---

2.    Why is AI Art Important?
AI technology drives social trend changes —— A tide of art democratization
Major turning points in art history often coincide closely with transformations in the history of technology. Innovations in art theory and form resonate with technological advancements, leading to the development or reconstruction of new social ideas and behavioral habits. In fact, the current era of artificial intelligence is not the first time that technological advancements have brought about new artistic practices.

In the current round of technological explosion driven by AI technology, the proliferation of AI is leading a new wave of art democratization, particularly with the introduction of models such as DALL-E, Midjourney, Stable Diffusion. These tools allow users to generate complex and beautiful works of art with simple prompts, significantly lowering the threshold for artistic creation and allowing ordinary people without professional training to participate in artistic creation. This not only changes the form of art creation but also expands the range of art creators, making art creation unprecedentedly easy and efficient, achieving the democratization of art creation.

The widespread use of AI technology may irreversibly lead to the democratization of art copyright. While AI art is becoming popular, it also sparks broad discussions about art quality, originality, and artistic evaluation standards. Notably, some critiques argue that art generated by AI may lack depth and genuine creativity because these works are often trained based on existing artworks and styles, presenting severe copyright infringement issues. This raises the question of who the real "author" of the work is: the AI developer, the artist using the AI tool, or the original artist providing the training data? This complex copyright attribution issue may prompt existing copyright laws to develop in a more inclusive and adaptable direction, considering the role of creative tools and their contribution to the creative process. AI can replicate and imitate various artistic styles, thereby creating a large number of artworks available for public access and use. The future may promote more copyright management models based on licensing and sharing, such as Creative Commons licenses, which can further facilitate the dissemination and reuse of human and AI artworks.

Each wave of large-scale art democratization has given birth to a new group of artists and collectors, the outstanding representatives of which often become the new nobles of the art world and take the stage of art history. Similarly, in the current wave of art democratization driven by AI technology, with the influx of a large number of users, admirers, and collectors, a group of artists with historical significance and displaying unique personal styles will gradually emerge. The value of their AI artworks will also gradually be recognized and discovered.

---

3.    Collection Value Assessment System
The value assessment of artworks involves several key factors, including historical and cultural value, circulation, and the reputation of the artist. These criteria are equally applicable to AI art. However, due to the prominent technological characteristics of AI art, when assessing its historical and cultural value, one should not only consider its position in the history of art, but also pay attention to its impact and contribution to the history of artificial intelligence engineering. Next, we will discuss in detail the value assessment system applicable to AI art.

3.1 Positioning in the History of AI Engineering
To accurately assess the position of artworks in the history of artificial intelligence engineering, it is necessary to have a deep understanding of AI technologies related to vision and their development. Before 2021, there were four main algorithms that significantly influenced the field of computer graphics, and their developmental history is worth noting: Generative Adversarial Networks (GAN), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Neural Style Transfer (NST). From 2021 to the present, the central figure in the field of AI generation is the Diffusion Model.

■ Generative Adversarial Networks (GAN)
Generative Adversarial Networks (GAN) are an artificial intelligence first proposed by Ian Goodfellow in 2014. GAN involves two neural networks, namely the Generator and the Discriminator, which compete with each other during the learning process. The task of the Generator is to create data that is as close to real data as possible, while the Discriminator's task is to distinguish between the generated data and real data. This setup makes GAN particularly suitable for generating complex data patterns, such as images, music, or text.

• Early Stage (2014)
  

The concept of GAN was originally proposed by Ian Goodfellow in his 2014 paper "Generative Adversarial Nets" when he was phd at the University of Montreal. This paper quickly gained widespread attention as it introduced a novel approach to training networks through adversarial processes, which was an innovative idea at the time.

• Rapid Development and Innovation (2015-2017)
  

In the following years, GAN technology experienced rapid development and diversification. Researchers introduced various variants such as Conditional GAN (cGAN), Deep Convolutional GAN (DCGAN), StyleGAN, CycleGAN, and Wasserstein GAN (WGAN). These variants improved the stability of the original GAN and enhanced the quality of generated data, enabling GAN to be used in a wider range of applications.

• Maturity and Applications (2018-present)
  

By 2018 and beyond, GAN technology started to be widely applied in commercial and research domains, particularly showing immense potential in areas such as art creation, game development, medical imaging, and data augmentation. In the field of image and video generation, GAN are capable of producing high-quality visual content, which is especially valuable when dealing with limited training data.

