Div

A generative AI web design tool that helps to tailor every design element of a website to specific audience

GenAI

Design Tool

Design System

Web App

Timeline

May 2023 - Aug 2023

Team

Effie Zhu / Isabel Li

Tool

Figma / RunwayML / ChatGPT / BroseAI / RelevanceAI

Role

Researcher / UI/UX designer / Prototyper / Project manager

Background

Gen AI should not replace the design process from start to finish. Instead, it should serve as a co-pilot within the workflow, with designers maintaining complete control.

Challenge

How might we help creatives better align their web designs to diverse user needs and expectations to improve audience engagement?

How might we help creatives better align their web designs to diverse user needs and expectations to improve audience engagement?

How might we help creatives better align their web designs to diverse user needs and expectations to improve audience engagement?

How might we help creatives better align their web designs to diverse user needs and expectations to improve audience engagement?

How might we help creatives better align their web designs to diverse user needs and expectations to improve audience engagement?

Our Solution

We developed key features for the GenAI model to identify and cater to different audience groups, guided by human input. The system customizes the website for users by analyzing their cookies and IP address, ensuring a personalized experience that aligns with each user's knowledge and needs.

Centralized Directory

Easy and comprehensive access to products, stores, and events information

Smart Navigation

Using users' input and advanced algorithms to provide the most efficient navigation

Personalized Assistant

Using real-time data to anticipate users' needs to provide recommendations

Prototypes

Userflow 1

Design Ecommerce Website Glamora

Userflow 2

Design
Personal Portfolio

Technology Centered Research

Initial Research

We did secondary research and a user survey around community of LLMs (Large Language Models), especially ChatGPT, which was defined as any human involved in the training of ChatGPT and any human impacted by the presence of ChatGPT.

Key Finding 1

LLMs have biased behaviors reflected by their trainers and assessors.

ChatGPT’s responses and content classifications, such as if a statement is hateful, show a liberal biases to certain topics even though it tells users it doesn’t have any political opinions.

Key Finding 2

People prefer ChatGPT’s work belongs to no one that belongs to everyone.

42.5% of the survey participants believed “the rights of the work of ChatGPT should belong to no one,” while only 29.3% responded “agree” to the statement “the content generated by ChatGPT belongs to all the users of ChatGPT.”

Key Finding 3

People don’t consider ChatGPT generated contents are of their own intellectual property.

Only 12.2% of the survey participants agreed that the contents they generate with ChatGPT are of my their intellectual property.

Primary Research

To understand how generative AI models work, we tried out existing models on RunwayML. During our explorations, we realized the model is not good at generating two-object combinations. We also learned that generative AI have pre-trained style. With the insights, we decided to train two models: one model of two people to see if the AI will recognize two people as individual human or as a whole, and the other model of an instagram artist’s style to see if the AI can recognize the specific style. We also tried Framer AI for website generation, but we realized its limitations in design and content.

Key Finding 1

The AI combined the features of two people from the input photos onto one person's face in the output.

Key Finding 2

The output pictures displayed features of other asian people from the pre-trained datasets, not just from our input.

Key Finding 3

In the output pictures, the nose is the facial feature that most closely resembles the one in the input photos.

Further Exploration

To further understand our findings, I found the facial recognition landmarks used by RunwayML in their Github. Opposite from my assumption, the nose has the least landmarks, which leads to my conclusion that nose is most accurate because it doesn’t move as much as other features. AI only recognize patterns repeated in the dataset instead of recognizing shapes as our eyes do.

Input
Output
Key Finding 1

The output photos only show a basic background and a few colors from the input photos' styles.

Key Finding 2

AI can only recognize repeated patterns from the input photos and apply them to the output.

Framer AI
Key Finding

The websites generated by Framer AI are limited in layouts and body copy content.

Design Brief

Based on our research and exploration, we wrote a design brief for our final project.

Ideation

We created low-fidelity wireframes for 5 use cases and did several variations for different use cases. The five use cases were around how GenAI might help the creatives in different stages of their design process, including research, visual design, prototyping, etc.

Proof of Concept

After ideation, we narrowed down our final design idea to using Gen AI to assist creatives in designing and editing websites more quickly and effortlessly. Therefore, it was crucial to confirm that this concept was feasible with generative AI technology. To validate our idea, we experimented with 8 Gen AI tools, testing whether they could be prompted to directly modify a website's code.

Proof of Concept Demo

Demo of the ChatGPT 4.0 reading and rewriting the .html code

UI Design

For the final user interface, we designed both light and dark mode for users to choose based on their preferences. We wanted our tool to look fluid, balanced, and muted, so it wouldn’t distract users’ attention and would keep the dynamic feeling of generative AI.

Light/Dark Mode Exploration

Style Guide

Iconography
Colors
Font Family

Varta

ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz

CTA Button

Heading 1

Heading 2

Body

Small

Semibold / 17

Light / 17

Regular / 15

Regular / 14

Regular / 12

Iconography