Intergalactic Frequencies

Series I - IX.

In a time when art and technology intersect at an unprecedented pace, Intergalactic Frequencies explores the merging of human creativity with machine intelligence. Centered on the concept of iterative generative processes, the project begins with a simple hand-drawn sketch and evolves into a surreal visual experience through advanced AI techniques. Inspired by the dynamic, otherworldly qualities of celestial phenomena, Intergalactic Frequencies questions the boundaries between human intuition and algorithmic creativity.

The motivation driving this exploration is twofold: First, to understand how AI can interpret, reshape, and iterate upon human-generated content, offering an enhanced form of digital creation driven by human intent. Second, to challenge common expectations of generative AI, especially in terms of abstract versus photorealistic output. While much generative AI research focuses on achieving photorealism, our work pivots towards abstraction, intentionally keeping guidance low to observe how the model interprets minimal prompts. Interestingly, we found that the less detailed the prompt, the more unexpected and creative the results became—an observation that hints at the model’s inherent ability to explore beyond the confines of precise human input. While not mathematically conclusive, this aligns with the idea that reduced prompt specificity can increase the diversity of generated outputs.

Throughout the iterative process, the model consistently introduced facial structures and recognizable elements such as buildings, despite no such features being specified. This tendency prompts an inquiry into the inherent biases of AI models, raising the question of whether AI is capable of true abstraction, or if its training data inevitably pulls it back towards representational forms. Furthermore, this project contemplates the aesthetic value of so-called "flaws" in AI-generated images—elements like distorted hands, morphed faces, and non-human features, which are typically dismissed in the pursuit of photorealism. We aimed to embrace these imperfections, exploring how they contribute to the uniqueness of the iterative process and the evolving visual narrative.

By stripping down the prompts and minimizing the algorithm's guidance, Intergalactic Frequencies seeks to reveal any latent creative potential within the model, offering insight into how generative AI can function not only as a tool for mimicking reality but as a tool in the creation of abstract, evocative art.

Product

Intergalactic Frequencies consists of a series of nine short sequences, generated through iterative processes using Stable Diffusion and ControlNet. At the core of the project is a hand-drawn ink marker sketch, which acts as the seed for all subsequent visual transformations. Over several weeks of experimentation with a single prompt, 'intergalactic frequencies', the project leverages the innate capacity of AI to refine and morph visual features, subtly shaping facial structures and dream-like landscapes from the abstract.

Through continuous re-tuning of guidance and prompt scales, and by feeding the generated images back into the system, a multi-layered visual narrative emerges—one that reflects the direction of the artist, with AI as a tool enabling novel transformations. The result is a sequence that pulsates with energy, like cosmic waves of frequencies passing through various dimensions, highlighting a new form of expression that is authored by human, using machine learning models as a tool.

The results were processed further and published online as experimental video shorts, available on youtube and instagram.

Process

The technological backbone of Intergalactic Frequencies is built upon several intertwined processes, each enhancing the artist's ability to iterate using AI-driven techniques:

• Generating from a Principal Sketch: The project begins with a hand-drawn ink sketch, a spontaneous imprint. This sketch is fed into Stable Diffusion and ControlNet to generate initial iterations of the sketch. By keeping the input image constant, the system gradually builds upon the underlying visual structure, offering new interpretations at every step.
• Selecting Process: Using a single prompt—intergalactic frequencies—the model produces a wide array of images over several weeks. From these, images are selectively chosen based on emergent aesthetic patterns. Interestingly, after several iterations, the model began to develop consistent facial structures, which were further enforced through parameter tuning and selective feedback loops.
• Sorting with TensorFlow: Once the preferred images were selected, TensorFlow's ResNet model was employed to sort and organize the sequences based on similarity. This allowed for the creation of smooth, cohesive transitions that maintained a visual narrative arc throughout the series, ensuring that the evolution of forms felt organic and intentional.
• Morphing between Frames with AfterEffects: To bridge the gaps between images, Adobe AfterEffects was used with time-warping techniques, creating an illusion of fluid motion between still frames. While some frames were skipped due to the nature of the effect, it highlighted the need for further development, particularly involving Generative Adversarial Networks (GANs), to enhance frame-to-frame continuity.
• Generating Transition Frames (Interpolation): The next phase involves developing a network specifically designed to generate intermediate frames between each sequence. This will allow for smoother transitions and a more cohesive visual experience, ultimately blending the images together in a seamless, evolving flow.