■ Convolutional Neural Network (CNN)
Convolutional Neural Networks (CNN) are a type of deep learning specifically designed to process data with explicit spatial relationships, commonly used for image and video analysis. CNN extract features from input data through multiple layers of convolutional layers, where each convolutional layer uses a set of learnable filters to capture spatial features. Non-linearity is introduced through activation functions, and pooling layers are used to reduce feature dimensions and enhance the model's generalization capability. Finally, learned features are transformed into the final output, such as classification labels, through fully connected layers. This structure enables CNN to excel in visual tasks as they can effectively recognize patterns and objects in images.

• Initial Concepts and Early Models (1960s-1980s)
  

1960s: Biologists and neuroscientists began to understand how the visual cortex of cats responds to various visual stimuli, inspiring early models of neural networks.
1980: Japanese researcher Kunihiko Fukushima invented the "Neocognitron," a biologically inspired artificial neural network used for handwritten digit recognition. The Neocognitron can be seen as a precursor to CNN, introducing hierarchical structures and the concept of local receptive fields.

• LeNet-5 and Practical Applications (1990s)
  

1998: Yann LeCun and his team designed LeNet-5, the first successful commercial application of CNN. LeNet-5 was primarily used for reading handwritten digits on bank checks. Its success demonstrated the effectiveness of CNN in practical image processing tasks and laid the foundation for future research.
Renaissance of Deep Learning and CNN Breakthroughs (2010s)
2012: Alex Krizhevsky's AlexNet achieved a breakthrough in the ImageNet challenge, marking the arrival of the deep learning era. AlexNet utilized ReLU activation function, dropout to prevent overfitting, and data augmentation. Its depth and complexity surpassed previous models by a large margin.
2014: GoogleNet (Inception v1) won the same challenge, introducing more complex Inception modules that optimized the utilization of computational resources.
2015: Microsoft's ResNet, by introducing residual learning architecture, enabled training of very deep networks (over 150 layers) and greatly improved image recognition accuracy.

• Diversification of Applications and Technological Maturity (2016-present)
  

With advancements in hardware technology and algorithm optimizations, CNN have been widely applied in various fields such as visual systems for autonomous vehicles, medical image analysis, video analysis, and natural language processing.

■ Recurrent Neural Networks (RNN)
Recurrent Neural Networks (RNN) are neural networks specifically designed to handle sequential data and are commonly used in fields such as time series analysis, speech recognition, language translation, and natural language processing. However, RNN also have unique applications in processing image data, particularly in tasks that involve analyzing the temporal dimension of image sequences or videos, such as video classification and action recognition.

• Initial Concepts and Key Breakthroughs (Pre-2000)

1980s: The fundamental theory of RNN was proposed in the 1980s by David Rumelhart, Geoffrey Hinton, and others. Initially, these networks were primarily used for simple sequence prediction tasks but faced challenges in training due to issues like vanishing and exploding gradients, which limited their large-scale applications.
1990s: In 1997, Sepp Hochreiter and Jürgen Schmidhuber introduced Long Short-Term Memory (LSTM), a specialized type of RNN designed to address the gradient vanishing problem in traditional RNN when learning from long sequences. The introduction of LSTM was a milestone in the development of RNN, enabling them to effectively handle longer dependencies in data.

• Applications of RNN and CNN in Image Processing (2010s)

In 2014, researchers started combining Convolutional Neural Networks (CNN) with Recurrent Neural Networks (RNN) for video processing and natural language processing (NLP). In video recognition and analysis tasks, CNN are used to extract spatial features from video frames, while RNN are responsible for tracking the temporal changes of these features, effectively capturing and interpreting the dynamic information in videos. This combination not only optimized the processing of video content but also enabled the generation of descriptive text from images, further enhancing the model's understanding of visual data.2015 and onwards: The alignDRAW developed by Elman and his colleagues at the University of Toronto marked a breakthrough in text-to-image generation. alignDRAW uses RNN to progressively draw images on a canvas based on textual descriptions, showcasing the ability to generate high-quality images from complex textual descriptions. Subsequently, the combined application of CNN and RNN became increasingly popular, particularly in fields like machine translation and automatic speech recognition (ASR).

■ Neural Style Transfer (NST)
Neural Style Transfer (NST) is a technique that utilizes deep learning to apply the style of one image to another, creating visually unique new images. This technology, for the first time, combines artistic style and content synthesis through neural networks, opening up new ways of communication between computer vision and art.

• Preliminary Exploration (Early 2000s)
  

Before the widespread application of deep learning and Convolutional Neural Networks (CNN) in image recognition, the field of image processing was already exploring basic methods for texture synthesis and style transfer. These methods primarily relied on traditional image processing techniques such as image pyramids or frequency-based analysis but were often limited to replicating patterns of simple textures.

