Glitch as Intention:

It challenges the rigid ideals of perfection that define contemporary computer vision systems and, instead, embraces the unpredictability and transience that lie at the heart of both human creativity and machine learning.

Glitch aesthetics usually focus on human artistic interpretation. This work, however, demonstrates how they are interpreted by AI models. This could explore the tension between human aesthetics and machine failure, which is rarely addressed in traditional glitch art.

There has been considerable discussion and numerous artistic explorations centered around the concept of wabi-sabi in digital art too. However, a critical gap remains: the lack of a structured approach to understanding how different types of computer vision glitches influence the final aesthetic of AI-generated artwork. Existing works often explore these distortions in an abstract or isolated manner, without systematically categorizing and analyzing their distinct visual impacts.

Additionally, existing works focus on two approaches: (1) prompting models to generate wabi-sabi-inspired aesthetics or (2) fine-tuning models to produce polished, curated imperfections. However, these approaches often overlook the core philosophy of wabi-sabi—embracing the incomplete, the fleeting, and the beautifully flawed.

• Analog distortions from hardware malfunctions
• Pixel scrambling
• Digital dithering
• Data moshing
• Video frame skips/tearing
• Data compression artifacts
• Color banding
• Digital channel shifts
• Noise and grain (analog)
• Over-exposure
• Signal noise (snow/static)
• Moiré patterns
• Combination of all the above

To illustrate the impact of each glitch type, we present a curated AI-generated artwork corresponding to each category. This structured approach not only provides a novel perspective on the interplay between computational errors and digital aesthetics but also establishes a foundation for future explorations in glitch-inspired AI artistry.

Wabi-Sabi is an aesthetic of incompleteness—one that celebrates irregularities, asymmetries, and the quiet dignity of decay. In the same way that a weathered teacup, its cracks traced with golden Kintsugi (金継ぎ), holds more depth than an unblemished vessel, AI glitch art invites us to appreciate the fractured and distorted. Pixelations, compression artifacts, and latent space errors are not failures; they are the fingerprints of the machine’s process, revealing an unfiltered glimpse into the imperfections of computational creativity. These artifacts, once dismissed as mere noise, become echoes of the unseen, urging us to find beauty in the unintended.

Wabi-Sabi also speaks to the impermanence of all things. Wood weathers, ink fades, and stone erodes—such is the nature of the world. Yet in the digital domain, where data seems immutable, glitches remind us that even information is subject to transformation and decay. Corrupt files, broken algorithms, and unintended distortions are the digital equivalent of time-worn temples and crumbling parchment. This project amplifies these ephemeral qualities, allowing the machine’s imperfections to surface, evolve, and ultimately dissolve into abstraction. The resulting visuals become a study of time, loss, and the fragile nature of digital existence.

Just as a Zen garden’s asymmetry is shaped by the randomness of falling leaves and shifting sands, AI glitch art emerges from the machine’s own biases, limitations, and emergent behaviors. The artist does not impose absolute control but rather engages in a dialogue with the algorithm, allowing its unpredictability to shape the outcome. This collaboration mirrors the organic spontaneity found in nature, where chaos and structure coexist in delicate harmony. By leaning into the unintended, we unlock a new mode of creation—one that acknowledges the machine not as a perfect tool, but as a partner in serendipitous discovery.

Wabi-Sabi values subtlety and restraint, where what is left unsaid is as important as what is spoken. In AI glitch art, missing pixels, fragmented forms, and distorted edges become poetic absences, inviting interpretation rather than delivering absolute clarity. Just as a traditional Sumi-e ink painting embraces negative space to evoke meaning, these digital distortions leave gaps that the human imagination instinctively fills. The imperfect, the incomplete, and the ambiguous become essential components of the aesthetic experience.

A machine's failure to perfectly reconstruct an image is often seen as an error—a deviation from the expected. Yet, just as an aged, cracked tea bowl holds more meaning than a mass-produced cup, AI’s imperfections tell a story. These glitches are not failures but markers of growth, adaptation, and the machine’s own fallibility. In embracing these computational ‘flaws,’ Glitch as Intention redefines failure as a form of artistic evolution—an acknowledgment that even within the realm of artificial intelligence, true beauty emerges not from perfection, but from the elegant embrace of the imperfect.

By aligning AI glitch art with the philosophy of Wabi-Sabi, this work invites viewers to reconsider their relationship with technology—not as a force striving for sterile perfection, but as a medium capable of expressing organic, transient, and deeply human aesthetics. In a world obsessed with high-fidelity images and seamless automation, Glitch as Intention asks: what if we found beauty not in the perfected, but in the broken? What if, instead of seeking to eliminate errors, we allowed them to guide us toward new forms of creative expression?

Data Collection and Preprocessing

Dataset is constructed from:

• Corrupt Image and Video Data: Scraped from archives of compressed, degraded, or intentionally distorted media.
• Glitches: Generated by applying algorithmic distortions such as pixel sorting, color channel shifts, and noise-based corruption techniques.

Each data sample undergoes preprocessing steps, including normalization, resizing, and structured augmentation to introduce controlled levels of randomness while preserving underlying glitch features. Data augmentation methods such as random cropping, spectral decomposition, and edge degradation are applied to simulate natural error propagation in digital media.

Model Selection and Training Pipeline

A Diffusion Model is employed to synthesize glitch-inspired visuals. The model architecture follows a multi-stage process:

• Feature Extraction: A pre-trained model extracts low- and high-frequency features from the dataset to inform glitch-based textures.

Generator and Discriminator Design:

• The generator is conditioned on noise vectors and input embeddings derived from the extracted features, allowing it to reconstruct and amplify glitch patterns.
• The discriminator employs a combination of perceptual loss and adversarial training to differentiate between human-created and AI-generated glitch artifacts.

Training Strategy:

• The model is trained over multiple epochs with progressively increasing corruption parameters to reinforce glitch characteristics.
• Fine-tuning is carried out with transfer learning techniques using a secondary dataset consisting of abstract art and photography.