
tors.
4 CLASS ASSIGNMENT FOR
MULTIPLE DISCRIMINATORS
When using multiple specialized Discriminators to
determine the authenticity of images generated by a
Generator, it is essential to assign each Discrimina-
tor a specific image class (or domain) to identify. In
this research, we propose two distinct methods for as-
signing image classes to each Discriminator. The fol-
lowing sections will provide a detailed explanation of
these two methods.
4.1 Class Assignment Based on
Hand-Defined Classes
First, we describe a method for assigning classes
to multiple Discriminators based on hand-defined
classes. In image generation tasks, it is sometimes
possible to roughly categorize the images we aim to
generate based on human senses. In such cases, the
method described in this section is useful.
Consider, for example, the task of generating face
images. In this scenario, we can divide the face im-
ages into some classes based on human senses. For
instance, we might set up four hand-defined classes:
male children, female children, adult men, and adult
women. Then, for assigning each class image to an
individual Discriminator, we need to divide the image
dataset into these multiple classes. In this research,
we use CLIP (Radford et al., 2021), a model trained
on a large dataset of text and images, to obtain the
relationship between each class label and its corre-
sponding set of images in a zero-shot manner.
Specifically, as shown in the upper part of Fig-
ure 3, we use CLIP Text Encoder to compute the text
features T
i
for the hand-defined N classes, Class i
(1, ··· , N). Next, we select an image x
j
( j =
1, · ·· , M) from the image dataset X that comprises the
images we aim to generate and calculate its image fea-
ture I
j
using CLIP Image Encoder. Then, by selecting
the text feature T
i
that best matches I
j
based on cosine
similarity, we assign image x
j
to Class i. By repeating
this process for all images in dataset X, we partition
X into N class-specific subsets X
i
(i = 1, ·· · , N), as
shown in the lower part of Figure 3.
In our research, the N Discriminators are each re-
sponsible for determining the authenticity of images
belonging to one of these N classes. This approach
leverages hand-defined categories combined with au-
tomatic class assignment using CLIP, enabling each
Discriminator to specialize in judging a specific sub-
set of images, thus enhancing the Generator’s perfor-
mance across diverse classes.
4.2 Class Assignment Based on Data
Distribution
We next describe a method for assigning classes to
each Discriminator based on the distribution of image
features within the training data. This method auto-
matically determines the optimal class distinction and
allows each discriminator to specialize on a subset of
images with similar characteristics. By using data-
driven clustering, each Discriminator can learn to dis-
tinguish its assigned classes more effectively, which
improves the overall performance of the Generator.
In this method, the image dataset X is automati-
cally divided into N datasets X
i
(i = 1, ··· , N) based
on its distribution. First, the M images x
i
(i =
1, · ·· , M) in the training dataset are input into CLIP
Image Encoder, resulting in M image feature vectors
I
i
(i = 1, · ·· , M). Then, as an initial state for cluster-
ing, the M image feature vectors I
i
are treated as M
clusters C
i
(i = 1, ··· , M) in the image feature space,
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
466

Figure 4: CLIP-based Discriminator Selector.
thereby forming the cluster set C as follows:
C = {C
1
, C
2
, . . . , C
M
} (1)
The distance d(C
i
, C
j
) between clusters C
i
and C
j
is defined based on Ward’s method (Ward, 1963) as
follows:
d(C
i
, C
j
) =
|C
i
||C
j
|
|C
i
| + |C
j
|
∥µ
i
− µ
j
∥
2
(2)
where, |C
i
| represents the number of data points in
cluster C
i
, µ
i
is the centroid (mean vector) of clus-
ter C
i
, and ∥µ
i
− µ
j
∥
2
denotes the squared Euclidean
distance between the centroids of clusters C
i
and C
j
respectively.
Then, the two clusters C
i
and C
j
with the smallest
distance d are merged to form a new cluster C
i j
.
C
i j
= C
i
∪C
j
(3)
As a result, the new cluster set C is generated as fol-
lows:
C ← (C \ {C
i
, C
j
}) ∪ {C
i j
} (4)
By repeating this operation until the number of
clusters reaches N, the image dataset X can be di-
vided into N clusters X
i
(i = 1, · ·· , N) based on the
distribution of image features obtained from CLIP.
5 DISCRIMINATOR SELECTION
Next, we describe a method for selecting which of
the N Discriminators should handle a generated im-
age G(z) from the Generator. In this research, we in-
troduce a new component, the Discriminator Selector,
responsible for making this selection. This Selector
identifies the appropriate Discriminator (and thus the
appropriate class) for each generated image based on
its image features. We implement the Discriminator
Selector using CLIP.