• Breakthrough Development (2015)
  

The key breakthrough in neural style transfer was introduced by Leon A. Gatys et al. in their 2015 paper titled "A Neural Algorithm of Artistic Style." They demonstrated, for the first time, the possibility of achieving advanced artistic style transfer using deep learning networks, particularly CNN. Gatys and his team utilized the VGG network, a CNN widely used for image recognition, to minimize the differences in feature representations between a content image and a style image by optimizing a loss function.

• Technological Advancements and Applications (2016-present)
  

Following Gatys' groundbreaking work, neural style transfer quickly became a research hotspot, inspiring a multitude of innovations and developments. Researchers proposed various improved algorithms that optimized efficiency, enhanced result quality, and expanded the range of applications. Style transfer has been applied not only in artistic creation but also in areas such as film production, advertising, video games, and enhancing user-generated content. Due to its visually striking effects, neural style transfer soon became commercialized. Several software and applications, such as DeepArt and Prisma, brought this technology to the hands of ordinary users, enabling them to easily apply the styles of world-class artists to their own photos.

■ Diffusion Model
The working principle of diffusion models involves gradually adding noise to the data until it reaches a completely random state, and then learning an inverse process to gradually remove this noise and recover the original data. These models have demonstrated remarkable performance in generating images, audio, and other complex data, garnering attention for their ability to generate high-quality and detailed samples. Diffusion models simulate the random diffusion process of data, showcasing excellence not only in the field of visual arts but also gradually revealing their potential in other domains such as natural language processing and molecular design.

• Initial Exploration and Theoretical Foundation (2015-2019)
  

The early research on diffusion models focused on the exploration of theoretical foundations and algorithms. It was first proposed by Sohl-Dickstein et al. in 2015, and subsequent studies gradually enhanced its practical applications in image generation.

• Technological Breakthroughs and Algorithm Optimization (2020)
  

In 2020, Jonathan Ho introduced an improved diffusion model called DDPM, which enhanced the generation quality of images. In 2021, the performance of the generative models was further optimized by introducing classifier-guided techniques, making image generation more precise.

• Integration of CLIP and Disco Diffusion (2021)
  

The release of the CLIP model provided a new approach for understanding the correlation between images and text in diffusion models. In the same year, OpenAI combined the language understanding and image matching capabilities of the CLIP model with the image generation capabilities of the diffusion model to develop CLIDE, with the aim of using CLIP's text-image alignment ability to guide the diffusion process, resulting in generated images that better align with the content and style described in the text. Disco Diffusion is a representative component in CLIDE.

• Introduction of DALL-E and Latent Diffusion (2021)
  

In January 2021, OpenAI introduced DALL-E, an image generation model based on CLIP and GPT-3, capable of generating creative images based on textual descriptions. The launch of DALL-E demonstrated the powerful capabilities of combining language models and image generation models.
In the same year, Latent Diffusion was introduced, optimizing data processing and generation efficiency, allowing efficient image generation in compressed latent space, paving the way for generating higher-resolution images in the future.

• Arrival of Stable Diffusion and Midjourney (2022)
  

Supported by the Latent Diffusion framework, Midjourney was released in July 2022, followed by Stable Diffusion in August 2022, significantly changing the landscape of AI art creation. The introduction of these two tools not only popularized high-quality image generation techniques but also greatly reduced the technical barriers, enabling a wide range of creators and developers to easily explore and realize their creative ideas, thereby promoting the democratization and innovation of AI art creation.

• Large-scale Applications and Future Prospects (2022-present)
  

With the rapid development in the fields of multimodal, 3D generation, and video generation, diffusion algorithms have been widely applied and explored. For example, in February 2024, OpenAI launched the Sora model, marking a major breakthrough for diffusion algorithms in the field of video generation.

■ Critiques: Does the inclusion of an algorithm output in significant research papers always indicate its collection value?
In reality, from the perspective of commercialization feasibility, many algorithm outputs that hold significant positions in the history of AI engineering are difficult to transform into artistic assets. To illustrate this situation vividly, this article selects positive and negative examples for analysis.

a)    Generative Adversarial Net
In the seminal paper "Generative Adversarial Nets" published by Ian J. Goodfellow in 2014, a series of datasets were used to train the Generative Adversarial Network (GAN), including MNIST, the Toronto Face Database (TFD), and CIFAR-10. The algorithm outputs are as follows.

