If Game Art is Advertising; GenAIArt is Market research

​​

​

​Game Art as Advertising 

The idea that art exists in a game to attract customers

This involves things that might seem obvious, like preproduction paperwork that helps the Art Director explore or validate their ideas as market feasible, or late production marketing materials. It also involves things that can be a little less obvious, like steering Art in a direction towards interesting Screenshots or making sure the HUD works well with Stream overlays, so that news outlets and streamers can do the advertising for you.

Why GenAIArt struggles: breaking the Sparkle Stick

A phrase that goes around gaming development is "hitting it with a Sparkle Stick", which describes the part of the project where the gameplay bones of it have settled and it's time to pour on the glitzy glamour. 

Hitting something with the Sparkle Stick has a cost associated with it. I mean this in the Performance and Financial senses, it's indeed expensive. But I also mean it in an opportunity cost sense, in that pacing sparkle throughout the gameplay experience in planned high-points and low-points can achieve larger artistic communication goals. (Ex: If we want the player to feel stronger over the course of the game, then their attacks can progressively get more sparkle.) Current GenAIArt is basically: "See this Noise? Remove the Noise from it. And while you're at it increase the probability it will end up looking like the common consensus of all internet pictures tagged 'Fox'". 

If you turn that probability high, then your results are all going to look the same. If you turn that probability low, you may get something that doesn't look like a Fox at all. There's a real tension between results being 'Monotonous but Coherent' and 'Varied but Nonsense'.  (This is why people say GenAI isn't capable of Creativity, Creativity is 'Varied and Coherent'.)​​

​Another way of look at that is: In GenAIArt something has to be commonplace in order to exist at all.

If you believe Game Art exists to attract customers, and you want to use GenAIArt that produces commonplace results, you end up in the position of having to make commonplace standout - especially as others using GenAIArt reinforces what commonplace is.

Sparkle Sticks sure do help things standout, but sticks are made out of wood, and it's only going to be hitting so many things before it breaks. Even if you had all the money/time/processing power in the world, you're still in between a rock and a hard place when it comes to larger artistic communication goals: either 'low-points are low' or 'start at 11 and have nowhere to go'.

​

"Discard the Bias, it's too Boring"

​​Instead it's better to recontextualize GenAIArt as the world's most accessible large-scale market research into Boring. (I do mean this as a good thing). During Production it's common to have Design Pillars, several short phrases that encapsulate what you want the resulting game to be like, so all the decisions people make in the day-to-day of their jobs can keep the greater goal in mind. If it's a foregone conclusion any result GenAIArt gives is too commonplace to advertise easily, then that means each Artist on the team has access to what an undesirable result looks like for any given assignment. Like a Design Pillar but in reverse, instead of encouraging people to point in the right direction, it discourages pointing in the wrong direction.  This is similar to "Discard your first idea, it's too obvious", a Rule of Thumb in Advertising to get around the Einstellung Effect and other cognitive traps. 

Use the same purposefully vague prompt across a variety of non-styled models, leveraging your human superpower of being able to pick out a pattern from limited examples to find similarities between the results. If you notice a similarity that isn't one of the words you included in the prompt, that's a bias worth examining or avoiding.

Example: "Create a screenshot for a videogame. In this videogame the character the player controls is a cunning fox, that is low on health. The fox is jumping over a fence obstacle. There is a single plum power-up that isn’t in the center of the image. There is a single intimidating Coyote boss enemy attacking the Fox. The background is a cornfield in autumn."

Notice how few images it took you to start seeing some underlying similarities.

Inferring a pattern from a few examples is something us Humans are naturally great at, because unlike GenAI we consider things like causality and applying lessons from other areas of life. (Ex: A Human wonders what effect the 'plum power-up' causes and applies knowledge from other experiences to figure it out. That's how the upper left plum can seem like a weapon, and the lower right plum can seem like health. In truth, GenAIArt is just putting the plum where icons probabilistically go. Which is why the 'fighting game' middle bottom plum is nonsensically drifting around where the combo counter would be.) Similarities:

• Lots of Brown  (despite not mentioning color)

• Cartoon style  (despite not mentioning style)

• 2d Side Scroller  (despite not mentioning game-type)

• Wooden Fence parallel to screen (despite not mentioning material or orientation)

• Fox on the left, Coyote on the right (despite not mentioning position)

• Afternoon (despite not mentioning time of day)

• Mountains in distance (despite not mentioning mountains)

Real World Example #1

If you were an Artist in charge of the Fence Asset, you'd be able to walk away from this with some idea that purely wooden fences that are colored brown might be so obvious it's boring. A quick google search may show that, in reality, barbed wire fences are often used to protect crops from animals. But that's getting into the next recommendation...

