The Reconstruction of Image Aesthetics Driven by the Development

of Artificial Intelligence Technology

Chengze Li

Beijing 101 High School, Beijing, China

Keywords: AI, Deep Learning, Datasets, Aesthetics.

Abstract: The remarkable growth of artificial intelligence (AI) has deeply affected contemporary art creation. Image

generation, as a result of the merging of AI and art, is reshaping the boundaries of art creation through a

distinctive method of creation and infinite potential. This paper reviews the technological changes in machine

learning and deep learning and their impact on the popularization of AI art creation. From the perspective of

painting production, relevant mainstream datasets, algorithms, and models are introduced, including

convolutional neural network models, generative adversarial network models, and diffusion models. This

paper further analyzes key element recognition in AI painting and image aesthetic quality assessment. AI not

only challenges the position of the subject in artistic creation but also brings about revolutionary changes in

artistic aesthetic standards and evaluative criteria. As artificial intelligence technology continues to develop,

its integration with art can effectively promote interdisciplinary collaboration and open up new possibilities

for innovation, expression, and cultural exchange in the digital age.

1 INTRODUCTION

Artificial intelligence (AI) is propelling a

technological revolution unprecedented in human

history. As pointed out by Paul Doherty, human

society is entering an era of coexistence and

symbiosis with machines (Doherty & Wilson, 2018).

In this revolution, the development of generative AI

art has achieved the transformation of AI from

imitation to innovative creation. Early computer

graphics laid the foundation for AI art, and in recent

years, breakthroughs in technologies such as deep

learning, convolutional neural networks (CNNs),

generative adversarial networks (GANs), and

diffusion models (DMs) have led to significant

progress in AI art in terms of creative means, style

expression, and aesthetics. While these works not

only introduce an unprecedented visual sensation to

people, they also inspire profound contemplation of

aesthetic concepts and artistic expression. As a

multidimensional perceptual cognitive activity, art

appreciation involves aesthetic judgment, semantic

perception, and emotional resonance, and AI is

gradually approaching it with computational models.

https://orcid.org/0009-0004-8181-8556

This article first reviews the evolution of artificial

intelligence technology, especially the application of

machine learning and deep learning in artistic

creation; then sorts out the commonly used datasets,

algorithms and models in AI art creation; then

discusses the specific application of technologies

such as convolutional neural networks, generative

adversarial networks, and diffusion models in image

generation; finally, analyzes the performance of AI

art in aesthetic reconstruction. This article attempts to

explore the mainstream technologies and cases at the

intersection of AI and art, and to outline future

technological prospects by integrating computer

science and aesthetics.

2 AI TECHNOLOGY CHANGES

AND ART

American philosopher John Searle divides the

development of AI into three stages: weak AI, strong

AI, and super AI (Searle, 1980). The mainstream

view holds that technology is still in the stage of weak

AI. In 1950, Alan Turing proposed the famous

Li, C.

The Reconstruction of Image Aesthetics Driven by the Development of Artiﬁcial Intelligence Technology.

DOI: 10.5220/0014362300004718

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 2nd Inter national Conference on Engineering Management, Information Technology and Intelligence (EMITI 2025), pages 527-534

ISBN: 978-989-758-792-4

527

"Turing Test," marking the beginning of the AI era.

Over the next 75 years, various AI technologies have

tried to pass this test (Mitchell, 2024). From 1950 to

1989, AI was in its early stages of exploration, with

technical models mainly being knowledge-driven

symbolic models and focused on narrow tasks.

Painting techniques were associated with simple,

rule-based symbols and geometry, such as in software

like Photoshop. Between 1990 and 1999, AI entered

the application stage, establishing data-driven image

models and early shallow statistical methods, such as

those used by Google. From 2000 onward, AI entered

the deep learning stage, where technical models

driven by knowledge, data, algorithms, and

computing power generated content, such as CNNs

and GANs. Since 2020, artificial intelligence has

entered the stage of large-scale training, as seen in the

emergence of the GPT series.

