Research on Artificial Intelligence Graphic Generation Technology

Youyi Zhang

Media Institute, Hankou University. Wuhan, Hubei, 430212, China

Keywords: Generative Adversarial Network, Variational Autoencoder, Diffusion Model, Graphics Technique.

Abstract: Generative artificial intelligence has been applied to many scenarios in society. It is an innovative technology

field, and its core principle is based on deep learning algorithms and neural network models. From the initial

generation of more blurred and simple images, it is now possible to generate extremely complex and highly

detailed images. The topic of this paper is the development and application of artificial intelligence graphics

technology. This paper lists the development history of image generation and GAN, VAE and Diffusion image

generation technology, but different algorithms have their own shortcomings. With the passage of time, the

algorithm is constantly optimized and the processing power is also enhanced. Contemporary generative

adversarial networks and variational autoencoders, for example, can be used in many aspects of image

generation. In the later stage, this paper also analyzes the possibility of combining image generation with

games. The development prospect of generative artificial intelligence is bright, and it will bring many

conveniences and changes to people's life and work in the future, and promote the development of intelligence

in various industries and different field.

1 INTRODUCTION

Since the birth of artificial intelligence, its functions

have experienced leaps and breakthroughs again and

again. From symbolic logic reasoning to perceptual

cognition, to natural language processing, learning

decision-making, and gradually realize autonomous

action and creativity, giving personalized service and

intelligent decision-making. It can realize the

functions of language synthesis, text generation,

image repair and generation, data prediction,

intelligent recommendation and intelligent painting.

In the present day, artificial intelligence has been

applied to natural language processing, machine

learning, computer vision, robotics, autonomous

driving technology, medical care, agriculture, gaming

and other fields.

This paper will introduce the development of

graphics technology. From the early computer

graphics and texture mapping phase (1960s-1980s) to

statistical learning and traditional machine learning

methods in the mid-2000s (1990s-2000s) to the

present generation adversarial networks (GANs) and

variational autoencoders (VAE) and Diffusion

https://orcid.org/0009-0006-2787-4615

models Diffusion. The image generation techniques

of GAN, VAE and Diffusion are also introduced.

During the 1960s and 1980s, key developments in

early computer graphics and texture mapping

included photorealistic graphics and texture mapping

methods.

There are light reflection models in photorealistic

graphics: Bouknight proposed the first light reflection

model, Gouraud proposed the diffuse reflection

model plus interpolation Gouraud shading treatment,

Phong introduced the Phong illumination model.

These models make computer graphics more realistic.

Solid modeling technology: Since 1973, the

University of Cambridge and the University of

Rochester have developed solid modeling systems

that allow users to build complex models from basic

geometry.

Ray tracing and irradiance algorithm: Whitted

proposed ray tracing algorithm to simulate the

interaction between light and object surface; The

radiosity method introduces multiple diffuse

reflection effects to enhance the realism of the

rendering (Yang, 2012).

However, early computer graphics and texture

mapping were limited by limitations in computational

Zhang, Y.

Research on Artiﬁcial Intelligence Graphic Generation Technology.

DOI: 10.5220/0013233800004558

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

In Proceedings of the 1st International Conference on Modern Logistics and Supply Chain Management (MLSCM 2024), pages 79-85

ISBN: 978-989-758-738-2

performance, rendering quality, texture mapping, and

user interaction.

Texture mapping methods include projection

mapping, which applying 2D textures to map 3D

object surfaces through Projector and UV Mapping.

Affine transform texture mapping and decal texture

mapping are two classical texture mappings. The

former scales, rotates, and distorts textures to increase

dynamic variation. The latter combines multiple

textures to create a rich visual effect.

During the 1990s and 2000s, Statistical learning

and traditional machine learning methods are used in

graphics techniques. Take the classic support vector

machine (SVM) as an example. The development of

SVM dates back to the early 1990s, when it was first

proposed by Vladimir Vapnik and his colleagues at

Bell LABS. The principle of SVM is to find a

hyperplane, which is a dividing line that divides the

sample data into two categories in two-dimensional

space. The goal of SVM is to maximize the interval

from the hyperplane to the support vector, and the

larger the interval, the better the generalization ability

of the model. Classification, regression and anomaly

detection are the main applications of SVM (Gaur &

Mohrut, 2019).

However, in the processing of rich information in

complex images, large amounts of data, high label

requirements, and in the processing of nonlinear

relations, timing information and large-scale data.

The effectiveness of statistical learning and

traditional machine learning methods is limited.

