‘10,000 Bowls of Oatmeal Problem’ is when a generator fails to produce visually unique outputs to a human eye, even if each output is unique in its computation. It draws its metaphors from bowls of oatmeals, as looking at 10,000 different versions of oatmeals does not intrigue the viewer to look at the granular details of each bowl of oatmeal.
I think this scenario can become a problem when you know that each output needs to be distinct from every other output. For example, if a generator produces game characters in a multiplayer game, each one has to be perceptually different from each other. Without the difference, players will become confused and frustrated, perhaps killing a player on the team by accident.
On the other hand, this problem can actually benefit scenarios when you want an output that changes over time. Procedural landscapes are a good example. When watching a procedurally generated landscape in continuum, you don’t want the user to be startled by starkly different images produced every 2 seconds. Instead, you want enough similarities between each image to maintain the identity of the place you produce.
Artistically, this problem can be solved by identifying the exact attribute you want to be different in each output. After narrowing down your goal, then you can go into the codebase to tweak the problem at hand technically.
Compton gives good advice to making generators: specific and concrete “artifacts”. Get expertise from someone who makes or knows those artifacts to form a list of characteristics to generate. Create a structure of rules that govern the limits of the generator’s variability and creation guidelines. The usage of Perlin noise and other randomness can be used to introduce parametric change with user input.
It is hard to answer Compton’s question of how to “debug” a generator which isn’t producing the creatures as predicted . To me, it requires technical knowledge of coding and a well made architecture so that it is easy to locate sources of variation in the generator. However, perceptual uniqueness is even harder to identify. This isn’t computer defined and is purely determined by the artist eyeballing the work to see if a human can feel interested in the different productions. For me, a periodic break so I can come back and look at the generated work with a fresh set of eyes helps me determine the perceptual differences in the artifacts produced.
Compton’s “10,000 Bowls of Oatmeal Problem” becomes important when it comes to the situation where multiples of individual items are rendered and used. As Compton explained in the article, every oatmeal piece in a bowl is unique and different from each other in terms of size, weight, location, orientation, etc. In other words, there cannot be two or more oatmeal pieces that are exactly identical when the oatmeal was poured into a bowl straight out of the package (randomness). Because there doesn’t exist two or more oatmeal pieces molded to be exactly the same intentionally(not random), every piece is independant and different, mathematically speaking.
This would be a problem, for example, when your intention of randomly generating mountains is to create a background of a game you’re developing and wanted to create a user experience of playing in a green mountain map. If the randomly generated mountains turn out to have different color, hue, saturation, shape that are too far away from the general image of mountains, then such intention wouldn’t be fulfilled by the random generation. This could potentially be solved by indicating certain features of the random generation before generating. To go back to the mountain example, we could limit the color spectrum to green and shape to more triangular looking. With such indication of certain features, we could then achieve what we intended to get out of the generator.
On the other hand, randomness and the uniqueness of the random generation would be beneficial when you intend to create a crowd of people, for instance. Often times a crowd of people in a film, animation, or game isn’t the main spotlight the developers want the users to focus on. The crowd of people are meant to create an impression that a lot of people/characters are there at the scene. Just like how it’s not common in real life to observe two or more people who look identical including the outfits, hairstyles, and body shapes at a crowded cafe, for example, it’d be funny if the scene you intend to create has a crowd made out of multiplication of a few the same exact looking characters.
Cate Kompton describes a problem surrounding the uniqueness of generative works and how uniqueness can be seen by an audience. A computer does not understand what we think makes something unique as from an algorithmic point of view, everything it makes is unique (she uses the example of 10,00 bowls of oatmeal that are all the same to a viewer but algorithmically different and unique to the computer). She explains that there are two main to resolve this issue, or at least help alleviate the problems it might cause. The first way is by use of perceptual differentiation, where we notice the uniqueness of something simply because it is imposed us as viewers and is large enough that it can not be ignored (like a forest). The other method is known as perceptual uniqueness, where someone will notice something not because it is imposing but because it looks slightly different than everything but not in a remarkable or alarming way. By giving your work strong personality and original features you can help prevent this problem.
Kate Compton’s 10,000 Bowls of Oatmeal Problem describes the potential for a generative algorithm to produce content or artifacts that are too similar for humans. The resulting mundaneness makes any output of a generator look too bland (ie. like oatmeal). This may be a good thing if you need to generate something like trees; each should be slightly unique, but 10,000 trees together becomes a forrest. If a given tree stands out too much, the forrest will not look right.
However, the problem of perceptual uniqueness arises when an algorithm’s output lack’s diversity. More precisely, consider a car generator. Cars are deeply personal items, with drivers often decorating them with bumper stickers, dashboard dolls, etc. It is difficult to explain why, but a human can tell if they’ve seen the same car drive by twice.