2.1 Machine Learning

Machine learning is a method by which computer

systems learn to perform a task using data to improve

at it. The history of machine learning, from 1950 to

1960 till now, has been punctuated by various

milestones like the invention of the Perceptron, a

graphic recognition machine designed to simulate

human neural control systems that gave birth to

neural networks. Machine learning development

benefits from three key ingredients: compute, data

and algorithms. Machine learning has made

tremendous progress in recent decades. In particular,

breakthroughs in natural language processing, e.g.,

the birth of ChatGPT, have enabled a new age for AI.

The theory of Artificial Systems, Computational

Experiments, and Parallel Execution (ACP) was

proposed by Li and others from the Chinese Academy

of Sciences. This approach proposes human

participation in control and learning, enhancing the

efficiency and fairness of machine learning, and

achieving closed-loop learning and optimization

results (Li et al., 2021).

If ACP is placed in the realm of art, human artists

(PA), virtual machine artists (MC), and physical

machine artists (MP) can collaborate on artworks

together. The PA is the starting point and criterion in

the process of artistic creation. The MC mimics the

PA's creation in virtual space, and the MP completes

the art in the real physical world under the MC's

guidance; its work is judged by the PA to optimize the

intelligent painting system. Under this framework,

artistic creation is no longer confined to individual

talent but rather the result of the collaborative wisdom

of humans, machines, and algorithms.

2.2 Deep Learning

Deep learning is a sub-field of machine learning,

which is a technique based on artificial neural

networks. It utilizes a multi-layer network structure to

learn high-level features of data and has achieved

tremendous success in image recognition, speech

recognition, and natural language processing. The

key technologies of deep learning include CNN,

GAN, etc.

CNNs and GANs are representatives of the use of

deep learning in the field of AI art. CNNs are key

deep learning models for image recognition and

classification. They can automatically learn image

feature representations and are widely used in tasks

such as image classification and object detection.

GANs comprise a generator and a discriminator,

which can produce new data samples through

adversarial training, enabling applications in image

generation and style transfer. Image recognition and

classification are the most popular applications of

deep learning, but they have limitations, such as high

data requirements and a potential for incorrect

interpretation.

Painting style transfer is an interesting research

task in computer vision. The algorithm based on deep

learning realizes style transfer by training neural

network models, and its core is two loss functions,

namely content loss and style loss. Content loss

quantifies how the content image differs from the

generated image, while style loss quantifies how the

style image differs from the generated image.

Minimization of both content loss and style loss leads

to the generation of stylized images that are both

content-wise similar to the target image and style-

wise similar to the target style image.

3 DATA, ALGORITHMS,

MODELS AND AI ART

The training data serves as the raw material for

training AI models. A significant amount of data is

needed for training the AI models, which are usually

gathered from online image databases, digital

collections from museums, user-uploaded images,

etc. Trained with large-scale data, the model is

capable of identifying and learning features and

patterns. For the models that create artworks, the

trained model learns about different artists' styles,

color palettes, composition, etc. During training, the

model continuously updates its parameters so that

higher quality and a higher similarity between the

EMITI 2025 - International Conference on Engineering Management, Information Technology and Intelligence

528

generated images and the corresponding ground truth

are achieved.

3.1 Datasets

The implementation of AI art features requires high-

quality datasets. Images in high-quality datasets

typically have high resolution, low noise, and rich

details, which are crucial for generating high-quality

artworks, while low-quality data can result in blurry

or distorted generated images.

After obtaining the original datasets, it is often

necessary to do some data cleaning work on the data

before model training begins. Data cleaning involves

removing duplicate, erroneous, or low-quality data.

The cleaned datasets ensure that the model is not

affected by useless or harmful data during training. In

addition to data cleaning, the data also needs to be

labeled. The labeled dataset can help the model better

understand the content of the data. When Madhu et al.

studied the dataset of ancient Greek vase paintings,

they added the bounding boxes of the figures and the

corresponding key points of body postures to the

original dataset. Ultimately, in the style transfer

process, compared with unlabeled datasets, the mean

average precision and mean average recall rate have

improved significantly by more than 6% (Madhu et

al., 2022).

Fairness and unbiasedness are important metrics

of a quality model, which can ensure the diversity and

representativeness of the datasets and also reduce the

bias of the model's generated works. Including works

of art from different genders, races, and cultural

backgrounds can prevent the generated works from

having a single perspective or discriminatory

misinterpretation. By July 2025, Wikiart had

collected around 250,000 artworks, including works

by more than 3,000 artists, but with a bias towards the

West, with half of them from Western Europe. If this

were the way to develop the global art model, there

would inevitably be deviations. Researchers need to

adjust and optimize the content of the dataset in a

timely manner to meet the requirements.