In the second section, the principles of GAN,

VAE and diffusion models are introduced. The third

section describes the application of GAN, VAE and

diffusion models in graphics technology and

discusses the possibility of combining the graphics

technology of artificial intelligence with medicine,

games and other fields. The fourth section is the

summary and future outlook.

2 OVERVIEW OF ARTIFICIAL

INTELLIGENCE GRAPHICS

TECHNOLOGY

2.1 GAN

In 2014, it was proposed by Goodfellow, and its full

name is Generative Adversarial Network. GAN

stands for GANs, and it is a type of deep learning

model. Its basic principle involves at least two

modules: a generator and a discriminator. Through

the mutual game and learning between these two

modules, the output is generated. The task of the

generator is to generate as realistic fake data as

possible in order to deceive the discriminator， it

receives a random noise vector as input，and through

a series of operations and transformations, it outputs

a newly generated data sample. The task of the

discriminator is to distinguish whether the input data

is real or fake data generated by the generator. It

receives both real data and data generated by the

generator, and finally outputs its own decision,

represented as the probability that the input data is

real. In this process, the generator and discriminator

engage in an adversarial training and learning process

(Chai & Zhu, 2019). Figure 1 is a schematic diagram

of GAN.

Figure 1: Schematic Diagram of GAN （ Photo/Picture

credit: Original).

The discriminator needs to continuously learn

from both real data and fake data generated by the

generator to improve its ability to distinguish between

true and false data. The generator adjusts its

parameters based on the feedback given by the

discriminator to generate more realistic fake data,

thereby increasing the probability of deceiving the

discriminator. As the training and adversarial process

progresses, the generator produces increasingly

realistic data, and the discriminator's ability to

identify true and false data also becomes stronger.

When a certain level is reached, the data generated by

the generator becomes so realistic that it is difficult

for the discriminator to distinguish it from the real

data. Through this adversarial learning process, GAN

is able to learn the underlying distribution of the data

and generate new data with similar characteristics.

The optimization process of the generator (G) and

discriminator (D) can be defined as a two-player

game with a minimax problem.

𝑚𝑖𝑛



𝑚𝑎𝑥





D, G



 𝐸

~ 









𝑙𝑔𝐷



𝑥





 𝐸

~













lg 𝐷



𝐺𝑧





)

The advantages of GANs in generating images are

evident, as they can produce realistic and creative

images. However, they also have drawbacks. The

training process is unstable, making it difficult to

control the quality. There is a risk of mode collapse,

MLSCM 2024 - International Conference on Modern Logistics and Supply Chain Management

limiting the diversity of the generated images.

Additionally, GANs can potentially be misused for

malicious purposes. Overall, while GANs possess

powerful capabilities, their shortcomings must be

addressed and improved upon in applications.

2.2 VAE

VAE were proposed by Kingma and Welling. VAE is

a generative model which is mainly used to learn the

potential variable representation of data and realize

the generation and reconstruction of data (Nakai &

Shibuya, 2022). VAE are shown in Figure 2.

Figure2: VAE (Nakai & Shibuya, 2022).

VAE is a model that combines variational Bayes

methods and deep learning. VAE represents image x

through latent variable z and plays a role in image

generation and feature learning. The aim of the model

is to maximize the log-likelihood log 𝑝



(𝑥) of the

image x and compute the approximate posterior

distribution ℒ(𝑞



(

𝑧

𝑥

)

) by variational inference.

Specifically, variational inference is achieved by

optimizing the lower bound ℒ(𝑞



(

𝑧

𝑥

)

). Including

constraint on the KL divergence between

ℒ(𝑞



(

𝑧

𝑥

)

) and the prior distribution 𝑝(𝑧) . The

lower bound consists of two parts. On the one hand,

it penalizes the difference from the prior distribution,

and on the other, it measures the quality of the image

reconstruction for a given potential variable z by

expecting 𝐸

(|)

[𝑙𝑜𝑔𝑝



(𝑥|𝑧)].

In the model, the potential variable z is assumed

to follow a normal distribution.

𝑞



(𝑧



|𝑥



) ≈𝑁 µ





, 𝜎







)

𝑝(𝑧) ≈𝑁 (0, 𝐼)

)





and 𝜎





are the mean and standard deviation of

the j potential variable given the graph 𝑥



. Through

deep learning techniques, the model optimizes the

parameters 𝜃 and 𝜙 , ultimately enabling efficient

representation of latent variables and high-quality

image generation.

VAE can generate new samples similar to training

data and extract high-dimensional data features. It is

easy to manipulate latent space, but the resulting

images are often fuzzy and lacking in detail, can

suffer from pattern crashes, are sensitive to

hyperparameters, and are expensive to train.

