Impossible Golf Courses

COYOTE CANYONS

UTAH, USA // 37.5930° N, 112.1871° W

PROMPT: /imagine: A stunning par-4 golf hole located at the Many Pools Trailhead in Zion, Utah, United States.” 

Once Padi Hills had been thoroughly explored, I began thinking about other landscapes that might accommodate a spectacular looking golf course. I had hiked around Bryce Canyon and Zion National Park on a trip to the US in 2018, and the scenery there had stuck with me ever since. After researching some reference locations and course design inputs, I started generating. The extreme, exposed landscape proved more challenging than the heavily forested Padi Hills, but through primarily a process of trial and error, a style emerged that felt plausible.

After exploring the Canyon floor and water hollows, the temptation to explore the upper reaches of the canyons was inevitable, and some extreme if not impossible holes started to emerge. I kept some of them in the series, and it certainly made good fodder for instagram (by far the highest performing series from the project). Some who discovered the post organically (as opposed to existing followers who knew a little more about the project) thought the course was real, and started suggesting buddies trips to Utah, before being disappointed to learn Coyote Canyons does not (and will definitely never) exist. But for my taste some of these shots dance a little too close to the line of implausibility. As a photographer that is where a good portion of the interest lies in this technology – can something generated inside a server somewhere be as good (or better than) reality? The answer for me is ‘not yet’ – the resolution and finer detail just isn’t quite there yet. But ‘when’ could be a lot sooner than many expect. Advances in this technology are being talked about in WEEKS, not months or years. I fully expect by the end of the 2023 it will be impossible to tell the difference, with the naked eye at least, between a real world photograph and an AI generated image. The implications of which are hard to predict.

How have you seen AI start to impact your work in building architecture, and do you see these potentially transferring across to golf course architecture? 

As the broad scale application of AI is fairly recent we’ve only just started thinking about the potential use of it in architecture. Previous to AI we have used algorithmic programs like Grasshopper that integrate with 3D modelling programs like Rhino and Revit to generate and iterate design options through parametric inputs. I would imagine if AI integrated into these programs we could generate design options at a much higher rate. There’s already AI based software that can spit out simple residential multi family buildings and office layouts from a series of drop down lists. The program claims it can also adhere to building codes and coordinate structural, mechanical and electrical services. This type of technology would seem to be particularly attractive to residential developers who want to generate buildings that have the highest ROI. My firm typically designs community buildings funded by municipalities who care more for the user experience where the measure of success is more qualitative and less quantitative in nature. I’d like to see how AI can design a building taking into more nebulous parameters like light quality and spatial composition.

What drove you to first start experimenting with AI generation models for golf course designs, and what was your reaction when you started to see what was coming back? 

I had seen some images posted on Twitter by a member of the virtual design community Peter Flory (of Lido Fame) and was kind of blown away and mystified by what he was getting. At the time I figured it was some complex software he was using. After some people in my office started experimenting with the program Mid Journey, which runs as a bot through the Discord app, I realized that the generation of these images was really as easy as typing in what you want to see. I figured I would give it a try and was amazed by what it could generate given all the layers of text information I would input.

Can you share some of the techniques and ideas you have used in your prompts to generate these images? 

The main function in MidJourney is “Imagine” where you type in a prompt and it spits out four images. You then select the image that you like and upscale it to a single detailed image. You can also take that image and create variations of it. Each render takes less than a minute so you can iterate very quickly. At first I was typing in some descriptions like “golf hole set amongst sand dunes”. The results were ok but they seemed kind of generic and a bit cartoony. I then continued to add more layers of description like “a hyperrealistic wide angle image of a par 4 golf hole set over rugged rolling terrain along grassy Pacific Northwest coastal dunes with dense cypress trees, long fescue grasses and gorse bushes. Bumpy terrain. Rugged bunkers cut into and blended into natural landscape. Windswept. Barren. Rumpled fairways. Stormy skies. Misty ethereal with diffuse light.” The more info I put in, the more granular and detailed the results were.

