Technical Imperfection in AI Media Art: How Generative AI Uncertainty Shapes Aesthetic Experience

Controlled Indeterminacy, Immersion, Playfulness, and Awe in AnimateDiff-Based Interactive Media Art

Jiyun Han and Yoomi Choi’s study brings distortion and instability—qualities usually treated as problems to be eliminated in generative AI—back into the artistic process. Rather than asking how accurately AI can generate images, the study begins with a different question: what happens to the viewer when AI becomes less accurate?

The authors are less concerned with the formal perfection of an artwork than with changes in perception and emotion during the viewing process. Their central concern is how low-level technical decisions such as parameter tuning can lead to higher-level experiences including immersion, curiosity, playfulness, and awe. To explore this, they combine artistic production with empirical audience research.

Designing the Conditions in Which Errors Can Occur

One of the key concepts of the study is controlled indeterminacy.

Technical imperfection here does not mean passively accepting whatever failed images an AI happens to produce and selecting interesting ones afterward. Instead, the artist deliberately changes parameters such as Sampling Step, CFG Scale, and Control Weight so that instability emerges within a chosen range.

The artist therefore does not directly determine every resulting image. Instead, the artist designs the range and conditions within which variation can occur.

The study also makes an important distinction between technical imperfection and visual indeterminacy. Technical imperfection refers to unexpected outputs caused by technological limitations, algorithmic conditions, unstable functions, or other aspects of the generative system. Visual indeterminacy refers more broadly to a condition in which the form or meaning of an image remains unstable and open to multiple interpretations.

Technical imperfection can therefore produce visual indeterminacy, but the two are not identical. Visual indeterminacy may also arise through deliberate abstraction, ambiguous composition, or interactive intervention.

This distinction subtly changes the position of the artist in AI art. Artistic intention does not require complete control over every result. Deciding where control should end and where machine-generated variation should begin can itself become an aesthetic decision.

The Parameters That Produce Better Images Are Not Necessarily Those That Produce More Interesting Experiences

The researchers created an interactive artwork titled Redrawing Classics using Stable Diffusion and AnimateDiff. Viewers select scenes from familiar classical animations, after which AI transforms them into continually varying visual sequences.

Choosing familiar animation is significant. Because viewers already have some understanding of the original imagery, they can devote more attention to the difference between the expected form and the form currently being generated.

When certain generative parameters are reduced, phenomena normally interpreted as declining image quality begin to appear.

Lower Sampling Step values produce rougher and less refined imagery. Lower CFG Scale values loosen the relationship between the prompt and the resulting image, increasing abstraction and unpredictability. When Control Weight becomes sufficiently low, similarity with the reference image decreases and forms begin to distort or collapse. Higher values stabilize the output, but they also reduce the intended degree of abstraction and uncertainty.

The desirable parameter setting in this context is therefore not necessarily the one that produces the cleanest image.

What matters is finding a range in which the viewer can still recognize traces of the original form while remaining unable to confidently predict what it will become next.

If randomness becomes too extreme, the viewer loses the reference point required for comparison. If the image becomes too stable, there is little reason to continue exploring it. The “control” in controlled indeterminacy exists somewhere between these two conditions.

What Held the Audience’s Attention Was Not Change Itself, but the Failure of Prediction

One of the most interesting findings comes from the distinction between continuous change and unpredictable change.

Viewers strongly recognized that the forms were continuously transforming. Yet simply perceiving continuous transformation did not show a significant relationship with emotional aesthetic experience. The same was true for the simple perception that the background changed randomly.

A different pattern appeared when viewers encountered forms and movements that developed in ways they had not expected.

Unexpected changes in form and movement showed significant positive correlations with general emotional experience (r=.339), playfulness (r=.438), and awe (r=.453).

This distinction reveals one of the study’s most important insights.

Change alone does not necessarily produce immersion. Exploration begins when the viewer forms an expectation about what will happen next and that expectation repeatedly fails by a manageable degree.

The viewer begins searching for the rules governing transformation rather than merely consuming a sequence of changing images.

Another condition appears alongside unpredictability.

The more viewers felt that changing backgrounds remained smoothly connected, the stronger the relationship with playfulness became (r=.483). Regression analysis also showed that smooth connections between background changes had the strongest explanatory power among the three variables examined for playfulness (R²=.233).

Yet faster transformation was not always better. Excessively rapid background transitions had a negative relationship with awe (β=-.282, p=.041).

Taken together, these results suggest that the aesthetics of indeterminacy cannot be reduced to increasing chaos.

Expectations need to be disrupted while the continuity of experience remains intact.

The viewer should be unable to fully predict the next image, but still capable of following the relationship between one moment and the next. Indeterminacy therefore has a rhythm.

Immersion Emerges as an Ongoing Act of Reading the Image

The emotional responses reported by viewers did not converge in a single direction.

Mystery accounted for 21.3% of responses and surprise for 20.5%, but confusion accounted for 17.2%, discomfort for 13.9%, and fear for 9.0%.

Despite these mixed emotions, levels of engagement were high. On a seven-point scale, concentration scored 6.32, the ability to ignore surrounding distractions scored 6.15, enjoyment scored 6.06, and curiosity scored 5.96.

The experience therefore did not depend exclusively on comfort or positive emotion. Visual exploration continued even when unfamiliarity, confusion, or discomfort entered the experience.

The interviews support this interpretation. Several participants reported that they could remain immersed in the changing visual elements even in the absence of a clear narrative. Others suggested that the contrast between the original classical animation and the AI-generated transformation caused them to reconsider the visual appeal of the original work.

Immersion in this study can therefore be understood less as disappearing completely into a fictional world and more as a temporal process of repeatedly asking:

What has changed now?

The authors similarly interpret immersion as an ongoing process of visual exploration and sustained attention. Intended imperfection encourages what they describe as a kind of perceptual defamiliarization, repeatedly requiring the viewer to look again.

What the Tables Reveal: Indeterminacy Requires Structure as Well as Intensity

The paper does not provide a single final table summarizing the entire study. Instead, Tables 20 through 23 effectively contain its clearest conclusion.

Table 20 shows that the simple presence of change or randomness has relatively little relationship with emotional experience. Unexpected transformation, however, is consistently related to emotional engagement. Smooth continuity between changing backgrounds is especially strongly associated with playfulness.

Tables 21 through 23 reinforce the same pattern through regression analysis. Unpredictable transformation positively affects general emotional experience, playfulness, and awe, while overly rapid background transitions can negatively affect awe.

The decisive issue is therefore not merely whether imperfection exists, but how it unfolds over time.

The viewer must be able to recognize transformation. The next result must remain difficult to predict. Yet the connection between successive moments must remain sufficiently legible to be followed.

Within this balance, technical imperfection moves from being an error to becoming a component of experience design.

The significance of the study lies in tracing a sequence that is often left implicit in discussions of AI art:

parameter settings → visual indeterminacy → perceptual exploration → curiosity, playfulness, and awe.

Rather than stopping at the observation that AI can produce strange or unexpected images, the study attempts to explain how those variations operate within the viewer’s experience.

This also points toward a broader shift in the role of artists working with generative AI. The object of artistic control may gradually move away from the production of a single finished image toward the range of possible images, the speed and continuity of their transformation, and the moments at which the viewer’s expectations fail.

The study is particularly useful because it captures this movement not only as a theoretical claim, but within an actual artwork and measurable audience responses.