2.3 Diffusion

Traditional models GANs learn image features by

creating two neural networks and pitting them against

each other. Gans also need to train generators and

discriminators. While tuning the loss function for

Gans is simple, the learning dynamics (including

trade-offs between generators and discriminators) are

difficult to follow, as are the problems of gradient

disappearance and pattern collapse (when there is no

diversity in the generated samples) (Liang, Wei, &

Jiang, 2020). As a result, training GAN models

becomes very difficult. Diffusion models are easier to

train than Gans to produce diverse and complex

images. The diffusion model can solve the problem of

GAN training convergence because it is based on the

same training data set. This is because it is equivalent

to presupposing the result of neural network

convergence, thus ensuring the convergence of

training (DHARIWAL, 2021), and can predict the

images generated in the middle during training and

give labels in order to classify the images. Next, the

gradient is calculated using the cross entropy loss

between the classification score and the target class

by using the gradient to guide the generation of

samples. In addition, after the introduction of

conditional control diffusion module, the diffusion

model can use heavy control to output a specific style

of image.However, the current diffusion model

algorithm also lacks certain theoretical guarantee,

high computation cost and large memory occupation.

3 GAN, VAE, DIFFUSION IMAGE

TECHNIQUES AND

VARIATIONS

3.1 GAN

Based on the principles of GAN, it can learn the

features and patterns of images as well as the

underlying data distributions. Based on the initial

version of GAN, many variations and improvements

have emerged, such as DCGAN, BiGAN, CycleGAN,

and many others. The following is an explanation of

these different GANs

Research on Artiﬁcial Intelligence Graphic Generation Technology

3.1.1 DCGAN

Deep Convolutional Generative Adversarial

Networks (DCGAN) incorporates convolutional

neural networks into the GAN model. It modifies the

convolutional neural network architecture to improve

the quality of the generated samples and the speed of

convergence. DCGAN possesses better capabilities

for generating images. DCGAN is able to generate

higher-quality images and related models, which to

some extent addresses the previous issue of instability

in GAN training (Liu & Zhao & Ye,2023). This is

because convolutional neural networks have a

powerful ability to process images. Today, DCGAN

has become a fundamental model in the field of image

generation. The generation effect of DCGAN is

demonstrated in Figure 3.

Figure 3: Generation Effect of DCGAN (Liu & Zhao &Ye,

2023).

3.1.2 BigGAN

BigGAN is the first to generate images with high

fidelity and low intra-class diversity gap. It

incorporates the concept of batch size into its training

process. Unlike traditional GANs, BigGAN enhances

the number of convolutional channels and grid

parameters. Additionally, it incorporates truncation

tricks and functions to control model stability. In

addition, there are also variations like BiGAN and

BigBiGAN with their own unique generation effects.

Figure 4 shows the generation effect of BigGAN.

Figure 4: Generation Effect of BigGAN (Liu & Zhao &

Ye,2023).

3.1.3 CycleGAN

CycleGAN is primarily applied in the field of domain

transfer. This refers to the process of transferring data

from one domain to another domain (DHARIWAL,

P, 2021). Its core idea is: Assuming there are domains

X and Y, map Y domain to X domain, and vice versa,

creating a cyclic process. Its core principle is

illustrated in Figure 5.

Figure 5: Core Principle of CycleGAN (Liu & Zhao &

Ye,2023)

CycleGAN can learn the latent distribution

characteristics of similar things and create a mutual

transformation between two domains. The premise is

that there is some commonality between the things in

these two domains, as demonstrated in Figure 6 with

the transformation from a zebra to a horse, and from

summer to winter.

Figure 6: Case Demonstration of CycleGAN (Liu&Zhao

&Ye, 2023).

3.2 VAE

The production of high-quality samples, the potential

space for interpretation, the ability to process missing

data, the smooth potential space, and the wide range

of adaptability and model flexibility all make VAE an

excellent performance in graphics technology.

3.2.1 Model of VAE-GAN

Xu and abdelouahed explored the reconstruction of

multi-spectral images (MSI) from RGB images in

their research (Yang et al., 2019). MSI is used in a

wide range of applications, including satellite remote

sensing, medical imaging, weather forecasting and

the interpretation of artworks. In this study, VAE is

combined with GAN. Specifically, the method

replaces the traditional autoencoder for VAE and

adds an L1 regulator.