As shown in Figure 4, for each of the N classes
defined in Section 4, we convert the images x
i j
∈ X
i
(i = 1, ·· · , N) into image feature vectors I
i j
using
CLIP Image Encoder. We then compute a mean fea-
ture vector I
mean
i
for each class as follows:
I
mean
i
=
1
M
i
M
i
∑
j=1
I
i j
(5)
where, M
i
denotes the number of data in Class i.
Then, as shown in Figure 4, the class c to which
the generated image I
gen
belongs is determined by
selecting the class with the minimum Euclidean dis-
tance between the feature vector of the generated im-
age I
gen
and the mean feature vector of each class.
This is formulated as follows:
c = argmin
i∈{1,...,N}
∥I
gen
− I
mean
i
∥ (6)
Each time the Generator produces an image, the
Discriminator Selector uses this criterion to assign the
generated image to a specific class. Consequently,
the corresponding Discriminator, specialized in that
class, is selected to evaluate the generated image.
In this method, an appropriate Discriminator is au-
tomatically selected based on the characteristics of
each generated image, resulting in improved discrim-
inative capability compared to using a single Discrim-
inator. Consequently, this enhanced discrimination
leads to improved generative capabilities of the Gen-
erator, as it receives more targeted and specialized
feedback from the selected Discriminator.
6 GRADUAL INCREASE OF
DISCRIMINATORS
In this section, we describe an approach for more ef-
ficient and stable training when using multiple Dis-
criminators. When assigning distinct specializations
to each Discriminator for adversarial learning, the ini-
tial stages of training can be challenging. At this
early stage, the Generator’s image quality is often
low, making it difficult to assign generated images to
the most suitable Discriminator (i.e., to determine the
correct class). This poses a stability issue in the early
training phases. Additionally, optimizing multiple
Discriminators sequentially can be time-consuming.
To address these challenges, our approach gradu-
ally increases the number of Discriminators over the
course of training. In the initial stage, we start with
one Discriminator adversarially trained against a sin-
gle Generator. Once the training has progressed to a
CLIP-MDGAN: Multi-Discriminator GAN Using CLIP Task Allocation
467

certain extent, we incrementally increase the number
of Discriminators. Each time we increase the number
of Discriminators, we do so by splitting each existing
Discriminator into two. Thus, the number of Discrim-
inators increases as 1, 2, 4, 8, and so forth. For each
split, the final parameters of the original Discrimina-
tor are used as the initial parameters for the two new
Discriminators. This inheritance of parameters en-
sures that the characteristics of the original Discrimi-
nator are retained in the split Discriminators, allowing
for stable handling of the expansion process.
There are several possible methods for determin-
ing the timing of Discriminator increases. In this re-
search, for simplicity, we chose to increase the num-
ber of Discriminators after a fixed number of training
iterations. Developing a method to automatically de-
termine this timing is an area for future investigation.
With this approach, we can avoid instability in
Discriminator selection during the early stages of
training. Additionally, as adversarial learning pro-
gresses more efficiently, this method enables perfor-
mance improvements in the Generator with fewer
training iterations.
7 TRAINING
Next, we discuss the training method for the net-
work. Here, we consider the case of adversarially
training a single Generator and N Discriminators D
i
(i = 1, ··· , N) using a training dataset divided into N
classes.
Suppose that an image G(z) generated by the Gen-
erator from noise z is assigned to a Discriminator D
i
by the Discriminator Selector. In this scenario, the
Generator G and the selected Discriminator D
i
engage
in adversarial training according to the following ob-
jective functions:
G
∗
= argmin
G
D
i
(7)
D
∗
i
= argmax
D
i
E
x∼p
data
(x)
[logD
i
(x)]
+ E
z∼p
z
(z)
[log(1 − D
i
(G(z)))] (8)
In this way, an appropriate Discriminator is se-
lected by the Discriminator Selector based on the
characteristics of the image generated by the Gen-
erator, enabling targeted adversarial learning to take
place.
8 EXPERIMENTS
Next, we present the experimental results using the
proposed method. In the following experiments, we
Figure 5: CelebA dataset(64×64 resized).
(a) Training images classified by hand-defined classes
(b) Training images classified by image distribution
Figure 6: Comparison of training images classified into 4
classes C
i
(i = 1, · · · , 4) by hand-defined classes and image
distribution.
target the task of generating face images using the
CelebA (Liu et al., 2015) dataset shown in Figure 5
as training data. In each experiment, the dataset was
divided using the proposed method, and a discrimi-
nator was selected using a Discriminator Selector to
train the GAN. For quantitative evaluation, we used
Fr
´
echet Inception Distance (FID) (Heusel et al., 2017)
to measure the similarity in the feature space between
the generated images and the real images.
