Where AI and generative art is heading? My exploration and opinion on artificial intelligence showing creativity on art forms such as oil painting, drawings, illustrations and many more. Questions in my mind that keep me wondering on the subject, and this interest is growing day by day. All prompts are reproduceable in this article. This is Part 1 to a series of blog.

Humans are creatives but machine is labour intensive. We have the inbuilt ability to understand complex ideas, and express deep emotions into images, sculptures and various art forms. Machine learning gurus are making AI models to think, act like humans. You never know in near future, a paint artist in the future may be avialable as a web services or a digital agent who can draw any number of oil paintings.

Current LLM models have extremely paced up the natural langaugae interface to the artifical intellegince generative processes. Now, you can more easily desrcibe what you want from the AI model to perfom. Generative models for images are also in active research, We see some great work in blending images, art work, profile pictures and back ground removing. Telling AI generative model to produce something similar to mona lisa works? No, or you should try that if dont agree here (Hold on, we attempted in next paragraphs).

Prompt engineering is when you give a command to generative models such as chatGPT to do a certain task. Similarly, DALLE, is image generator, it will generate a picture from the model that had learned from thousands of pictures, and generalizing the components, elements to whatever the neural net weights have being assigned to layers. And ofcourse a billion of parameters are used in learning that model. We will try some prompts.

I performed a set of observation/experimentation using free credit by hotpot.ai (They are using stable diffusion). Overall, it was a good experience. Thanks to them. What I was looking for to see how my prompt are ending up in the outputs. I have a model to choose and a text prompt. The longer and meangingful the text, the better would be the result. Longer prompts also means you have to know what you are looking for. I kept adding one word specific here and there to get what I wanted. Choosing a model, you have a specific set of theme in your mind: I used 3 different prompts and 3 different models out of 18-20 variations that I created. I must be clear here, the intent is totally educational and find out what can be acheived.

Last thing: hotpot interface was simple and useful. I use all my credit in one go: After each prompt, the session loaded with all my previous attempts with edit options. I tried so many variation that I forget which one is which? You know what I mean. I use Edit to reload and check the composition of prompt and model I used. Awesome work!

Prompt 1: “Mona Lisa : hand made art versus AI generative made art comparison” with model “Photo Portrait 1”
AI generated Mona lisa art image
Mona lisa compare original versus ai generated one. Source Credits: hotpot.ai/art-generator How appealing in mona lisa?
Prompt 2: “Human made art versus Ai generative made art comparison” with model “Hotpot Art 10”
Generative AI is spoiling art images
Human art versus AI generative art: a comparison Source Credits: hotpot.ai/art-generator The model used has picked two images and kind of comparison since they are side by side. Both are AI generated: Could you distinguish which one ‘human art’?
Prompt 3: “hand made art versus Ai generative made art comparison” with model ‘Illustration Art 2’
Generative AI art images
hand made art versus Ai generative made art comparison Source Credits: hotpot.ai/art-generator

I come to think that how creativity is being used by such models. Is it possible, that models learn from bad art images and produce something good out of it? My experience over couple of weeks on the model led me to some thinking about human aspects of the A.I image generators. I list here:

Generative AI is spoiling art images

  1. AI generative image models fail to blend emotions Human think, choose colours to highlight emotions, whereas a perfect prompt will blend two images into one. Try a couple of images first please, I tried it. Nah, my artistic sense do not accept it. Blending emotions is difficult.
  2. AI generative image can not merge two concepts into one It can do merging, but AI modeler will pick up input data that produce the quality to acheive such a task. Yes, I agree, a lot of work is really impressive in terms of AI improvements, not impressive in art form.
  3. Can human experience be compared to generative ai model A lot of paintings are drawn to get the experience, until you get a master piece. While several compute hours and a budget in millions could give you a model. Humans learns from mistakes, so does the generative models. Is the generative model 'thinking' the same way as human? I can agree, generative model can extract 'patterns' in bulks of image faster than humans but using them creatively is still under question. A program is supposed to do labour intensive task fairly in less time. However, the experience brings in good art judgement and the minute details of association and blending elements in the artistic process is irreplaceable.

Machines sholud learn the creative process. Is there any ‘thinking’, involved?. I think there is more of direction effort needs to be put in. My selection of different prompts made me think still I have to figure out what I want to achieve: hit and trial. Image model has vast knowledge of images and I am the one who is unable to figure out, what to ask or how to ask, to get the right answer? This is what I felt. Ah, Is it that I would have to learn from A.I, how to even ask What I want to ask? Here are my thoughts and reasons:

  1. learn the process of image creation Huge data describing the natural language, how elements, shadows, postures, face lift, light, and prespective are making an impact in over all goal of the artist. Get huge set of examples for the expressions.
  2. learn more for emotions Anger, Laughing, sadness and smile how to add in the complex images which tells a story. When such prompts are given, the model should explain it as well what element it has picked to bring the user request 'prompt'. See prompt 4 below.
  3. AI Tools for niche task can be helpful Making a big Sentiment tagger for all languages, and fine tuning to a specific may not give that good results. Specific task based artificial intelligence tool could be helpful in assisting humans in their work.
Prompt 4: “long shot image where I can see bride smile at the funeral with raining at the background but crying groom” with model ‘Illustration General 2’
Generative AI art images
long shot and raining was well ‘understood’ while smiling at funeral is forced at all characters. The situation in itself a little tricky. Source Credits: hotpot.ai/art-generator

A.I is disrupting the scene world wide. I am in Pakistan, and yes, following it is not as easy as it should be. Hotpot.ai came in handy. It’s not that I did not tried any other resources, the interface make sense and simplest was rather cool for me. I am going to further write on my explorations on chatting and image generators.

Finally, I myself oggling at the A.I scene, expecting more amazing things to happen. Really it is hard to catch up with trends and fads. If you like what you read, please sign up BilaUnwan Newsletter on weekends.