Best wishes with AI.

2023 was the year of the advance of AI . In 2024, cooperation between humans and AI will continue to develop. The possibilities of AI seem endless. But how well does it really work? Or how well does it really work? Does AI know creativity? Will AI replace the role of a designer? We conducted two test cases on how AI can be applied.

Generative design

Earlier, we tested with Generative Design. Here you develop a “ruleset” for your design task, within which computing power with algorithms gives substance to the set framework.

This is a good and meaningful variation on designs. This rule set can range from:

• programmatic principles (at x area floor split into multiple building volumes)
• numerical ratios (at x m2 GLA X number of toilets)
• optimizations: x m2 of building volume with minimal shadowing on surroundings

This involves using software and the computing power of the computer as an assistant: quickly generating and evaluating alternatives. After each iteration, both the software and the designer learn from the delivered output in order to redefine the input. With Generative design, there is often a clear picture of the desired output: it is a very specific question with a clear product as an answer. Here the design tools from the field are used: Autodesk Revit, Dynamo (with Revit API). Often the output is a step in the design and development process.

Designing with AI

With AI, the output can become very different: for example, this is suitable for getting visually strong results quickly. But will I get predictable results with ai? Can I influence that enough? The image that often arises of AI is that there is a degree of self-thinking in it. But this is incorrect on two levels: first, self-thinking is purely based on “machine learning” through a lot of historical data. It is only as good as the data used. The designer or engineer sets the goals, gives the instruction and only then does AI take over. We did a test case with our New Year’s greeting to make this principle visible.

Design process with ai .

Ai generates an image based on an instruction. We asked BING creator to create an image similar to the one we previously developed ourselves: an astronaut with a shopping spree in his visor.

With the assignment “take a picture of an astronaut with a shopping mall in his visor” comes the result below. Not exactly what we’re looking for: it’s too much of a shopping mall and not enough Christmas atmosphere.

Working with AI is like a dialogue: you indicate preferences for certain images, ask to improve a variant or adjust the instruction. The assignment was changed to ‘take a picture of an astronaut with a shopping street in his visor. Make the shopping street in Christmas atmosphere and show the moon’.

The result is already better, but still not quite what we are looking for. It is a little too much astronaut, a little too little shopping street. The moon is too big and the image is too square, not suitable as a banner. The assignment has been changed: ‘take a picture of an astronaut with a shopping street in his visor where the astronaut is in the middle of the shopping street. Make the shopping street in Christmas atmosphere.’

The result is a considerable improvement: the shopping street has a central perspective and is clearly in Christmas spirit. But actually we want the astronaut himself to be in space. The assignment was changed to: ‘take a picture of an astronaut floating in space with the reflection of a shopping street in his visor, where the reflection is in the center of the shopping street.’ Make the shopping street in Christmas atmosphere with the moon in the sky at the end of the shopping street. Make sure the visor fills most of the image.’

With this we achieve a good result. We have close up of the astronaut with a shopping spree in his visor. What this process shows is that a fine result can be achieved: but the outcome is as good as the input. The more specific the input (describing the goal and desired outcome) the better the outcome.

Applied AI

The only requirement currently unfulfilled is that it must also be suitable as a banner. BING Creator cannot meet this requirement. In Adobe Photoshop, the “generative infill” feature is an application of AI. Widening the image creates space for the infill. With the prompt ‘deep-space background like photos from the Hubble telescope,’ we get the infill as desired. Again, it took some refinements to get the right result.

This shows that AI is not a one-to-one replacement, but a tool that we need to learn how to deploy. Just as 3d models have come to support the production of 2d line drawings, AI can support the design process by quickly showing results. In a future blog, we will explore how this can be deployed in interior design. For now: best wishes from AI!