Wang, working on an algorithm model based on

an improved NTS network, was unable to obtain

feature evidence from existing public datasets and

relied on self-collection and organization to build a

new dataset. Eventually, his model achieved a

recognition rate of 95.1% for Thangka art style

effects. The model's recognition effect on the Oil

Painting dataset was only 73.8% (Wang, 2022).

3.2 Algorithms and Models

In the field of AI, there are two closely linked but

different notions: algorithms and models. An

algorithm is a series of clear and orderly steps and

principles for solving a specific problem and

completing a specific task. It can be an algorithm or a

program. A model is a mathematical model or

calculation framework formed by learning from

training data with a specific algorithm. A model is the

output of running an algorithm on training data,

which captures the characteristics of the data and can

be used for prediction, classification, or data

generation. Frequently used models include the linear

regression model, decision tree model, CNN, GAN,

diffusion model (DM), variational autoencoder

(VAE), etc.

3.2.1 Convolutional Neural Network Models

CNNs are the core model in deep learning for

processing grid-structured data and have developed

into a variety of classic architectures over the years.

Some of CNN's classic models are widely used in the

field of AI art, such as the VGG model, ResNet

model, InceptionNet model, etc. The VGG model,

proposed by the Visual Geometry Group at the

University of Oxford, is characterized by "small

convolution kernels + great depth". By stacking

convolutional layers with multiple 3×3 small

convolutional kernels and 2×2 max pooling layers to

build a deep architecture, it can automatically and

efficiently extract features from lower to higher levels

of the image (Simonyan & Zisserman, 2014).

VGGNet performs well in tasks such as image

classification. It can apply the style of one image to

another, generating works with specific artistic styles.

The ResNet model, proposed by Microsoft,

addresses the vanishing gradient problem in deep

networks through "residual connections" and, for the

first time, increases network depth to over a hundred

layers, solving the degradation problem of deep

networks and enabling the training of extremely deep

networks. Versions such as ResNet50 are often used

for classifying artworks. By learning the features of

paintings of different styles, it can classify and

recognize artworks and also provide a feature

extraction basis for style generation in artistic

creation.

The InceptionNet model was proposed by Google.

In 2015, Google introduced DeepDream, a deep

learning-based image processing technology based

on pre-trained Inception series models, which

amplifies specific features in images through

The Reconstruction of Image Aesthetics Driven by the Development of Artiﬁcial Intelligence Technology

529

techniques such as gradient ascent to generate images

with a fantastical and surreal style.

Researchers have been optimizing the model. The

Valentin model, based on the ResNet architecture,

was trained on 40,000 images. It improved

classification quality through a variety of methods,

including data augmentation, scheduler optimization,

parameter selection, and loss function tuning. At the

same time, analytical preprocessing was carried out

on the existing painting image dataset. The model

achieved a classification accuracy of 51.5% for five

types of paintings. To reduce overfitting and

computational time, the authors truncated the weight

parameters of the neural network. For categories with

significant differences, the model accuracy improved

significantly to 91.25% (Kovalev & Shishkin, 2020).

In the comparative study by Sethi et al. on the

architectures of VGG-16 and VGG-19, the

comparison of different performance parameters was

discussed, which shows which parameter is more

convenient to use for which particular use case. The

study results show that after 1,000 iterations, the style

loss of VGG-16 and VGG-19 has almost similar

performance, but the difference in the value of

content loss is 73.1%; however, VGG-19 performs

better because it minimizes both content loss and

style loss (Sethi et al., 2022).

Yin et al. combined the historical experience of

CNNs and proposed an algorithm for DeepInversion,

reusing pre-trained mature CNN architectures such as

ResNet50, VGG16, and InceptionV3 as knowledge

sources. This is a data-free knowledge transfer

method, with an accuracy improvement of more than

10% compared to the original model (Yin et al.,

2020).

3.2.2 Generative Adversarial Network

Models

GANs are an algorithmic model involving adversarial

training between a generator and a discriminator.

