Recently there has been a hot discussion about AI image generators in social media. The main reason is because of a painting called “Théâtre D’opéra Spatial” by Jason Allen, who won first place in the “digital arts/digitally-manipulated photography” category at the Colorado State Fair Fine Arts Competition. However, after the result emerged after the prize was awarded, Allen posted excitedly about his win on Midjourney’s Discord server on August 25. The post went viral on Twitter after a few days, and it aroused the anger of many artists because Allen uses an AI generator — Midjourney to create the image. Allen said that he claims that he submitted his work using the phrase “via Midjourney”, but it seems to have been omitted when his work entered the competition for judging. It’s unclear if the judges knew that the piece had been created by an AI art generator or if the judges did not know the generator’s name.

Anyway, it is not a matter for me, as if the AI generator becomes more accepted in the future, and I think it certainly will, I am more concerned about it will affect the ecosystem of current creative industries. When I browse through Facebook and found most of the industry’s practitioners have negative feelings towards the emergence of AI generators, I have a different view towards it. I think AI generators can be a good reference tool for creative practitioners. It can help solve many miscommunications at the beginning of the work. It is a common situation in the commercial world between artists and clients who both have vastly different understandings of various arts-related terms.

This news also inspires me to think about my research focus, the Hong Kong Actions. So, I started to do the experiment with Midjourney. Firstly, I have a question about how the AI generator understood the term ‘action movie’. The best way to answer this question is to try it directly inside the generator. When we use any AI generators, we need to know the process called ‘prompting’. Prompting means the word string you type in the AI generator to describe the arts you want it to generate. If you have experience with it, you will know that there is a learning curve to using it. You need to give a few tries to let the AI understand what you mean before you get your ideal image. I called it a conversation between you and the AI. Here is the first prompt I type in Midjourney:

/imagine prompt action movie –ar 16:9

It means you want the generator to give you an image of an action movie when a 16:9 aspect ratio. I requested the 16:9 view because I want it more looks more like a captured image from a movie. The following image is the result.

The result is fairly good, but the AI only ‘thinks’ the action movie is about an explosion in the city. For a better prompting string, it needs more keywords in a string, so I started to try with ‘people fighting’ and ‘action sequence’. It is much fairer for the AI to show its understanding.

As we can see, the AI understood more about what I wanted. The result is more like an image from an action movie. After a few more tries, the action sequence is a needless word to generate a similar image, so I dropped it and changed it to people fighting scene only. Then I started to query what is the difference if I add Hong Kong to the prompt.

The AI generator gives me a similar image but only add Hong Kong city scene in the background and change the costume of the characters. I continue to try with a more specific scene to expect a different result. I changed the fighting scene to two people fighting scene. Here are the results without and with Hong Kong in the beginning.

The two results seem quite different. The one without Hong Kong in front of the action movie seems extremely focused on showing two people fighting. However, the image prompt with Hong Kong seems not to show two people only. I have a few months of experience with Midjourney and I knew that it would have different results on some occasions. I tried a few more times on both results as shown in the below slides, first one is without Hong Kong.

We can easily spot one interesting difference in both cases. The action movies without Hong Kong are more emphasize the fighting between two people and the result of them. The prompt with Hong Kong action shows two people fighting in a crowd fight or a gang fight. Also, there are some occasions that the AI tried to create a closeup shot for the fight, but the background still has some gangs. I cannot get an image of only two people fighting.

Based on the resulting images, I start to raise the following questions:

1. Why is the word ‘Hong Kong’ has more weight than the weight of ‘two people’ for AI? (For the AI different words have different influences on the result, it is expressed as the weight in the balance.)
2. How can the AI relate ‘Hong Kong Action Movie’ to the closeup shot? Did the AI understand Hong Kong action movies have a shot size bias compared to a general understanding of action movies?
3. Are there any more attributes of Hong Kong action movies that I can analyze from AI? and how?

After this small test, I will certainly continue the investigation of AI and Hong Kong action movies. I hope I can have a chance to investigate not by image only. My research focus is the action-narrative. There are some other tests using AI image analysis like OpenPose that can help me understand how the machine understands human action and the idea of expression through human motion and action.

Please leave me comments or keyword suggestions for me to continue the test. Finally, I leave a few more interesting results with the prompt keywords for your reference. Hope you will like it.