How to control AI in architecture

I recently had the opportunity to help my friend Maxime and his studio : Maxime Lavaure Architecte —an architect looking for a faster, more creative way to do prospection—by combining AI tools and Photoshop. His objective? To quickly generate conceptual sketches and designs in 3D that communicate his architectural vision

The requirement : From a 3D sketch of a house near Toulouse, I need a light atmospheric image all while maintaining high quality.

This is the before after

Here’s the process we followed:

• 3D Foundation Maxime started by creating a basic 3D model of the house. He described the setting—a scenic hillside location near Toulouse—and showed me reference materials he wanted to incorporate.

• Midjourney for Inspiration With a clear idea in mind, I used Midjourney to capture the desired mood. This helped me refine the visual direction before diving deeper into the AI workflow.

• Mood & Hand-Painting To establish a calm atmosphere and highlight both the architecture and the view, I hand-painted colours and rough textures, including where the trees and greenery would go.

• Pulze Project Dream Iterations I then fed my vision into Pulze’s Project Dream in two rounds:

• Targeted Refinements After upscaling, I refined specific areas like the bedroom and living room with smaller iterative steps—this didn’t require another upscale and allowed for precise detailing in no time.

In just about an hour, we went from a rough 3D model to a polished, high-resolution concept that Maxime can use for presentations and client prospection.

Key Takeaway: AI is amazing, but is hard to control. I found that project dream give me the control between the prompt, the original image and the reference. For each iteration I take only what I need, and the further I go in the process, the less creative I let the AI be. I can control this easily in the software of Pulze.

This process works for me, it evolves all the time as the software programs do

What do you think about this workflow? The hardest part is to actually understand the limits of AI.

Here is a video of the same image done in Runway ML (an image to video tool)