Choe Gun

<Chaos with Breathing> is an experimental project that generates new audiovisual imagery by extracting the sensory structures embedded in painting, installation, sculpture, and object images from the exhibition dataset provided by Total Museum, and transforming them into AI generative rules. In this process, AI was not employed merely as a tool for image enhancement, but as a core generative apparatus that analyzes the visual characteristics of the dataset as a sensory language and translates them into reproducible visual structures. The central focus of the work lies not in directly replicating dataset images, but in identifying recurring sensory features and transitional structures across images, and converting them into sensory codes and halftone screen generation rules.

was developed by constructing a sample set drawn from painting, installation, sculpture, and object images within the dataset, and analyzing the visual characteristics of each image and their interrelations through a sensory language framework. In particular, the space between artworks was conceptualized as “spacing.” Here, “spacing” does not refer to mere emptiness, but to intervals of sensory gap, breathing, transition, collision, and resonance that emerge between distinct works. Accordingly, the dataset was interpreted not simply as a collection of individual images, but as a structure encompassing the sensory relationships generated between them.

To interpret these “spacing” intervals, sensory language—including onomatopoeia, mimetic words, adjectives, and state descriptors—was extracted. Expressions such as “crisp,” “dense,” “crunchy,” “subtle,” “sharp,” “seeping,” and “floating” were used as analytical axes that determine the direction of visual transformation.

The analysis began with a six-axis sensory code system:

• Crisp: a feeling of clear openness
• Subtle: a quietly permeating state
• Crunchy: a brittle, breaking sensation
• Dense: a sticky, lingering state
• Heavy: a downward pressing weight
• Sharp: a piercing, striking sensation

It was then expanded into a ten-axis system for greater precision:

• Crisp / Subtle / Crunchy / Dense / Heavy / Sharp
• Soft: a gently enveloping texture
• Vivid: a sharply striking clarity
• Seeping: a slowly spreading state
• Floating: a buoyant, suspended sensation

Finally, for direct application in generative rules, these were compressed into a five-axis transformation model:

• Ventilation / Residue / Fragmentation / Viscosity / Weight

Thus, the dataset analysis process consisted of:
analysis of individual visual features → interpretation of “spacing” between works → extraction of sensory language → encoding into sensory axes.

Sensory Translation Algorithm
The sensory translation algorithm of converts extracted sensory language into rules for visual image generation. The images were first quantified along sensory axes, then compressed into the five-axis model, and subsequently mapped onto CMYK/halftone plate structures.

The correspondence rules were defined as follows:

• Ventilation → intensity of the Yellow (Y) plate
• Residue → intensity of the Magenta (M) plate
• Fragmentation → intensity of the Cyan (C) plate
• Weight → intensity of the Black (K) plate
• Viscosity → adjustment values for dot density, diffusion, and pressure

Halftone screen angles were fixed as:

• C = 15°
• M = 45°
• Y = 75°
• K = 90°

Rather than treating each sample as an independent endpoint, a sensory interpolation method was introduced to address the “spacing” between samples as a generative interval. This interpolation was designed as a nonlinear transformation in which each axis evolves at a different rate. For example, fragmentation, vividness, and floating decrease rapidly in the early phase; viscosity increases from the मध्य phase onward; and weight intensifies more strongly in the later phase.

The algorithmic structure of can be summarized as:
dataset images → extraction of sensory language → scoring along sensory axes → compression into five axes → conversion into CMYK/halftone rules → nonlinear interpolation between samples.

AI Generation Method and Models
The process utilized OpenAI ChatGPT and an image generation model (DALL·E-based). The generation pipeline was designed so that the sensory structures extracted from the dataset functioned as the core parameters of the prompts.

Prompts were structured around three primary components:

1. Dataset-based sample designation
   Painting, installation, sculpture, and object images were defined as reference points (e.g., Sample 1, Sample 2, Sample 3).
2. Input of sensory axes and plate rules
   The five sensory axes and CMYK/halftone plate specifications were incorporated. For example: high ventilation, high fragmentation, low viscosity, low weight; CMYK angles set to 15/45/75/90; with fine, sharp dots and low pressure.
3. Specification of visual generation methods
   Descriptions included halftone screens, CMYK dot screens, print textures, dense dot structures, pressure variation, radial halftones, and airy layered screens.

Thus, the prompt was not a simple stylistic instruction but a structured translation of dataset-derived sensory analysis into generative rules.

Manual Intervention and Digital Retouching
The process did not conclude with AI-generated halftone and CMYK dot screen images. Instead, it incorporated a stage of manual intervention through direct hand retouching. Using Procreate, this phase was conceived not as mere correction, but as a transformation and reconfiguration of the AI-generated print screen structures through manual input.

In particular, this process drew from the working method of Ayako Rokkaku, reinterpreted through digital retouching. Rather than imitating the final visual appearance, the approach sought to emulate the artist’s gestural movement—spontaneous intervention, surface traces, and the continuously shifting and spreading state—within a digital environment.

Accordingly, the generated images underwent the following process:
dataset-based sensory analysis → AI-generated halftone/CMYK imagery → manual retouching → reconstructed visual composition.

Output Format and Final Result
The final output takes the form of an MP4 video (43 seconds), rather than static images, unfolding sensory transitions between samples and transformations of halftone screen structures along a temporal axis. Instead of directly presenting the original dataset images, the video constructs a continuous sensory field composed of intermediate states generated according to sensory axis codes and halftone plate rules.

In this process, the dataset functions not merely as a reference, but as the core system driving sensory analysis, quantification, plate translation, AI generation, manual retouching, and video transformation. AI is thus positioned not as a post-processing or enhancement tool, but as a central generative mechanism that translates extracted sensory structures into new visual outputs.

Tools and Format

• AI Models: OpenAI ChatGPT; DALL·E-based image generation
• Digital Retouching: Procreate
• Output Format: MP4, 43 seconds

Technical Summary

• Dataset-based sample extraction
• Sensory language analysis
• Construction of 6-axis, 10-axis, and 5-axis sensory models
• Conversion to CMYK/halftone plate structures
• Nonlinear sensory interpolation between samples
• AI-generated imagery
• Manual digital retouching of halftone/CMYK images
• Video output