Zaš Brezar wrote two pieces on generative AI in landscape architecture, one about Midjourney shortly after its launch in 2022 and another on a suite of generative AI tools in the summer of 2023. More than a year has passed, so what is the stage of AI in and outside the profession and discipline of landscape architecture?

Many firms are now incorporating Generative AI into their workflow. Firms such as SWA have been able to fund research fellows exploring generative AI. Anecdotally, I have learned that other firms have similar internal initiatives. One trend is using LoRA (Low-Rank Adaptation of Large Language Models), a lightweight training technique that can “fine-tune” one’s Stable Diffusion models to generate images in a certain style. In one of the SWA’s research projects, LoRA was used to reproduce landscape design sketches in the style of Jim Lee, an SWA principal with a signature style of using blue-coloured pencils. “Now we have a mini Jim.” is the research team’s tagline.

Yuqing Wu, a friend of mine who works in Field Operations, recently invited me to participate in a Chinese online panel to discuss the application of AI in landscape design. Wu showcased several use cases of AI in their day-to-day work to increase efficiency in iterative design processes. For example, one popular use case is to transform a rendering into different seasons in seconds. Another use case is to generate material studies using one base image. Wu also uses their fine-tuned Stable Diffusion model to turn old project renderings into watercolor style for marketing purposes.

Material and seasonality study using one base image.

Co-presenting in the panel was Xun Liu, a lecturer at Southern California University and PhD candidate at the University of Virginia. Part of Liu’s work investigates cutting-edge usage of AI for landscape architecture. She showcased a workflow combining real-time camera feed with Stable Diffusion. Designers can modify real models and even everyday objects like plates and fabrics to generate photo-realistic architectural renderings, all in real-time. Both of their work moved beyond the concept phase, which many have reported helpful using generative AI. Their work offers a glimpse of a future where human-machine collaborative intelligence is enabled by AI-powered design workflow.

Real-time Ideation using StreamDiffusion. Image/video credit: Xun Liu

Our Chinese colleagues appear to be more entrepreneurial in this space. Designers can now easily download LoRA models trained by Chinese designer communities from Civitai.com. Some even offer short online courses on AI design workflows for under $100. However, let’s not forget that many US-based design firms have long outsourced their rendering tasks to Chinese rendering firms for a cheap price and a fast turnover. In a way, AI has exacerbated the rat race among Chinese rendering companies to produce images faster.

One could argue that the industry now finally realizes the implications of generative AI for landscape architecture design; we are moving from “early-stage exploration” to a more “strategized development and adoption” of AI into day-to-day design workflow. Perhaps the most pressing question we should ask now is: How should landscape architecture—a field where technology has always been intrinsic to its practice—respond to the emergence of AI tools?

Today, if we look at Théâtre D’opéra Spatial, the award-winning AI-generated artwork that triggered the AI art boycott at the end of 2022, one would probably say confidently, “Oh, this image is definitely generated by AI!” Indeed, after two years of training—not AI algorithms but human eyes this time—many of us can probably differentiate AI-generated images with just a glance. While fascinated by how “trainable” human eyes are, I want to emphasize instead the co-evolution of human and machine intelligence manifested here. I call it “co-productive intelligence”, a concept explored in my recent book Cybernetics and the Constructed Environment. This concept helps analyze a popular question that haunts everyday designers: will AI replace us human designers?

Cybernetics and the Constructed Environment by Zihao Zhang

Many have offered insights into this question, ranging from pessimistic to optimistic perspectives. My view essentially argues that this AI-human rivalry is a false dichotomy in the first place; only when we bypass this constructed myth can we ask better questions regarding AI and design. The question “Will AI replace humans?” is based on the premise that humans are the measure of intelligence and AI is essentially about mimicking human mental capacities. Indeed, the famous Turing test in his 1950 paper “Computing Machinery and Intelligence” essentially set up how AI has been discussed for the past 70 years—human is the measure of intelligence, and the goal of AI research is thus to mimic human intelligence to a point the behaviours are not distinguishable anymore. This form of behaviourism ignores the intrinsic difference between an artificial neural network—how most AI systems are made of these days—and a person.

Here, I need to introduce a touch of posthumanism. We have long used humans as the standard for measuring other forms of intelligence; for example, we might say a dog is equivalent to a three-year-old child. However, dogs would surpass us if intelligence were defined and measured by olfactory ability. An analogy can be drawn to think about artificial neural networks—a network of mathematical functions (neurons). Without getting into the details of how AI systems work, the point here is that human intelligence and machine intelligence are not comparable. Maybe they can be compared in terms of the effectiveness and speed in achieving certain goals; however, this does not mean we should use this comparison to reduce humans to a collection of utility functions for achieving goals. Isn’t it sad to imagine a perfect human as the golden standard to measure intelligence and then use that standard to see the world around us? Doing so misses the possibility of seeing other forms of intelligence that would surprise us. Finally, let’s not forget the danger of this form of human exceptionalism and its problematic consequences, for example, the racist, sexist, and classist claims made based on IQ testing.

Let’s use a thought experiment to bring this logic closer to the design professions. In comparing AI to a designer, a capitalist firm owner might immediately consider the idea of a “perfect employee” who can work long hours on endless iterations; then, the owner will invest in developing AI algorithms to replace those who cannot meet the standard. If we agree that this approach is inhumane, then we need to embrace a design culture that promotes human-machine collaborative and co-productive intelligence, viewing them as complementary rather than comparable and thus replaceable. When iteration becomes easy and inexpensive with AI tools, perhaps it’s time to rethink the “iteration culture” in design practice—a strategy now often used to justify hours billed to clients by showcasing the volume of work. Instead, the advancement of AI give us a chance to revalue subjective intuition and judgment as essential qualities in designers.

Finally, I will critique and caution against the proliferation of the picturesque in today’s AI boom, as my co-author, Shurui Zhang, and I have argued in Representing Landscapes: Visualizing Climate Action (Ed. Nadia Amoroso). After nearly 30 years of theoretical development and experimentation, contemporary landscape architecture is poised to move beyond “designing beautiful scenes” toward “designing processes and frameworks” that encourage landscape emergence and evolution. Projects like Freshkills Park and Living Breakwaters exemplify this paradigm, focusing on long-term planning and continued adaptation through learning by doing. This shift is essential for addressing complex issues like the climate crisis, uniquely positioning landscape architecture to lead in climate adaptation efforts.

However, generative AI made image-making easier, and the proliferation of images exacerbated the already problematic reductionist approach to contemporary landscape design—returning the imagination of landscapes into two-dimensional, mono-sensorial visual representations of picturesque scenes to be constructed and measured against. Picturesque becomes anathema to contemporary landscape architecture theory and practice that see landscapes as ever-changing dynamics of ecosystems emerging and re-emerging over time. A picturesque veneer freezes time, reduces landscape processes to visual tactics, and denies possibilities of interpretation. While we celebrate this AI boom, shouldn’t we also worry about the theoretical backlash that returns the profession to “landscape painters”?