"Knowing what doesn't work, is the first step to knowing what does"

Early on in a project having some idea of what not to do can really narrow down the option paralyzing possibility space. This is better than using GenAIArt to give you ideas directly, not only for the Sparkle Stick reason above, but also because using it this way is futureproof.

GenAIArt has a problem because it can accidentally learn from AI Generated Art instead of Human Art. This is math-wise, a bad move, and leads to ever more boring results.​​​​​

Whether this eventually gets solved or not, doing 'Not Boring' is a pretty safe process to implement, because it'll be just as useful today as it will be 3 years from now. 

Notice how poorly GenAIArt handled "a single plum power-up that isn't in the center of the image".

When a Human sees that sentence with a Negative, we automatically think of all the other locations the plum could be (ex: upper left, right middle, etc.)

GenAI doesn't have a way to cycle through the other options. To GenAI, a Negative just negates or reduces another word, that doesn't tell it what the word becomes after it's been reduced or negated. To be fair, Negatives are linguistically complicated, as it isn't always as simple as the opposite. (ex: when we Humans say 'Not Hot' we don't mean 'Cold', we mean 'Room Temperature'. This is why some models offer Negative Prompts, which helps them avoid results with 'Cold'...but that doesn't necessarily lead to results with 'Room Temperature'.) For us Humans, knowing the other options means a Negative is an opportunity.

Flip the similarities into negatives, leveraging your human superpower of rapidly cycling through other options when presented with a negative to find ideas to explore.

• Not Brown: Blue, Yellow, Orange, Red.

• Not Cartoon style: Storybook Style, Illustrative Style, Craft Style (Clay, Toys)

• Not 2D Side Scroller: 3D, 2.5D, First Person, Open-World.

• Not Wooden Fence, Not parallel to screen: Wire Fence, Metal Sheet Fence, Perpendicular.

• Fox Not on the left, Coyote Not on the right: Coyote behind, Coyote head attacks from field

• Not Afternoon: Morning, Evening, Midnight.

• Not Mountains in distance: Barn in distance, Parallax Rows of Corn in distance, Rolling Hills in distance.

Real World Example #2

If you were an Artist in charge of the Fence Asset, you might note a Barbed Wire Fence reads as more dangerous and is better suited to being intuitively understood to be an obstacle. You also might note if the Fence Assets were oriented Perpendicular to the screen, it might be easier to understand they are meant to be avoided by Jumping.

This has the most benefit as an early process. It's probably best formally added to the Art Bible as a part of Market Research, and then rolled out as a process with standardized instructions of what prompt/s and services Artists should use to periodically check.

​References:

GenAIArt Images all throughout these posts are from: Leonardo.ai, Adobe Firefly, Midjourney, ChatGPT, Deepseek, Gemini, Nightcafe, Kling, and various open-source models from ComfyUI.

Biology of the eyes example: https://en.wikipedia.org/wiki/Chromostereopsis Neurology of visual processing example: https://en.wikipedia.org/wiki/Checker_shadow_illusion Psychology of how we group information example: https://en.wikipedia.org/wiki/Gestalt_psychology

Einstellung effect: https://pubmed.ncbi.nlm.nih.gov/18565505/#:~:text=The%20Einstellung%20(set)%20effect%20occurs,within%20their%20field%20of%20expertise. Casual Reasoning:

https://en.wikipedia.org/wiki/Causal_reasoning Domain General Learning: https://en.wikipedia.org/wiki/Domain-general_learning Model Collapse: https://www.nature.com/articles/s41586-024-07566-y Doggo image: ​https://techcrunch.com/2024/07/24/model-collapse-scientists-warn-against-letting-ai-eat-its-own-tail/