Research

AI Image Competition: Kai Nishimura Gives a Walkthrough of Round 1

At PLP Architecture, we are hosting an in-house AI Image Generation competition where our team members utilize AI technologies to respond to a series of architectural and urban briefs. The initiative gives us a platform to train in using AI tools, explore the opportunities of using AI in an architectural context, and to have a bit of fun while we’re at it.

For Round 1 of our in-house AI Image Generation competition, Kai Nishimura speaks to us about the processes behind the image that he created. Over to Kai …

Step 1: Planning the Workflow

To begin with, I used both ChatGPT 4o and Midjourney to create my images. I did this in two different steps. The first was to dig deeper into the brief and create useful prompts, then was to generate images and tweak the prompts to get increasingly closer to one that answered the brief.

Step 2: Exploring the Brief

Initially, I used ChatGPT to:

Expand on the brief to make an urban planning strategy.
Use this to make a scenario plan and go into more detail about the components of the new city. I used ChatGPT to build on the brief and actually start thinking of what this new city would look like.
Turn this into prompts for Midjourney.

I started by asking ChatGPT to take our brief and propose an urban solution to it.

ChatGPT then broke this down into four phases, exploring everything from stakeholder engagement and site selection, to exploring sustainable mobility and community involvement.

I then asked ChatGPT to use this framework and describe what this new city would be like ‘in as detailed and concrete terms as possible’.

Step 3: Refining the Prompts

Then ChatGPT was asked to turn this description into a Midjourney prompt, specifying some of the criteria mentioned in the brief. With a little bit of adjusting, I arrived at a prompt that I was happy to start off with.

Step 4: Generating Image Options
From here, the process was to generate images and adjust the prompt as necessary.

This is critically where the AI and human designer work hand in hand. Whilst Midjourney can generate options, it is for the human to decide which design solutions actually go some way to tackle the problem and are implementable in real life.

I generated a series of images, adjusting the prompt each time, until I had developed a solution that I thought addressed the brief suitably.

Stage 5: Final Edits
The final stage was to use the ‘upscale’ and ‘in-paint’ functions to both increase the resolution and to make changes to specific parts of the image. In this case, that included adding in things like the mountainous background.

Take Aways
Although I thought my final image tackled the brief well, I felt that I wanted to make adjustments to make it more beautiful, realistic or cinematic. But, I was unable to do this by editing it in Midjourney. Any attempt to add some adjectives to the prompt and regenerate it, completely changed the image. In this case, Stable Diffusion would have been a more useful tool.

For the next round, giving a specific context and site boundary would create a more realistic workflow by more closely simulating a real project.