They are a machine learning model proposed by

Goodfellow et al. (Goodfellow et al., 2014). The

technical framework of GANs is shown in Figure 1.

GANs are based on the unsupervised learning of two

artificial neural networks called the generator and the

discriminator, both of which are trained under the

concept of adversarial learning.

Figure 1: GAN infrastructure (Goodfellow et al., 2014).

In the art of aesthetics, GANs can be modeled as

two confronting artists and critics interacting with

each other to improve the art itself and to make the

work extremely realistic. In GANs, the generator tries

to generate new data so that it appears to be real data

without knowing how the real data is generated;

however, it aims at making the generated data

realistic as if they are actually real and hard to

discriminate. At the same time, the discriminator

behaves like a connoisseur.

Liu et al. adopt the structure of GANs to study

visual effects created by the seamless merge of

Unity3D. The generator and the discriminator both

utilize the Transformer model, which handles

sequential data and long-range correlations

effectively. The obtained experimental results

indicate that the inception score (IS) is 8.95, with the

Frechet Inception Distance (FID) being 20.1,

showing good diversity and image quality (Liu et al.,

2025).

3.2.3 Diffusion Models

In the machine learning domain, DMs are generative

models that can produce new data similar to the

training data. The main idea lies in using random

processes to start from a simple and manageable state

and gradually introduce increasingly complex

structures through some process, thereby generating

data similar to the target data.

The extended model can be thought of in art as

creating an artwork from nothing, bringing structure,

details, etc., little by little from a chaotic, unordered

mess to the piece itself. This is commonly done in

EMITI 2025 - International Conference on Engineering Management, Information Technology and Intelligence

530

several steps or "time steps," beginning from a state

that is always simple and introducing some

randomization with control at each time step, so that

the final data gradually takes on a structure

resembling the target data. DMs are similar to an

artist's step-by-step iteration in the creative process:

starting with a simple initial thought, through

continuous adjustment and addition of details, and

finally rendering a beautiful work.

Figure 2: Data mapping of pixel space to Latent space (Rombach et al., 2022)

4 AESTHETIC

RECONSTRUCTION OF AI

ART

In the traditional sense, aesthetics usually refers to the

aesthetic experience, emotion, and evaluation of the

object of aesthetic appreciation. The emergence of AI

art has raised a series of aesthetic questions and

controversies. AI is both a connoisseur and a creator.

It integrates technologies such as computer vision,

machine learning, and deep learning to perform

feature analysis, style recognition, emotion

interpretation, and value assessment on a vast number

of artistic paintings, while generating a large number

of works. The 20th-century thinker Benjamin once

believed that the contemporary era was the "age of

mechanical replication" of art, and the "aura" of

artworks was vanishing (Benjamin, 1936). However,

the advancement of AI technology is unstoppable,

and the quality of its works even surpasses that of

human-created ones. Théâtre D'opéra Spatial (Spatial

Opera Theater), a painting by Jason Allen, was

generated by the AI drawing tool Midjourney and

retouched in Photoshop. This award-winning work

has sparked discussions about the subject of art, and

the boundaries of artistic creation are constantly

expanding and blurring.

4.1 Identification of Key Elements

The Stable Diffusion Model is one of the most

popular AI painting base models today. The latent

method on which the Stable Diffusion Model is based

is the Latent Diffusion Model (LDM). It was

proposed by Robin Rombach and Patrick Esser of AI

video company Runway, who are also important

participants in the subsequent development of Stable

Diffusion. The technical structure of LDM is shown

in Figure 2. Its near-optimal balance between

reducing complexity and retaining details

significantly enhances visual fidelity (Rombach et al.,

2022). From an artistic perspective, if a painter is

creating a watercolor painting, through the diffusion

model, it is like the painter constantly adding ink and

noise, making the painting appear somewhat blurry.

At this point, the stable diffusion model acts as a

protective layer that moves the image from the pixel

space to the latent space, and then adds ink and noise

to the latent space. This protective layer ensures that

the changes do not have an irreversible impact on the

painter's work. Ultimately, it makes the presentation

of the painting orderly.