Available on mintbase.xyz

Process intergalactic frequencies - Renders with extended, experimental prompt, first attempts

Example of input parameters for the above structures:

"prompt": "full lenght shot, mycelium structures, frequency of enlightenment, growing shapes, vanishing point, architecture, abstract collective intelligence, sketch, consciousness, fungal structure, non-finito, concept art, octane render, Line Art, Geometric, Ink, Linocut, Synthwave, Melancholic, by Yoshitaka Amano, Intricate Environment",
"num_inference_steps": 100,
"guidance_scale": 15,
"sampler_name": "dpm_solver_stability",
"use_stable_diffusion_model": "sd-v1-5"

Test render after reducing the prompt to "intergalactic frequencies" and tuning parameters.

Example of input parameters for the above render:

"prompt": "intergalactic frequencies",
"num_inference_steps": 60,
"guidance_scale": 7.5,
"sampler_name": "lms",
"use_stable_diffusion_model": "sd-v1-5"

Process intergalactic frequencies - Advanced findings after iterative generation and parameter tuning, using generated images as input for ControlNet (Canny, OpenPose, Depth and others).

After dialing into a certain desired aesthetic, the model generated more detailed and interpretative structures, even though the prompt was very open and non-descriptive.

Sorting with TensorFlow

In testing video sequence results, we explored both the VGG16 model and ResNet50 for image similarity estimation, ultimately opting for ResNet50 due to its superior performance and quality in rendering consistent visual sequences. ResNet50 provided a more refined ability to capture and align the intricate features of each frame, ensuring smoother transitions between images.

An alternative approach was also tested, which involved breaking down the input images into tiled sections and clustering these tiles based on feature similarity. The idea was to sort frames by comparing smaller regions, hoping to achieve more nuanced transitions by isolating specific areas of interest. However, this method proved computationally intensive and yielded inconsistent results, as the clustering of tiles often led to fragmented sequences that disrupted the overall fluidity of the video.

link to notebook example >>>

Ultimately, this approach was abandoned in favor of a more holistic sorting technique with ResNet50, which better preserved the continuity and integrity of the evolving imagery.

Policy

This project encourages us to consider the evolving role of technology as a tool in the artistic process, particularly in the context of Intergalactic Frequencies. The AI was used as an instrument—much like a brush or chisel—allowing for new creative outcomes based on human guidance and decision-making. While the AI provided unexpected visual interpretations, it remained a tool in service of the artist’s vision, rather than a creative agent in its own right. This distinction is crucial as we contemplate the future of artistic authorship in an era increasingly influenced by technological mediation.

The work also touches on ethical considerations surrounding the use of AI in creative fields. As AI tools grow more sophisticated in generating visual content, the question arises: how do we maintain the value of human creativity in a landscape where machines can produce convincing artistic output? The role of the artist remains central to Intergalactic Frequencies, guiding the AI’s iterations and making crucial creative decisions, highlighting that human intent is irreplaceable in meaningful artistic production.

By employing AI as a means to extend creative possibilities, rather than replace human input, this project underscores the symbiotic relationship between artist and technology. The AI’s capacity for generating surprising and unconventional results serves to expand the boundaries of human creativity, but always within the framework set by the artist. This work looks toward a future where technology can enhance artistic exploration, while ensuring that human imagination and agency remain at the core of the creative process.

Intergalactic Frequencies stands as an exploration of how digital tools can push the limits of what is visually possible, without compromising the integrity of human-driven artistic vision.