8.1 Image Generation Accuracy of
Generator
In this experiment, we compared the image genera-
tion accuracy of the generator by setting the num-
ber of discriminators to 1, 2, and 4. Of these three
cases, the case with 1 discriminator is the conven-
tional method.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
468

(a) 1G1D (b) 1G2D (c) 1G4D
Figure 7: Images generated from 1G1D, 1G2D, and 1G4D.
(a) hand-defined (b) data distribution
Figure 8: Images generated from 1G4D trained by using the
hand-defined classes and the data distribution based classes.
For dataset division based on hand-defined
classes, we set the following four classes and divided
the dataset using CLIP Text Encoder and Image En-
coder.
• ”Man with a smile”
• ”Man with a Neutral-face”
• ”Woman with a smile”
• ”Woman with a Neutral-face”
Example images in the four classes after division
are shown in Figure 6 (a). Each row is a group of im-
ages corresponding to ”Man with a smile”, ”Man with
a Neutral-face”, ”Woman with a smile”, and ”Woman
with a Neutral-face” respectively. From these images,
we find that by using CLIP Encoder, the images can
be roughly classified correctly into the four classes.
Next, we show the results of automatic data seg-
mentation based on the distribution of the image fea-
tures in Figure 6 (b). Each of the four rows repre-
sents one of four classes of images after segmenta-
tion. From the top row, these images appear to be
roughly divided into Western women, Western men,
Asians, and Athletes. These results are different from
the hand-defined classes in Figure 6 (a), but we find
that the images have been segmented into groups with
similar image features.
Based on these data classifications, we compared
the image generation ability of the generator when ad-
versarial learning was performed between one gener-
ator and four discriminators (1G4D) with adversarial
learning with one discriminator (1G1D) and adversar-
ial learning with two discriminators (1G2D). In the
case of 1G2D, we transitioned from 1G1D to 1G2D
in 50 epochs. In the case of 1G4D, we transitioned
from 1G1D to 1G2D in 50 epochs, and from 1G2D to
1G4D in 100 epochs.
Table 1: FID score of generated images by the Generator
trained with each configuration.
hand-defined data distribution
1G1D 95.71 94.74
1G2D 90.33 89.85
1G4D 90.03 87.37
Table 2: FID score of generated images by the Generator
trained with each configuration.
FID
1G4D (4) failed
1G4D (1,4) 96.42
1G4D (1,2,4) 87.37
Figure 7 shows example images generated from
1G1D, 1G2D, and 1G4D respectively. Figure 8 also
shows a comparison of images generated from 1G4D
using hand-defined classes and using data distribution
based classes.
To quantitatively evaluate the quality of these gen-
erated images, we computed the FID values of the
generated images. The results are shown in Table 1.
Table 1 shows that 1G2D and 1G4D can generate
generators with higher performance than the conven-
tional method 1G1D, and that the more discriminators
are added, the higher the performance of the genera-
tor. Also, comparing the second and third columns
of this table, we find that automatic data classifica-
tion based on the distribution of the data can generate
a generator with higher performance than the hand-
defined classes.
8.2 Gradual Increase of Discriminators
Next, we investigated the difference in image genera-
tion results between learning while gradually increas-
ing the number of discriminators and learning with
an increased number of discriminators from the be-
ginning of learning. We compared three cases: (1)
1G4D from the beginning (1G4D (4)), (2) starting
with 1G1D and changing to 1G4D at the 50th epoch
(1G4D (1,4)), and (3) starting with 1G1D, changing
to 1G2D at the 50th epoch and changing to 1G4D at
the 100th epoch (1G4D (1,2,4)). The FID score of
the generator trained in each configuration is shown
in Table 2.
In the case of IG4D (4), the image generation by
the Generator in the early stages of learning was low
accuracy, so the discriminator selection could not be
performed appropriately and training failed. On the
other hand, comparing the results of 1G4D (1,4) and
1G4D (1,2,4), we find that 1G4D (1,2,4) has a higher
generation accuracy of the Generator, and that gradu-
ally increasing the number of discriminators can gen-
CLIP-MDGAN: Multi-Discriminator GAN Using CLIP Task Allocation
469

erate a better Generator.
These results show that when using multiple dis-
criminators, gradually increasing their expertise can
result in better adversarial learning.