The unconventional prompts in alignDRAW reveal the immense potential of AI algorithms in exploring surrealistic art. These absurd prompts serve not only as technical tests for the algorithm but also unintentionally pave the way for new avenues of artistic expression. AI has the ability to break free from the logical constraints of the real world and create imaginative and innovative images. This practice demonstrates the unique value of AI as a tool for artistic creation, capable of translating human creative ideas into visually captivating artworks, thus expanding the boundaries and forms of art.

b)    A Book from the Sky
A Book from the Sky is the first GAN art project created by AI art pioneer Gene Kogan in December 2015 using a Convolutional Generative Adversarial Network (DCGAN). In the description of this project on his personal website, it is explicitly mentioned that it was inspired by the artwork "A Book from the Sky" created by Chinese contemporary artist Xu Bing in 1988. It was also one of the earliest AI artworks with a clear artistic orientation created using DCGAN, shortly after the release of the DCGAN paper in November 2015. In December 2015, this artwork caught the attention of Yann LeCun, one of the Turing Award winners in 2018 and one of the giants in the field of AI, who gave it a special mention on Facebook.


■ Summaries
The development of AI art has sparked a new wave of artistic democratization, with its starting point being the introduction of the DALL-E model by OpenAI in January 2021, followed by the emergence of Stable Diffusion and Midjourney. The availability of these accessible AI models has enabled the general public to engage with and use AI art, which was previously unimaginable. Prior to this, due to the high technical barriers, AI art models were primarily used by professional artists and researchers. Therefore, in terms of the widespread adoption of AI technology, the development of artificial intelligence can be clearly divided into the era of AI art technology before 2021 and the era of popularized AI art from 2021 onwards. In the era of AI art technology before 2021, many AI artworks were the output results of breakthrough algorithm models that held significant historical positions and contributions in the history of artificial intelligence.
In the era of AI art technology, 2015 was undoubtedly a turning point for the field of AI art, particularly evident in the introduction of the Diffusion model and advancements in Generative Adversarial Networks (GAN) and Neural Style Transfer (NST). In that year, alongside key technological innovations in Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) and their notable academic achievements, AI also made a qualitative leap in artistic creation applications. Important progress during this time included the introduction of the groundbreaking deep learning model ResNet and the birth of alignDRAW, the first system capable of generating images based on textual descriptions.

It is worth noting that 2018 was a crucial year for the development of GAN in the field of AI art. Since that year, GAN technology has made significant progress in generating high-quality artwork and has started to be widely applied in commercial and research domains. According to interviews conducted by the author with AI art pioneer Gene Kogan and statements made by Ivona Tau during an AI art course held at the Vertical Gallery, it can be observed that starting from 2018, an increasing number of artists have entered the field of AI art and begun using GAN technology for their creations. This phenomenon demonstrates that there were relatively few artworks created using GAN before 2018, thus giving these early works a higher degree of scarcity value in the market.
From the perspective of asset issuance and operation, the low survival rate of AI art from 21 years ago can be attributed to factors such as limited issuance, difficulties in copyright verification, and a lack of artistic qualities. Firstly, as mentioned earlier, the high barriers to creating AI art at that time resulted in a low number of assets being issued. Additionally, the process of assetizing AI artworks from 21 years ago often involves copyright verification, which typically requires artists to provide academic publications, social media records with timestamps, or website documentation. As many AI artworks were not given sufficient attention at the time, it became challenging to prove their creation time, thereby hindering the conditions for assetization and further reducing the survival rate of these artworks. Even for AI output images with verified copyrights, it is necessary to consider whether they possess sufficient human subjectivity or artistic intent to evaluate their artistic value. Due to these demanding conditions, the number of early AI output artworks suitable for art commodification is scarce, further proving the collectible value of AI artworks from the pre-2021 AI technology era. Regarding the asset pricing of these early AI artworks, one can consider using the academic contribution of the underlying algorithm models as an evaluation criterion, quantified by factors such as the impact factor of the model's paper. This approach not only scientifically reflects the innovative value of the artworks but also provides a new pricing reference for the art market.

3.2 Influences in Art History
Some AI artworks have left a unique mark in art history due to their legendary stories. One prominent example is the artwork "AI Generated Nude Portrait #7" created by artist and programmer Robbie Barrat, also known as Lost Robbie.

In 2018, at the inaugural Art+Tech Summit held by Christie's, the conference focused on exploring the intersection of emerging technologies and art. During the London event, the digital art marketplace SuperRare gifted each attendee with a gift card containing an AI artwork created by Barrat using GAN algorithms. At that time, blockchain-based tokenization of digital art was still a novel concept, and non-fungible tokens (NFTs) were not yet widely known. The artwork "AI Generated Nude Portrait #7," specifically created by Barrat for this event, was divided into 300 unique pieces, and attendees could claim ownership of the artwork online using a special gift card.

     


©️ 2024 N Pole Art. All rights reserved