In this article, we’ll experiment with the idea of using a generative model as a modern paintbrush. An abstract piece of ‘artwork’ from scratch, relying entirely on a chat window – is it possible without an expensive model? Scroll down to see the results and let me know your thoughts. Purposefully ran this entire experiment using the free tier of Gemini. Wanted to demonstrate exactly what the most accessible baseline tool can achieve.
Generative models act as complex black boxes. They run on probability, vast neural weights, and a fair amount of digital chaos. So, even if you copy my prompts word for word. You certainly won’t get the exact pixel output. Thus, I won’t share all the prompts and experiments here, but I’ll give you an adequate idea to experiment with. The iterative technique worked for me on my various attempts. It reliably steers that chaotic output relatively close to the original creative vision.
The Concept: The Chaos of Human Focus #
Let’s explore the bizarre nature of human focus. To make it more random for the AI, let’s consider the daily Muslim prayer, for example. Since I’m a Muslim, praying five times a day for roughly ten to twenty minutes is a daily part of life. It serves as a dedicated pocket of peace. Standing in prayer, intending to talk to the Creator (we Muslims call our God ‘Allah’), if it’s done properly, it clears the burden from the mind. But sometimes, the brain is already too restless; it immediately rebels. It decides this exact moment is perfect for a random system check. You suddenly remember you need milk or groceries for your kid. You recall exactly where you left that missing keychain. Or, if you are a programmer, you might subconsciously start debugging a stubborn race condition in your latest codebase. You do not even realize your mind was wandering until you hear the congregational prayer is ending already! Even in this short span of time, concentration quickly evaporates if you’re not focused. We should visually capture that tiny window of slipping time.
Let’s scale this idea up further by looking beyond for a moment. Life itself heavily mirrors those short moments in prayer. We only get a finite span of time on this earth. Yet we constantly wander off-topic. We let trivial daily frictions dictate our entire existence. Someone bumps your car in traffic. They yell at you. They wildly insult your family. The actual event lasts two minutes. You let that anger hijack your entire day. You mentally replay the argument while eating dinner over and over, ignoring that same family! You miss the opportunity to spend a quality moment entirely. The altercation ended hours ago, yet you let it continue stealing your joy.
We waste our actual lives obsessing over irrelevant noise. Our generative artwork needs to capture this exact tension as well. It must firmly contrast the quiet necessity of focus against the loud chaos of distraction.
Sharpen the concept #
Now that we have the big picture, let’s think about the composition and paint a mental picture first. We bring the traditional Arabic right-to-left composition to make it more relevant and thematic. We start from the far right side, expressing tension and chaos. We will also use Arabic calligraphic-style borders to keep the chaos contained without going berserk. Then, at the bottom right side, we represent the entering moment of prayer with a symbolic mosque entrance.
The random thoughts or life struggles can be represented with waves in a sea of thoughts, and the wandering, confused mind can be represented using clouds. Clouds can also represent gloomy sadness or calmness in some cases, so we’ll heighten the chaos by adding swirling vortices, giving a stormy feel to the space hovering over the mosque entrance.
Since the sea of thoughts and cloudy minds are connected, we put two prominent connected dots linking them together.
Next, let’s compose an hourglass, representing the short prayer time, from which the clouds are draining some sand out. We can always improvise later if we want to.
For the last part, we will represent the life cycle with various phases of the Moon on the left side. It should be symmetric. Life itself has a symmetry in it; when you are an infant, you depend on your parents, and when you are extremely old, you again need to rely on someone to care for you.
The challenge #
Now, if we give the idea ‘as is’ to Gemini, ChatGPT, Grok, or Claude, and ask them to one-shot it, the output may not be what we expect it to look like. Each one of them will go crazy with it; even if you use the same prompt over and over, the output will be random and different each time. Moreover, the art style won’t be consistent; in many cases, I’ve seen parts of the art use a different art style and combine them in a way that makes no sense from an artist’s perspective. I will add some generated images at the bottom of this article with one-shot attempts; just notice how they are composed, and you’ll see what I mean.