The Reconstruction of Image Aesthetics Driven by the Development of Artiﬁcial Intelligence Technology

531

Figure 3: Common Classic painting styles

The complexity of painting, which is the diversity

of styles and the various tools, materials, and media

chosen by different artists, makes it one of the most

challenging tasks in computer vision. The AI model

integrates several key dimensions for style

recognition. Among the color features, classicism has

low-saturation tones, such as the blue-gold palette in

The Rape of Europa (Figure 3). Impressionism

features high contrast and vivid colors. Modernism,

on the other hand, has bold, pure tones. Post-

impressionism features collages and mixed colors. In

stroke analysis, CNNs analyze the direction,

intensity, and pattern of strokes, such as short strokes

with high rotation relative to the smooth texture of

realism. The ResNet model distinguishes between

Expressionist rough strokes and Impressionist loose

strokes. In terms of composition, classicism

emphasizes symmetrical order, while Impressionism

emphasizes natural spontaneity. Modernism is

characterized by perspective subversion, while post-

impressionism is characterized by fragmented

recombination. In semantic features, classicism

includes mythological figures, while postmodernism

features popular symbols. Technically, the dot-stroke

features of Impressionism enhance the accuracy of

style classification. In Warhol's Marilyn Diptych, AI

detects repeated print blocks that link it to 20th-

century screen printing, in line with postmodernist

diversity, fragmentation, and anti-traditionalism.

For the oil painting style classification task,

researchers first extract the painting features and then

model and classify the features. Folego et al., in their

research on the authenticity of Van Gogh's paintings,

divide each painting into nonoverlapping image

blocks, keeping 196.3 pixels per inch. They use a

multimodal fusion method that combines manual

features with CNN features. In analyzing artistic

style, short and fast brushstroke direction features can

provide the most useful information to identify Van

Gogh's works, while high-saturation contrasting color

features provide more discriminative ability to

distinguish between Van Gogh's different artistic

periods. This method achieved an accuracy rate of

92.3% in distinguishing Van Gogh's works,

performing better than traditional methods that only

used color or brushstrokes (Folego et al., 2016).

4.2 Aesthetic Evaluation of Images in

the Context of AI

Mainstream models such as GANs and DMs still have

problems such as distorted detail rendering and

mechanical

style transfer, making it difficult to

Figure 4: A postmodernist makeover of the Mona Lisa by an AI called Dreamina

EMITI 2025 - International Conference on Engineering Management, Information Technology and Intelligence

532

accurately capture the deep emotional expression of

artworks. When AI stares at The Rape of Europa, it is

unable to distinguish mythological implications from

chaotic pixels. When it stares at Warhol's Marilyn

Monroe, it sees repetition rather than an attack on

consumerism. Figure 4 shows a similar style

transformation of the Mona Lisa after the style of

Marilyn Diptych was identified using an AI called

Dreamina. Although the work retains postmodernist

features, the new work has problems such as

mechanical replication and incomplete composition.

Image Aesthetic Quality Assessment (IAQA) is a

computer vision topic that consists of multi-

disciplinary knowledge. Essentially, it aims to

approximate human subjective judgment of an

image's "aesthetics" through algorithms to realize

automated analysis and assessment of an image's

aesthetic quality. It makes the abstract concept of

aesthetics measurable and computable. The

mainstream method is the deep learning method,

where convolutional neural networks such as

VGG/ResNet automatically learn the deep aesthetic

features of images; with a trained model using a

labeled large-scale training dataset, the models learn

humans' aesthetic preferences from the training data.

Chounchenani et al. summarized the research on

IAQA in input-processing-output deep learning over

the past decade. Based on the Atomic Visual Actions

dataset, the experiment compared different deep

learning methods, including ResNet-50, VGG-16,

InceptionNet-V3, and others, in image aesthetic

quality assessment, with the evaluation method

achieving the highest accuracy rate of 91.5%. It is

noted that deep learning models still need further

improvement (Chounchenani et al., 2025).

The integration of IAQA and style transfer aims

to apply the "aesthetic judgment capability" of IAQA

to achieve better generation effects in style transfer.

Before style transfer, IAQA analyzes the aesthetic

defects of the original image, such as imbalanced

composition and lack of contrast, to clarify the

optimization targets of the style transfer algorithm.

IAQA can also act as an aesthetic supervisor to

constrain style parameters after style transfer.

