Experience of using artificial intelligence in the work of an artist

In the art community, debates have long been ongoing about what the growing use of neural network-based artificial intelligence (AI) will bring: the death of creativity and creative professions, or a great boon that eases the workload for everyone, both creative and non-creative.

The topic of AI is highly interesting to me, so I want to share my experience using this new tool, along with the results and conclusions I’ve drawn. Perhaps this might encourage someone to take a different view of the «terrible and soulless machine.»

The Birth of an Idea

For a long time, I had an old artwork that had caused me significant difficulties and dissatisfaction during its creation. Unsurprisingly, I wasn’t too happy with the final result either. The only thing I could say about it was, «Well, at least I found the strength to finish it after six months of attempts.» Most of the time I’d just open the file, stare at it, make a couple of brushstrokes, and then close it for weeks. In that sense, I was glad that this nightmare had finally reached a somewhat acceptable outcome.

However, months or even more than a year later, looking back at the piece no longer brought me satisfaction. It’s normal for me to revisit old but decent works and make necessary corrections. I thought it was time to do the same with this piece. Yet, I had no idea where to start. Just like during its initial creation, I felt overwhelmed by the sheer number of details I had envisioned and the mistakes I had accumulated along the way. Unfortunately, I didn’t have anyone around who could offer advice.

That’s when the idea of using AI for help came to me. For me this wasn’t a completely new thought. Before I had already used ChatGPT for studying foreign languages and resolving certain psychological challenges. Around the same time I came across a post on social media by an artist who had experimented with using AI in her work. She would set the base colors for a character sketch, and the AI would suggest rendering options. As her digital assistant, she used the AI tool from the mobile photo editing app Meitu.

The Process of Work

The Meitu app offers a vast array of photo editing features, but my goal was specifically to test its AI tool. Finding it was easy — the app developers made sure to promote the new feature prominently, placing its icon right at the top of the main menu. Once you upload the desired image, the AI processes it into five different styles. Four of these lean toward a relatively anime-like aesthetic, while the fifth is also somewhat anime-inspired but distinctly incorporates traditional Chinese elements, including cultural motifs and colors. (This is how this function worked back at May 2023. Now it has way more variety as well as restrictions.)

For obvious reasons, the «Chinese» style didn’t suit my needs. Two of the «anime» styles distorted the original image too much, leaving the remaining two as the most functional for my purposes. These two styles are the ones I predominantly use. After uploading my original image, I selected two variants for further consideration:

From a distance and with a slightly unfocused sight, the neural network’s work looked fantastic! But, as many people have probably already noticed, the longer and closer you examine the results, the more flaws you find. Let’s analyze the outcomes in more detail. What did the AI manage to identify in my painting?

• The character’s gender — female.
• The position of the hand(!) in one of the versions.
• Clothing elements — white shirt and crimson pants.
• Interior elements — a table, chair, framed picture, and even a shadowed door and a cable on the floor.
• Some environmental elements — a rapier, sheets of paper on the table, a cloak hanging on the back of the chair, and a bottle in one of the variations.

Now, let’s move on to the errors and flaws:

• Humanization: This didn’t surprise me, as this AI tool is primarily designed for processing photos of humans. As a result, unnatural skin tones and other non-human elements (ears and tail under the chair) were “eaten” by the AI, leaving only two light streaks above the head where the ears had been. At this point, nothing remained of my original character — just an abstract anime girl.
• Body Parts: In the second variation, the face was somewhat recognizable, albeit heavily stylized in an anime aesthetic. However, in the first variation both the face and visible hand were significantly altered. The AI merged the hand with the cloak hanging on the back of the chair, transforming it into a sort of shawl draped over the character’s shoulders.
• Clothing: While similar, the clothing in the AI-rendered images was still different from what I had originally envisioned.
• Environmental Elements: Recognition of the surroundings was inconsistent. For example, in the second variation, the AI successfully identified a bottle on the table, but in the first it turned into some sort of vintage lantern. About the futuristic blaster — it might partially be my fault, as I’m not yet skilled at drawing weapons. However, even if I had rendered it better, I doubt the AI would have recognized it as a futuristic blaster rather than, say, an antique revolver. Meanwhile, this blaster was a crucial part of the character’s story. Instead of the blaster the AI rendered a cloth or papers in its place. The same goes for the dagger and pen — key narrative details were lost.

Did the Neural Network Misinterpret My Character?

Yes, it did. Did it misinterpret specific elements of the environment? Absolutely. Did it make my drawing 100% better, to the point where I could confidently post it on social media? Partially. If I were looking for an abstract result or could convince myself that this was exactly what I wanted, I might have been satisfied. However, I wanted this to remain my specific character in their specific outfit and with specific details in the image. As such, I could only use these AI-generated images as references. But even that is no small thing!

Refining the Artwork

I decided to lighten the overall shadow, work more meticulously on the chair and the cloak draped over it, add highlights to the rapier, and simplify the shadows on the hair, shirt, sock and shoe. This brought me to the following result:

Next, I uploaded intermediate versions of the artwork into the app two more times. This was partly to refine some details and partly out of curiosity — I wanted to see how the neural network would interpret my work after the adjustments.

What Changed?

• The quality of the face worsened in both variations.
• The bottle appeared in the first variation but turned into a lamp in the second.
• A pen emerged on the table in the first variation.
• The textures of the walls, floor, and door became more detailed in both variations.
• The pants changed color — brown in the first variation, black in the second.

At this stage, I decided to enhance the door texture, reduce the hair volume while adding more highlights, and refine the folds on the fabric of the shirt, pants, and the shoe on the floor. This resulted in the following outcome:

Next Stage of AI Processing:

What Changed?

• The faces improved in both variations.
• The bottle transformed into a lamp in both variations.
• In the second variation the paper sheets and pen disappeared from the table. On the first the pen turned into another sheet or an envelope.
• The color of the pants returned much closer to the original in both variations.
• Reflections appeared on the floor in both variations.
• The shoe on the floor, which had remained a shoe until now, became an abstract color blotch in both variation.

Final Result

In the final stage, I made the last adjustments: refining the folds on the pants, reducing the size of the hand, adding a fallen piece of paper to the floor, incorporating more blue tones into the shadows, and intensifying the glow effect on all illuminated objects. These changes brought more «life» and «air» into the artwork. After all these corrections based on the AI’s «advice,» I decided to leave the piece in a state that satisfies me much more:

Conclusions

So, what conclusions did I draw after completing this work?

1. AI finishes your work instead of you only if you don’t know what you want.
   If you do know, it will still finish it — but not in the way you want. Unfortunately, this particular AI has no adjustable parameters. Other neural networks do have them, allowing for better control over the outcome. However, even those currently lack sufficient precision.
2. AI is a decent critic and assistant.
   You can turn to it for «advice,» especially when you lack access to a teacher or another competent individual (and don’t feel like listening to anonymous opinions from the internet). From this, another conclusion follows.
3. Artificial intelligence can be used as an auxiliary learning tool.
   The advantages of AI as a critic and learning instrument include:
   • Impartiality
   • Non-judgmental feedback
   • Emotionless corrections
   • Relative variability in corrections
   • Non-mandatory application of all suggestions

Final Thoughts

My experience has shown that an artist can effectively use artificial intelligence in their creative process while maintaining their vision and personal style. By treating AI purely as a tool and a support mechanism, it is possible to improve both the quality and productivity of one’s work.