How do we make it consistent? #
You could obviously draw a few foundational lines on paper/ipad and feed it as a reference image to the model and let it improvise on it.
What I found working for myself is keeping things minimal till the full composition. I start with a few stroke line arts, breaking the concept into small, digestible micro compositions. This works as a foundation for the LLM to constrain its thoughts around a targeted goal.
Since, for this experiment, we want to keep everything confined within the chat window, we will start by doing some basic line drawing. See the examples below:
A minimalist line art of a mosque entrance with as few strokes as possible. Only a few strokes to represent the archway and a few more strokes at the bottom to resemble some stairs. Put some cloud vortices above it
Minimalist line art of a Sea.There are four surfer waves raised above the water level. Following one another, going from right to left. Use as few strokes as possible.
Let’s merge these two into a single composition
Imagine a rectangular frame containing the abstract art, The attached sea is at the bottom line of the frame, while the attached archway is at the bottom right corner. Expand the clouds to fill the area above, keeping empty spaces on the left side.
By repeating the above process, we can generate the rest of the items, like the hourglass and others; we can combine all the micro compositions, stacking one on another. If the AI starts hallucinating, retry the same steps, tweaking a few words.
After some iterations, the composition already started to look more meaningful
By sticking to minimal lines and targeted spatial instructions, we limit the moving parts, making it much easier for the model to adhere to the original vision.
Next, we can ask the model to draw prominent lines for the base shape and some supporting textured lines and add more detail to the composition, which ends up as below:
Now, Let’s add the colors that I imagined with a separate prompt. In this prompt, I specifically limited Gemini to use the two primary colors and stick with those. The result is as below:
To me, it looks great, pretty close to what I actually imagined! It is purposefully chaotic; notice how your eyes are getting distracted by several elements at once, exactly how we are distracted if we are not focused. But if you start looking from the mosque entrance, you pass through the swirls and waves, and your eyes settle on the two prominent dots. You momentarily regain your consciousness from the distractions, but just like the vortices of cloud pulled the sands out of the hourglass, you spent quite a bit of your time on those distractions.
Now look at the shape of the sands that have already fallen inside the hourglass; notice how they match the symmetrical shapes of the moon phases, indicating these tiny moments that slip into distractions are also part of your life.
There is still room for improvement; for example, the prominent two dots that I originally planned aren’t placed accurately yet. But overall, this is great!
After generating the artwork, I also tried experimenting with some fun stuff, like how it will look if we turn it into a rug or carpet, for example:
Now let’s look at some of the one-shot attempts that I made with various AI models.
One-shot attempt of Gemini #
Simply described the concept and then asked Gemini, ” Can you create an image?”
Seems like Gemini focused on the storm too much, making it too violent. The sky is torn apart in visible blood, while the mosque is engulfed in both the sea and clouds.
ChatGPT’s attempt at one-shot #
Similarly, I’ve given the exact same prompt as Gemini to ChatGPT; the output is wildly different:
Notice, ChatGPT’s output is more realistic than Gemini’s by default without adding any additional keywords.
Claude’s attempt at one-shot #
This is done on the default claude.ai chat with default settings, i.e Sonnet 5.0, etc with the exact same prompt as the other models.
This reminds me of Microsoft’s og MS Paint for some reason.
Grok’s Attempt at One-Shot #
Grok went one step further and attempted to label each moon phase in Arabic.
What if we mention Abstract art instead of an image? #
The output changes drastically again. For example, Gemini spits out the image:
Even though we didn’t explicitly ask any of these models to include Arabic text, they’re adding Arabic words as calligraphy. Gemini is still overdoing the storm and sea.
What if we use advanced prompting techniques? #
Yes, there are ways to get something more meaningful if you append some keywords and stylistic preferences to your prompt; for example, LLMs behave a little bit better if they are told to mimic famous art styles, like if you say draw like van gogh/picasso, etc.