Additionally, IAQA can learn different users'

aesthetic preferences, thus enabling style transfer to

generate results that conform to specific users'

aesthetic preferences.

5 CONCLUSIONS

This article summarizes AI technologies and

applications related to painting art, including machine

learning, deep learning, etc. It further analyzes

mainstream datasets, algorithms, and models based

on AI in the field of painting. Finally, this paper

describes AI's style transfer and aesthetic

reconstruction in painting art. Although AI shows

great potential, existing technologies still have much

room for improvement in terms of data quality, model

interpretability, and practical application. There are

three levels of room for technological progress. The

first is the optimization of existing AI datasets,

computing power, and models. The second is the

cross-integration and application of multiple models

to enhance big data training capabilities. The third is

the integration of AI with other technological fields.

Technological advancements in the field of AI art can

combine knowledge from both technical and artistic

fields, and drive AI's advancement in the field of art

based on "aesthetic" target requirements.

REFERENCES

Benjamin, W. (1936). The work of art in the age of

mechanical reproduction. New York.

Chounchenani, M. D., Shahbahrami, A., Hassanpour, R., &

Gaydadjiev, G. (2025). Deep learning based image

aesthetic quality assessment: A review. ACM

Computing Surveys, 57(7), Article 183.

Daugherty, P. R., & Wilson, H. J. (2018). Human +

machine: Reimagining work in the age of AI. Harvard

Business Review Press.

Folego, G., Gomes, O., & Rocha, A. (2016). From

impressionism to expressionism: Automatically

identifying van Gogh's paintings. In 2016 IEEE

International Conference on Image Processing (ICIP)

(pp. 141–145). IEEE.

Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B.,

Warde-Farley, D., Ozair, S., ... & Bengio, Y. (2014).

Generative adversarial nets. Advances in Neural

Information Processing Systems, 27.

Kovalev, V. Y., & Shishkin, A. G. (2020). Painting style

classification using deep neural networks. In 2020

IEEE 3rd International Conference on Computer and

Communication Engineering Technology (CCET) (pp.

334–337). IEEE.

Li, S., Wang, Y., Wang, X., & Wang, F. (2021). Mechanical

design paradigm based on ACP method in parallel

manufacturing. In 2021 IEEE 1st International

Conference on Digital Twins and Parallel Intelligence

(DTPI) (pp. 1–4). IEEE.

Liu, Q., Li, J., & Hu, W. (2025). Exploration of cross-modal

AIGC integration in Unity3D for game art creation.

Electronics, 14(6), 1101.

Madhu, P., Villar-Corrales, A., et al. (2022). Enhancing

human pose estimation in ancient vase paintings via

perceptually-grounded style transfer learning. ACM

The Reconstruction of Image Aesthetics Driven by the Development of Artiﬁcial Intelligence Technology

533

Journal on Computing and Cultural Heritage, 16(1),

Article 16, 12–17.

Mitchell, M. (2024). The Turing test and our shifting

conceptions of intelligence. Science, 385(6710).

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., &

Ommer, B. (2022). High-resolution image synthesis

with latent diffusion models. In 2022 IEEE/CVF

Conference on Computer Vision and Pattern

Recognition (CVPR) (pp. 10674–10685). IEEE.

Searle, J. R. (1980). Minds, brains, and programs. The

Behavioral and Brain Sciences, 3, 417–457.

Sethi, P., Bhardwaj, R., Sharma, N., Sharma, D. K., &

Srivastava, G. (n.d.). A deep learning-based neural style

transfer optimization approach. Intelligent Data

Analysis, 29(2), 320–331.

Simonyan, K., & Zisserman, A. (2014). Very deep

convolutional networks for large-scale image

recognition. CoRR, abs/1409.1556.

Wang, H. (2022). Research on painting image style

classification and application based on deep learning

(Master's thesis). Qinghai University.

Yin, H., Molchanov, P., Alvarez, J. M., Li, Z., & Kautz, J.

(2020). Dreaming to distill: Data-free knowledge

transfer via DeepInversion. In 2020 IEEE/CVF

Conference on Computer Vision and Pattern

Recognition (CVPR) (pp. 8712–8721). IEEE.

EMITI 2025 - International Conference on Engineering Management, Information Technology and Intelligence

534