Like Kate Compton writes, considering dealbreakers of artifacts at scale is one way to try to combat a lack of perceptual uniqueness. Another strategy is to categorize artifacts into types/tiles (ie. SUVs, German, electric, diesel, sports, etc.) From a more artistic point of view, adding certain defects (ie. dust, rust, scratches, peeling paint, etc) to an artifact may make it more perceptually unique.
Compton’s “10,000 Bowls of Oatmeal Problem” talks about an aesthetic challenge that is often seen in creating generative content. While an algorithm might be able to generate different results, they may not always be perceived as different. Based on the purpose of the generative content, uniqueness has different levels of importance in determining the effectiveness of the generator.
“Perceptual differentiation” and “perceptual uniqueness” are the two criteria that Compton used to evaluate the generative results. “Perceptual differentiation” only requires the results to be not identical. In cases that need a large group of generative results without the need of highlighting specific characteristics, like a crowd of people or a school of fish to fill as background, differentiation is enough. While for cases that require distinct differences, like generating game characters, uniqueness should be taken into consideration while setting up the generative process.
To overcome the “10,000 Bowls of Oatmeal Problem”, we need to consider the purpose of the generator, whether the results should be seen as unique or simply identical, and adjust the constraints of the generative process based on the requirements.
Generative art has advantages and limitations – the advantages are that you can create consistent content, the limitations are that you create consistent content.
This might be a problem if you were trying to generate images of random but different themed scenery/backgrounds for a game (e.g. a cave, an underwater location, a fiery location). It might be difficult to generate images with specific themes and details. Generative art might make it difficult to create content that is stylistically unique.
A scenario in which this wouldn’t be a problem is if you were creating scenery or backgrounds for a generic location in a game. In this case, the generated images don’t need to have anything inherently interesting or special about them, they just need to not be distracting. (You could also create generic images to show on your computer lock screen. This also reminds me of the screensavers computers used to have.)
Some strategies for overcoming this problem would be to make additional changes to the content to make it unique, either with artistic decisions by hand or by manipulating each generated image with a different, unique algorithm.
The “10,000 Bowls of Oatmeal Problem” is a common problem which artists face when creating generative artwork, where the idea of perpetual uniqueness may only be correct in the micro sense but not the macro sense. You may have 10,000 “different bowls” of oatmeal, where oatmeal in the bowls are composed differently, but may be too similar to one another to differentiate in the bigger picture.
In some cases, the lack of perpetual uniqueness may not be a problem when an artists goal is to create a more homogenous set of objects defined by specific characteristics, specifically when the objects are not of significance. An example would be creating a crowd of people in a stadium, where the features of each person do not need to be clear and can thus appear somewhat similar from one another.
In other cases, perpetual uniqueness becomes an issue when a set of important objects are all distinct, but lack an interesting appearance. Though uniqueness is present in every character, the traits and characteristics fall short of being exciting or interesting due to the lack of memorable traits and symbols that give each character the unique personality. Successful generative artwork would have to depend on the artist’s creative decisions that will make every iteration memorable through a defining personality.
Compton’s “10,000 Bowls of Oatmeal Problem” illustrates the possible problems a creator can run into while utilizing generativity into their work. The problems can be faced is when the creator constructs so many artifacts making each of them unique in their own way but it can become difficult to be perceived the same way by the user/audience.
We can look at this idea as we look at the construction of violins. Each crafted violin may produce a different or unique sound. However, on the outside, they are all designed the same way and are utilized the same way. Here comes Compton’s concern where the unique qualities of a generated object will remain unknown without acknowledging its good/bad attributes. In this case, depending on the user, an individual violin can maybe produce more richer, deeper tones than the violin next to it. Whether that is a good/bad attribute will be determined by both the user and the maker but either way it will help identify perceptual differences within the clutter of instruments.
In order for the artist to overcome this and generate a copious amount of artifacts that truly vary from each other, they must be able to recognize the different types of perceptual differentiation that can be experienced by the user and enhance them until the users find it recognizable.
In Kate Compton’s reading, the “10,000 Bowls of Oatmeal Problem” describes an issue in generative work when an algorithm can produce a ginormous amount of artifacts that are each unique but are not perceived as uniquely different from the audience perspective. Compton provides an example that if she created 10,000 bowls of oatmeal and each grain of oat was different, according to the algorithm, it is unique, but perceived from an outside view, it all ends up looking the same.
When you need to generate thousands of artifacts that will have to vary slightly, you can choose between perceptual uniqueness or perceptual differentiation. Perceptual differentiation is an easier level to succeed at if the environment doesn’t need the artifacts to be highly memorable (i.e. trees in a landscape or a very large crowd.) Perceptual differentiation is when the user can tell from a glance that there is a difference between the artifacts, but it isn’t that significant. On the other hand, perceptual uniqueness is the contrasting view between remembering a main character versus remembering a face in the crowd. The artifact must have a distinct character personality, making it more memorable than the other artifacts.
To overcome this problem, it is important to understand your audience and know the key characteristics of the artifact you are generating. Humans like readable meanings and identifiable personalities, and is a great strategy to start with overcoming this problem.