Best prompts and settings

[Illes Papp]

Do you have the "perfect prompt" or have you found wonderful balance of the settings that works for you and want to share with others? Don't keep your best prompts and settings for yourself, but share it with the community and check what others are discussing.

[Aixtreme]

I'm more afraid that we're training architects who can operate an AI but can no longer design.

Here is an example from my students for a beach house. The designs were mercilessly bad and all of them would have failed purely in terms of design and concept. But the renderer produced excellent results.

 

[Jeno Barta]

Yes, AI is a tool, not a full solution. We have seen these transformations in the past, for example when graphic designers changed from analog to digital. They got insane benefits, like they could undo, the got a lots of automatic tools etc., but it does not replaced talent, and profession. 

There are a lots of people, who think AI image generators will generate images that can be monetized, or implemented directly. For example in the domain of video games / game design there were a lots of graphic designer layoffs, as big companies thought they dont need them. Sure an AI can generate images, but these images wont be perfect, and it can fool a viewer just for a glimpse, but not for a second time where it checked in detail. These images cannot be used directly, they are not asset ready. But they can be if a professional graphic designer starts inspiring from them, and modify them. 

Now this is how AI become a tool, not a full solution. The images it generates can fuel inspiration, or give us a nice base to start out from. In the domain of architecture I know people who like to inspire from orher building, from nature, or they play around with physical models, wood, and glue to concept. Inspiration can come from many sources, and I think AI Visualizer can be one of them. 

For example there are these beach houses. If they are modeled as they are on the images, they wont make any sense structurally, architecturally for sure. But an architect who has its foundations, and knows the ins-outs of building planning, can pull some inspiration out of them, and adjust the form, structure accordingly to make sense. Same as a graphic designer grabs an image that AI generated, and fixes the artifacts on it, modify some aspects.

AI is a tool, a very good tool that develops rapidly, but it wont be able to create anything like a human being, or a group for a very long time, especially on a very complex domain like architecture. But lets use the tools with sense, and pull out the best from it. 

[Jp1138]

I basically agree with you that there are other areas of the program that need more atention, or even better uses of AI for more efficient workflows. However, I think this AI visualizer "experiment" must have taken very few resources of these other areas, as it basically takes stable-difussion and makes the minimum programing to get an image from archicad into it. The fact theat it even comes without an install program show this more like an experimental feature that a complete one. At least that is the way I see it, a low resource incursion in the AI field so Archicad and Graphisoft can say there are there too. Somewhat of a mix of publicity stunt and first steps to thread the waters. At least that what I hope it is 😅

[Barry Kelly]

Maybe this should be call "AI Inspiration" rather than "AI Visualiser".

The visualisation of the model in the images it produces can be nothing like the original seed model or what is asked for by the prompts.

But the inspiration that can be gained from those images could indeed be useful.

Barry.

[Jeno Barta]

Hey Jp1138!

You are absolutely right that it's not a huge programming task to connect a foundational model through API into Archicad, but actually it has it's complexities. On the surface it seems very easy, but actually it raised a lots of questions. However the hard part was developing a parameters set, and translate them to like 2-3 representative UI elements. When you change different parameters in AI Visualizer, it does not set just one parameter in the foundation model, it's sets multiple. So finding these combinations, finding a proper underlying prompting was a pretty big challenge, and took a lots of iterations. Not mentioning that we changed from Stable Diffusion 1.5 to SDXL mid development. We took the path that we does not fine tuned a model, but create an underlying parameter set that fits our goal: "generate high quality building concepts from mass models, in almost any style our Users can come up with". You can try this out, with deleting everything from the prompt field. If you do so, AI Visualizer will still generate buildings. 

For example in the prompting we had to address, that we are working with different model styles, morphology, forms, and even colors, and AI Visualizer have to work properly with one grey cube, and a colorful high fidelity building also. 

[Aixtreme]

One of the most possible upcoming next things here might be to work with template pictures of existing buildings. A lot of generative AI allow to upload a sample image as a template for the generation process. 

What do you think about for example to upload images from Mies van der Rohe and generate appropriate AI renderings?

[scoz]

사실 AI VISUALIZER는 한계가 있고 결과가 한결 같을 때
도 있지만, 기초가 되는 모델링이 충분하면 만족스러운 결과
도 나오고 있습니다 그쪽에서는 어느 정도 만족하고 있습니다.
간단한 영감으로 사용할 수 있다고 생각합니다.

Archicad modelingArchicad AI VisualizerArchicad AI Visualizer

[Barry Kelly]

That is much better.

Still strange that some elements like the 3 windows on the left stay pretty much the same.

And the step to the top of the front facade.

Other elements completely change.

I would say the logo would be pretty important to keep and that has gone altogether.

Barry.

[scoz]

If I repeat it many times, I get a similar image
But there's another weird part