MLSCM 2024 - International Conference on Modern Logistics and Supply Chain Management

Two classical datasets CAVE and ICVL were

used in the study. Four quantitative indicators: root

mean square. Square error (RMSE), normal or

relative root mean square. Square error (nRMSE or

rRMSE), peak signal-to-noise ratio. Ratio (PSNR)

and Structural Similarity Index (SSIM). These four

quantitative indicators are applied to the comparison

of VAE-GANs with CNNs and cGANs. Table1

shows the comparison of VAE-GANs with CNNs and

cGANs.

As can be seen from Table 1, when rebuilding

RGB from MSI, RMSE is reduced by 66%. This

shows that the method is more effective in capturing

and reconstructing image features, and can recover

image information more accurately.

Table 1: Comparison of VAE-GANs with CNNs and

cGANs.

Metrics Berk Kin Ours

Approach CNNs cGANs VAE-GAN

Ratio of training

and testing

N/A 50%:50% 50%:50%

RMSE~ (0-255) 2.55 5.649 1.943

RMSE~ (0-1) 0.038 N/A 0.0076

PSNR 28.78 N/A 42.96

SSIM 0.94 N/A 0.99

3.2.2 VAE with Priori Segmentation

Nakagawa and Haseyama et al. combined VAE with

image segmentation prior in their research

(Nakagawa et al., 2021). VAE is an independent

potential variable 𝑧



and 𝑧



for foreground and

background region learning respectively. Model is

shown in the figure7. Table 2 shows the Estimation

Error under different potential variables.

Figure 7: VAE-based model by splitting an image into

several disjoint regions (Nakagawa et al., 2021).

Table 2: Estimation Error of latent variable (The direction

of the arrow indicates whether the estimate error is better to

increase or decrease).

Model: Ours (𝛽 = 100)

Input Estimation Error

None 19.48%

Z 17.40%±0.037%

𝑧



17.40%±0.040%

𝑧



19.48%±0.031%

As can be seen from the data of 𝑧



and 𝑧



Table 2, the study of Nakagawa and Haseyama

verified the disentanglement and transferability of

VAE.

The above two studies demonstrate the

effectiveness, effectiveness, separability and

transferability of VAE.

3.3 Diffusion Model

With the rapid development of diffusion models, their

potential in text image generation has increased

significantly. According to the architecture, the

researchers propose a text image generation method

based on diffusion model. These methods include

cascaded based diffusion models, unCLIP priori

based diffusion models, discrete space based

diffusion models and potential space based diffusion

models.

The cascaded based diffusion model mainly uses

guiding strategies to create high-resolution images.

First, text conditions are entered into the diffusion

model to capture the overall structure of the image

content.

After that, a high-resolution image is generated by

an upsampled diffusion model to improve detail and

maintain authenticity and variety. For example, the

GLIDE model proposed by Nichol et al. (NICHOL et

al., 2022) adopts the cascade diffusion method. First,

Transformer is used to encode the text, and the

encoded text embedding replaces the class

embedding in the ADM model to transform the rough

image with 64×64 resolution. Then, an upsampling

diffusion model is trained to improve the image to

256 ×256 high resolution, and the image details are

refined. In addition, the implicit classifier guidance

strategy used in the training process can ensure the

diversity and fidelity of images while supporting

flexible text prompt generation. However, the

generation of complex prompts still faces challenges

(GAO, Du, & Son, 2024). The GLIDE model

proposed by Nichol et al. adopts the cascade diffusion

method (NICHOL et al., 2022). Firstly, Transformer

is used to encode the text and the encoded text

Research on Artiﬁcial Intelligence Graphic Generation Technology

embedding replaces the class embedding in the ADM

model. The upsampled diffusion model is then trained

to 256 × 256 high resolution to refine the image

details. In the training process, the implicit classifier

guidance strategy can ensure the diversity and fidelity

of images, and support flexible text prompt

generation. However, generating complex text

prompts remains a challenge. As shown in Figure 8.

Figure 8: Cascade based diffusion model （Photo/Picture

credit: Original).

3.4 Discussion

GAN, VAE and diffusion models have their own

advantages. In terms of generation quality, GAN and

diffusion models perform better. In terms of stability,

VAE and diffusion models are better. VAE is simpler

in terms of complexity. In terms of speed, VAE and

GAN are generated more quickly. According to the

different characteristics of GAN, VAE and diffusion

models, they have different effects when combined

with other algorithmic models. As mentioned in the

third section of the graphics technology, DCGAN is

mainly used to generate high-quality images.

BigGAN focuses on generating high-resolution and

diverse images. CycleGAN is used for unsupervised

image style transformation. VAE-GAN works well in

situations where learning is required. The

combination of VAE and image segmentation prior

has better disentanglement. The cascade based

diffusion model can ensure the diversity and fidelity

of images and support flexible text prompt generation.

