Samuel Cheung

One of the highlights of working with City Tech’s Center for Holistic Integrations, directed by Professor David Smith, and participating in City Tech’s Undergraduate Research Program, was exploring the intersection of AI-generated art, physical fabrication, and human creativity. Images play a crucial role in how we perceive and engage with the world, offering new perspectives and expanding the way we tell stories—especially when blended with traditional formats like shadow puppetry.

During last year’s research cycle, I focused on how AI-powered image generation—particularly using tools integrated with ChatGPT—can support visual storytelling. While AI can produce compelling visual content, it also comes with limitations that require human oversight and creative direction. Image generation has been essential in helping me create immersive settings for my projects, especially in Twine, a text-based game development tool, where visuals are otherwise absent. Integrating AI-generated imagery allowed me to introduce consistent character references and enhance the narrative experience.

Character development proved to be one of the more challenging aspects. AI-generated characters often lack visual consistency and coherence, which can disrupt immersion. To address this, I employed two design approaches:

Template Characters: Structured with front, side, and back views to ensure consistency throughout various puppet and digital scenes.

Generative Characters: Built from a core design that could adapt to dynamic storytelling needs, while still maintaining recognizable traits.

Additionally, AI-assisted design was crucial during the fabrication phase of our shadow puppets. Starting from hand-drawn or AI-refined sketches, we used chipboard—a versatile material—for most humanoid puppets due to its flexibility and stage performance reliability. We also experimented with acrylic, but due to its weight and fragility, we limited its use. Improvements in laser-cutting accuracy and design templating helped reduce the risk of wardrobe malfunctions that had occurred in previous performances.

In the fall, I pushed ChatGPT’s image generation capabilities further, evaluating how well it retained visual traits over time. Later, in the spring, we refined this process through Scenario AI, which helped maintain a unified art style across characters and backgrounds—key to achieving a cohesive visual identity.

During the Spring 2025 semester, the focus of my research was on improving the visual consistency of assets created with AI for our shadow puppet production. This involved refining how prompts were written and introducing a more structured method for generating characters and backgrounds.

One of the key changes was the use of base prompts and a fixed visual template to guide the generation of character images. In earlier projects, characters often looked different from scene to scene due to inconsistent AI outputs. By reusing a controlled set of prompts and maintaining the same visual references, the results this semester were more stable and consistent.

The approach also improved the backgrounds used in the production. Compared to previous iterations, the visuals this time showed fewer rendering issues and felt more cohesive. Most of this came from adjusting prompt phrasing and selecting outputs that matched a defined visual style.

Rather than using AI freely, the process this semester focused on narrowing control—using the tool less for experimentation and more as a support to maintain design continuity. This helped reduce variation and made the assets easier to integrate into the final performance.