9 CONCLUSION
In this paper, we proposed a method to improve the
image generation ability of a generator by adversari-
ally training a generator using multiple discriminators
with different expertise. In particular, we proposed a
method to give multiple discriminators independent
expertise by dividing a dataset so that they have inde-
pendent image features and selecting images for each
discriminator by using CLIP. In addition, we showed
a method to gradually increase the number of discrim-
inators in order to eliminate the instability of training
that occurs when using multiple discriminators.
Experimental results showed that when dividing
the classes that each discriminator is responsible for,
more appropriate expertise is given to the discrimina-
tors when dividing based on the distribution of image
features obtained by CLIP rather than based on fea-
tures thought by humans. It was also revealed that
it is more efficient to give expertise while gradually
increasing the number of discriminators.
REFERENCES
Albuquerque, I., Monteiro, T., Doan, T., Considine, B.,
Falk, T., and Mitliagkas, I. (2019). Multi-objective
training of generative adversarial networks with multi-
ple discriminators. In Proc. International Conference
on Machine Learning.
Brock, A., Donahue, J., and Simonyan, K. (2019). Large
scale gan training for high fidelity natural image syn-
thesis. In International Conference on Learning Rep-
resentations (ICLR).
Choi, J. and Han, B. (2022). Mcl-gan:generative adversarial
networks with multiple specialized discriminators. In
Proc. Conference on Neural Information Processing
Systems (NeurIPS).
Dinh Nguyen, T., Le, T., Vu, H., and Phung, D. (2017).
Dual discriminator generative adversarial nets. In
Proc. Conference on Neural Information Processing
Systems (NIPS).
Durugkar, I., Gemp, I., and Mahadevan, S. (2017). Genera-
tive multi-adversarial networks. In Proc. International
Conference on Learning Representations.
Ghosh, A., Kulharia, V., Namboodiri, V., Torr, P., and Doka-
nia, P. (2018). Multi-agent diverse generative adver-
sarial networks. In Proc. IEEE Conference on Com-
puter Vision and Pattern Recognition (CVPR).
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B.,
Warde-Farley, D., Ozair, S., Courville, A., and Ben-
gio, Y. (2014). Generative adversarial nets. Advances
in neural information processing systems, 27:2672–
2680.
Guzman-Rivera, A., Batra, D., and Kohli, P. (2012). Mul-
tiple choice learning: Learning to produce multiple
structured outputs. In Proc. Conference on Neural In-
formation Processing Systems (NIPS).
Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and
Hochreiter, S. (2017). Gans trained by a two time-
scale update rule converge to a local nash equilibrium.
Advances in Neural Information Processing Systems,
30.
Hoang, Q., Nguyen, T., Le, T., and Phung, D. (2018).
Mgan: Training generative adversarial nets with mul-
tiple generators. In Proc. International Conference on
Learning Representations.
Karras, T., Laine, S., and Aila, T. (2020). Analyzing and
improving the image quality of StyleGAN. In Proc.
IEEE/CVF Conference on Computer Vision and Pat-
tern Recognition (CVPR), pages 8110–8119.
Kingma, D. P. and Welling, M. (2014). Auto-encoding vari-
ational bayes. arXiv preprint arXiv:1312.6114.
Ledig, C., Theis, L., Husz
´
ar, F., Caballero, J., Cunning-
ham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J.,
Wang, Z., and Shi, W. (2017). Photo-realistic single
image super-resolution using a generative adversarial
network. Proc. IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), pages 4681–4690.
Liu, Z., Luo, P., Wang, X., and Tang, X. (2015). Deep learn-
ing face attributes in the wild. In Proc. International
Conference on Computer Vision (ICCV).
Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh,
G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P.,
Clark, J., Krueger, G., and Sutskever, I. (2021). Learn-
ing transferable visual models from natural language
supervision. In International Conference on Machine
Learning (ICML). PMLR.
Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C.,
Radford, A., Chen, M., and Sutskever, I. (2021).
Zero-shot text-to-image generation. arXiv preprint
arXiv:2102.12092.
Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and
Ommer, B. (2022). High-resolution image synthesis
with latent diffusion models. In Proc. IEEE/CVF Con-
ference on Computer Vision and Pattern Recognition
(CVPR), pages 10684–10695.
Ward, J. H. (1963). Hierarchical grouping to optimize an
objective function. Journal of the American Statistical
Association, 58(301):236–244.
Zhang, H., Xu, T., Li, H., Zhang, S., Huang, X., Wang,
X., and Metaxas, D. N. (2017). Stackgan: Text to
photo-realistic image synthesis with stacked genera-
tive adversarial networks. In Proc. IEEE International
Conference on Computer Vision (ICCV), pages 5907–
5915.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
470