All three models also have advantages in areas

such as gaming and healthcare. GAN generates high-

quality images that enhance the gaming experience

and improve the accuracy of medical diagnoses. VAE

has the flexibility to provide games with different

styles of design and to generate medical images in

different pathological states. The outstanding

performance of diffusion model in generating quality

and detail recovery is conducive to the de-noising and

recovery of realistic game publicity images and

medical images. Artificial intelligence graphics

technology is not only used in the field of games and

medicine, but also has good applications in

agriculture, meteorology, architecture and other

fields. With the continuous progress of modern social

science and technology, the fields related to images,

graphics and vision will provide sufficient space for

the development of artificial intelligence graphics

technology. It is not only limited to the improvement

of the quality of production content, but also to the

addition of advanced functions such as artificial

intelligence-related automation and interaction.

4 CONCLUSIONS

This article introduces the history of artificial

intelligence in the development of graphics

technology, it introduces the meanings of GAN,

VAE, and DIFFUSION technologies, along with

some application scenarios and extended discussions

and reflections. The future demand for image

accuracy continues to increase, and generation is a

rapidly developing area in artificial intelligence.

Based on deep learning algorithms and neural

network models, people can be assisted in generating

higher-quality images, significantly enhancing the

overall experience. Furthermore, through continuous

learning, AI and large models can produce a diverse

range of image types. Although there were certain

issues with early generative models, over time, VAE

and GAN have leveraged the concept of game theory

to drive the development of the entire field of

adversarial artificial intelligence. Currently, they are

mainly applied in areas such as technology

integration, innovation model optimization and

improvement, application domain expansion,

personalized and customized services, as well as

security and compliance. While they offer numerous

benefits, they also spark additional considerations and

reflections, for instance, addressing resource issues

can be constrained by limitations in datasets, a more

comprehensive system is needed to enrich the models,

reducing instability and confrontation are some of the

securities or deepfake-related social issues that can

arise from AI's involvement in image generation.

Appropriate laws and regulations are also needed to

impose constraints. In conclusion, generative

artificial intelligence has a bright development

prospect. In the future, it will bring many

conveniences and changes to people's lives and work,

driving the intelligent development of various

industries and different fields.

REFERENCES

Yang, S., 2012. Research and Implementation of

Illumination Model in Photorealistic Graphics

Technology. Xidian University.

Gaur, K. & Mohrut, P. 2019. A review on Hyperspectral

Image Classification using SVM combined with

Guided, Filter 2019 International Conference on

MLSCM 2024 - International Conference on Modern Logistics and Supply Chain Management

Intelligent Sustainable Systems (ICISS), Palladam,

India, pp. 291-294

Chai, M. T., & Zhu Y. P., 2019. Research and Application

Progress of Generative Adversarial Networks

Wang M. Q., Yuan W. W., & Zhang, J. 2021 A Review of

Research on Generative Adversarial Networks (GANs).

Computer Engineering and Design, 42(12): 3389-3395.

Nakai, M. & Shibuya, T. 2022. Efficiency of

Reinforcement Learning using Polarized Regime by

Variational Autoencoder, 2022 61st Annual

Conference of the Society of Instrument and Control

Engineers (SICE), Kumamoto, Japan, pp. 128-134.

Liang, J. J., Wei, J. J., & Jiang, Z. F., 2020. A review of

generative adversarial networks Exploration of

Computer Science and Technology.

DHARIWAL, P., 2021. NICHOL A.Diffusion models beat

GANs on image synthesis.

Liu H. D., Zhao X. L., & Ye H. P., 2023. A Review of GAN

Model Research. Internet of Things Technologies,

13(01): 91-94.

Yang, G. Lu, Z. Yang, J. & Wang, Y. An Adaptive

Contourlet HMM–PCNN Model of Sparse

Representation for Image Denoising, in IEEE Access,

vol. 7, pp. 88243-88253, 2019.

Nakagawa, N. Togo, R. Ogawa, T. & Haseyama, M., 2021.

Disentangled Representation Learning in Real-World

Image Datasets via Image Segmentation Prior, in IEEE

Access, vol. 9, pp. 110880-110888.

Nichol, A. Q., Dhariwal, P., Ramesh, A., et al., 2022.

GLIDE: towards photorealistic image generation and

editing with text-guided diffusion models, Proceedings

of the International Conference on Machine Learning.

Gao, X., Y., Du, F., Song, L., & J., 2024. A review of

comparative research on text image generation based on

diffusion model. Computer Engineering and

Applications, 1-23.

Research on Artiﬁcial Intelligence Graphic Generation Technology