AI-Driven Generative Project
Runway AI
Design Brief
For this project I was asked to experiment how AI-driven generative work can support my design practice, specifically focusing on my senior capstone, and document the process. It was an exercise to discover how AI-based tools can support my efforts - in research, ideation, prototyping, testing, communication, and other tasks in my project.
My aim is to create marketing collateral images that highlight the conversation ceramic pieces I created for my capstone project.
Software:
The Tool: Runway AI
I explored Runway’s Gen-4 model, a text-to-image generator that uses AI to create visuals based on prompts and reference images. One of its standout features is References, which allows you to generate consistent characters or settings across multiple images using a single reference image.
This tool is designed to preserve key traits—like facial features, clothing, or environments—while letting you change lighting, style, or background through your prompt. It’s particularly useful for building cohesive visual narratives or branding materials where consistency matters.
Runway recommends using high-quality, well-lit references and detailed prompts for best results. While References currently focus on character and location preservation, the platform notes that more reference types and features are in development. Because it’s still a relatively new tool, Runway encourages experimentation. Runway AI also provides a guide on Creating with Gen-4 Image References.
The Process: Iterating with Runway’s Gen-4 Model
As I worked on creating marketing visuals for my capstone project, I experimented with Runway's Gen-4 model using the References feature. The goal was to generate consistent, styled images of ceramic pieces I designed—playful teacups with conversation starter prompts—across different settings and aesthetics. Below is a breakdown of my process, including the prompts I used and what I learned from each iteration.
1st Attempt
Prompt:
A @dinnerscene featuring @conversationceramics in an intimate @tablesetting
Reference images 1) dinnerscene 2)conversationceramics 3)tablesetting
Result:
The model combined several reference images into one composite scene, but the result lacked clarity and polish. I realized the prompt was too vague and didn’t provide enough visual or emotional detail.
1st Result
2nd Attempt
Prompt:
Two ceramic @cups with playful conversation quotes, held by friends laughing together in a sunny backyard brunch scene.
Reference image 1) cups
Result:
This version captured the mood well—friends laughing in a sunny brunch setting—but the cups were strangely oversized, and the food consisted only of fruit, which felt off-brand.
2nd Result
3rd Attempt
Prompt:
Two ceramic tea @cups with playful conversation quotes, held by friends laughing together in a sunny backyard brunch scene. The photo has a film photography feel, with warm lighting. The brunch scene is aesthetic and minimal.
Reference image 1) cups
Result:
Adding stylistic details helped refine the output. The lighting and composition improved, but the scale of the cups was still too large.
3rd Result
4th Attempt
Prompt:
The @teabowls are displayed on a pedestal at a gallery exhibition alongside other ceramic pieces featuring conversation prompts written on them as seen in the reference photo.
Reference images 1) teabowls 2) reference photo
Result:
I pivoted to a more curated, exhibition-style image. The gallery setup looked clean, but again, the teabowls were disproportionately large.
4th Result
5th & 6th Attempts
Prompt:
The @cups are tea bowl size and appear next to the @flowers in a beautiful tablescape.
Reference images 1) cups 2) flowers
Results:
The first image was close to what I envisioned, though the second cup generated looked off. When I tried prompting the model to remove it, it instead added a teapot spout, diverging from the reference entirely.
5th Result
7th Attempt
Prompt:
A close-up product shot of a ceramic @product with a conversation starter prompt.
Reference image 1) cups
7th Result - First Iteration
6th Result
Result:
This was the most successful result. The image closely resembled my original product, with no bizarre additions or distortions. The visual clarity and focus made it ideal for a marketing asset.
7th Result - Second Iteration
Reflection
One of the most challenging parts of working with Runway’s Gen-4 model was learning how to write effective prompts. I initially underestimated how much the wording, order, and specificity of a prompt could impact the output. Many of the flaws I saw—like oversized cups, irrelevant objects, or missing design details—weren’t just technical failures. They were often tied to the way I was communicating with the model.
As someone new to prompt engineering, I found it difficult to balance being descriptive without overwhelming the model. Sometimes I was too vague; other times, too specific in ways that caused the AI to focus on the wrong things. Each attempt taught me how sensitive the tool is to phrasing—and how much trial and error is required to get results that feel intentional.
While Runway’s References feature showed promise in maintaining aesthetic and character consistency, it often struggled with accuracy—especially in product-focused scenarios where scale and detail really matter. In the end, the tool worked better as a conceptual aid than as a source for final assets.
This process helped me realize that using AI creatively isn’t just about having access to advanced tools, it’s about developing a new kind of literacy. Prompt writing is its own craft.