There is a second function within Midjourney called “Blend” which I’ve found gives you a bit more control of the output. Rather than a text prompt, you select two images to blend together. For this function you can use both the images created from previous renders as well as  real-world images of golf courses. For example I could take an image of the 10th hole at Pine Valley and cross it with an image of the 12th hole at Augusta resulting in a vaguely familiar hybrid. I’ve also found you can save images of just landscapes from the internet and then blend with an image of a golf hole to allow you to choose the landscape that gets reflected in the image. After many AI renders you end up with a palette of images that you can combine in different ways depending on what you are trying to achieve. For example, I might want one of my renders to have more fairway so I’ll take the original render and blend it with an image of the 18th hole at St Andrews. What’s great about this function is that you can start to manipulate the results rather than simply describing what you want and having the AI do it all. I’m hoping in time we’ll see more AI functions that will give us even more creative control over the images.

The visual impact of the generated images that come back can be quite striking, and perhaps a starting point for some ideas, but this probably has limited use in a practical sense. What do you think is the next step of AI in golf course design (and potentially construction) that will make a truly useful tool in the process?

Beyond purely aesthetic matters I don’t see why AI couldn’t be used for more advanced spatial problem solving like routing a golf course. With the Lidar data available now you would think that it would be possible for it to read the terrain and generate a sequence of holes based on a set of predefined  parameters like number of holes, number of par-3s, 4s, 5s, maximum length of a course, maximum cross slope for a fairway, hole proximity and safety, blindness, orientation to specific views, prevailing winds. It might be a challenging bit of programming for the software to make this work but it doesn’t seem too far off.

That said, I’m not convinced it will spit out something better than what Tom Doak could produce but maybe it could be used as a tool for early exploration of design options. For sites like Sandhills where there were literally hundreds of potential routings you would think it might be a way to generate options very quickly. As I noted above in regards to architectural design,  AI might be limited with respect to the qualitative aspects of the routing and design process and perhaps will never replace the well trained hand and eye of a golf architect completely.  Most of greatest golf architecture is the result of an architect being on site with an intimate knowledge of and hands-on connection to the site and its great to see a new generation of architects carrying on that tradition. When I listen to Bill Coore talk about spending countless days just walking around a site its hard to argue with that approach when you look at the results!

There might be some more technical applications of AI if the course is prebuilt in a digital model, for instance a course could use AI software to look at site drainage resolution and analysis.

Routing a golf course is the sort of deep thinking problem that tends to separate the truly great designers from the pack. It seems to require great imagination and intuition. Is this perhaps where AI could be of greatest benefit? In what practical ways might it help this process?

At this point I think it works well as an idea generator. For example you could take a picture of the raw landscape of a site and blend it with a hole with a bunkering scheme that you like, the result could be an early visual concept for the course. We use the term precedence a lot in architecture and its basically using completed buildings for design inspiration. With AI we can now create a realistic precedence without requiring that the building exist in real life. I think the same could be applied to golf design. Furthermore, in architecture studios one thing that is often done early in the design process is to find ways to break out of your normal way of thinking. For example you’ll be asked to draw a section view of a building and then part way through they’ll ask you to shift your thinking and make a plan out of the section view. The process is disruptive but can lead you down unexpected paths that you wouldn’t necessarily arrive at on your own. Perhaps the AI process could be used to help some designers think about things in a different way and break out of the box a bit.

With bulldozers and shaping equipment now offering millimeter level precision, it is feasible that an entire course can be constructed from a 3d model, as we have now seen in practice with The Lido at Sand Valley. What possibilities does this open up in terms of automation and integration with AI?

If the hypothetical software existed to generate an AI course routing from Lidar data you could feasibly create and build a course almost completely via AI with minimal human input; barring the limitations as described in the response above. To be honest, the wholesale AI approach, much like building architecture created from drop down lists, sounds kind of scary and could result in a renaissance of bad golf architecture with AI generated courses delivered efficiently but perhaps lacking the nuance and delight usually brought through the collaboration between the golf architect and their associates on site. Perhaps this would usher in a generation of courses analogous to the post war era where the promise of technological advancement in golf design created a greater detachment from the site resulting in courses devoid of a sense of place. I think AI should be used where it can elevate design exploration and perhaps expedite some of the more tedious technical aspects of golf design and construction.