AIGC Series – AI-generated Art’s Technology Development & Commercialisation

AI-generated art has become popular since 2022, which is regarded as the first year of the AIGC boom. Stability AI seized the opportunities brought by the AIGC boom and raised USD 100m in October last year. Its valuation once reached USD 1bn. Could you update us on the breakthroughs in the efficiency and technologies of AI art generation algorithm models over the past 1-2 years? What are the growth drivers for the AIGC industry?

Could you elaborate on AI art generation tools, including open-source models such as Midjourney, Disco Diffusion, Stable Diffusion and DreamBooth? Could you compare these models in terms of background, level of openness, deep learning features as well as advantages and disadvantages?

You mentioned diffusion-based and non-diffusion-based AIGC algorithm models. What are the respective advantages and disadvantages of these two types of models? Do you think any type of model will become the mainstream in the AIGC industry, or will there be assorted models on the market?

You mentioned that Chinese companies may adjust the parameters of AIGC algorithm models they use. Could you compare different open-source AI art generation platforms in terms of the time and costs needed for image generation or model training?

It is reported that the first AI art generator companies are already going out of business, such as the overseas StockAI. Some Chinese companies might have encountered the same situation. Some reports cited high operating costs and difficulties in monetisation as the causes. Domestic players need to train the models, deploy computing power and acquire talent in the initial stage. What are their costs and cost structure?

The Chinese-style art generator model “Taiyi” is the first open-source Chinese model based on Stable Diffusion and received much press coverage. What is your opinion on Taiyi? How did International Digital Economy Academy (IDEA) train and develop this model? How do you estimate the cost of operating this model?

The typical Chinese AIGC platforms only add a translation plug-in to the original English models, whereas Taiyi allows direct Chinese input. How does this empower domestic AIGC platforms? Or does it have a limited impact because translation plug-ins already do a decent job?

Domestic AIGC players include Baidu’s Wenxin Yige, Yijian AI, Printidea.art, Wujie AI, your employer Picup.ai, Tiamat, Wenge Group’s Hanmo Danqing (翰墨丹青), Chushou AI (触手AI), 6pen, Draft, Lingjing, Meitu and Wondershare’s Filmora. Which do you think are notable at the moment? What are the current competitive tiers and landscape?

What models do domestic AIGC companies adopt? Do they deploy multiple models? According to your knowledge, is the algorithm model behind Baidu’s Wenxin Yige more special? What do you think of the advantages and disadvantages of different players in adjusting algorithms and model training?

What are the core competencies of tier 1-2 players that use Stable Diffusion as the base model? What are their investment focuses?

You just said that AIGC monetisation is now concentrated on individual customers. What type of customers will pay for AIGC services? How big is the consumer market size? What are some opportunities in business-oriented AIGC and what is the potential market size?

For clients such as brand enterprises, what benefits can they enjoy from purchasing AI-generated paintings in terms of reducing costs and improving efficiency, compared with cooperating with marketing companies or purchasing commercial paintings or designs? Are AI-generated art services attractive to them?

Are there copyright and legal issues involved in applying AIGC to brand marketing? After all, brands would publish their AI-generated marketing content on the market, though that kind of content is not called “commercial paintings”. Will AI-generated advertising plans or designs bring potential legal risks?

In terms of AIGC players’ operating costs, you just said that it takes 2-3 months and about RMB 0.8m-1m to develop a dataset that consists of data structures and hardware. Do you mean acquiring source pictures incurs little cost? Could you elaborate on what data structures and hardware contain?

In the process of AIGC algorithm model optimisation, do AIGC companies choose to adjust the Stable Diffusion model, or to improve the code generation? Or do they only conduct data classification and training, instead of deeper model optimisation? Could you introduce how Baidu, Volcano Engine’s AI art generator, Printidea.art and Wujie AI conduct their own model optimisation? Do they adjust the Stable Diffusion model? Or do they give priority to support annotation for generating nice images with various styles via tools, other than making considerable optimisation of models?

It’s quite clear now, and I asked those because I am not very familiar with this field. Let’s take a look at commercialisation now. As an individual user, I tried several different AI art generators and input several keywords of different styles, but the content generated was quite unsatisfactory, some of which were pretty terrifying. Could you please share with us what are individual users’ evaluations of tier 1 AI art generation platforms? Do those platforms meet their needs? How high is the individual users’ willingness to pay for such platforms? What is the current commercialisation progress in the business-oriented market? Are there any business users willing to pay for AI art generators and how much will they pay?

Could you introduce the DAU of leading players in the AIGC industry? What is the ratio of paid users? What is the number of DAUs and paid users of leading players in the industry in 2022?