Right-to-left abstract composition, contained within Arabic-decorated borders, the bottom border contains a sea with four softer-edged surfer waves, bottom-right mosque entrance archway with stairs, stormy cloud vortices filling up the above, two prominent connected dots linking clouds and sea, hourglass with sand draining into clouds placed following the rule of thirds at the middle left side of the composition, far left symmetrical round shapes representing the various moon phases, conceptual abstract art, 2D.
Right-to-left abstract composition of Picasso, contained within Arabic-decorated borders, the bottom border contains a sea with four softer-edged surfer waves, bottom-right mosque entrance archway with stairs, stormy cloud vortices filling up the above, two prominent connected dots linking clouds and sea, hourglass with sand draining into clouds placed following the rule of thirds at the middle left side of the composition, far left symmetrical round shapes representing the various moon phases, conceptual abstract art, 2D.
Right-to-left abstract composition of Van Gogh, contained within Arabic-decorated borders, the bottom border contains a sea with four softer-edged surfer waves, bottom-right mosque entrance archway with stairs, stormy cloud vortices filling up the above, two prominent connected dots linking clouds and sea, hourglass with sand draining into clouds placed following the rule of thirds at the middle left side of the composition, far left symmetrical round shapes representing the various moon phases, conceptual abstract art, 2D.
Many factors instantly give away that these are all AI slop. Even for an untrained eye, these look absurd or unusual.
Some more advanced one-shot attempts with detailed stylistic references #
We can append some keywords to optimize the output and mimic artistic styles a bit more closely; for example, see below:
Right-to-left abstract composition, contained within Arabic-decorated borders, the bottom border contains a sea with four softer-edged surfer waves, bottom-right mosque entrance archway with stairs, stormy cloud vortices filling up the above, two prominent connected dots linking clouds and sea, hourglass with sand draining into clouds placed following the rule of thirds at the middle left side of the composition, far left symmetrical round shapes representing the various moon phases, Ukiyo-e, traditional Japanese woodblock print, flat bold colors, prominent black outlines, dynamic composition, Edo period aesthetic
Ukiyo-e is a genre of Japanese art featuring woodblock prints and paintings that flourished from the 17th through 19th centuries. The generated image has some vibe of the art style, but the mosque is unusually highly detailed and doesn’t blend with the actual composition. Also, for some reason, the sand appears to be leaking downwards.
Pablo Picasso, Cubism, fragmented reality, geometric shapes, multiple perspectives, bold abstract lines, analytical cubism, asymmetric, Right-to-left abstract composition, contained within Arabic-decorated borders, the bottom border contains a sea with four softer-edged surfer waves, bottom-right mosque entrance archway with stairs, stormy cloud vortices filling up the above, two prominent connected dots linking clouds and sea, hourglass with sand draining into clouds placed following the rule of thirds at the middle left side of the composition, far left symmetrical round shapes representing the various moon phases2d abstract art
Like the way it connected the two prominent dots; however, the calligraphic border with Arabic text and the moon phases don’t blend well. Nevertheless, these are a bit better.
Closing thoughts #
We are living in the AI age. I still remember when we giggled at Prisma turning awkward selfies into digital Van Goghs or at the viral hype around early Midjourney. Today, the landscape looks entirely different. We now routinely command superior models to build photorealistic worlds and render flawless typography in seconds.
For years, digital artists emptied their wallets for high-end tablets just to run Procreate, or memorized obscure keyboard shortcuts to perform vector magic. Some fiercely debated raster versus vector graphics. Now, modern multimodal LLMs casually perform tasks once considered unimaginable just half a decade ago.
Carefully craft a descriptive sentence, and your computer ‘hallucinates’ a masterpiece.
Off-topic: If you are a complete beginner to prompt engineering or agentic coding,
you might find the following books interesting
(skip if you are already an expert)