AI Tools I Tested and Rejected — And Why

Over the years of working in landscape design, and then developing formspace.design, I tried many AI tools, hoping to find effective solutions for design. My experience showed that, despite their impressive capabilities, most of them proved useless or even harmful for the specific tasks of landscape design. The reason is simple: they don't understand the unique requirements, context, and subtleties of our field. I will explain what types of AI I tested and why I ultimately abandoned them, focusing instead on creating a specialized tool.

My Journey to AI in Landscape Design

My interest in artificial intelligence is no accident. After decades of managing the company "Home and Garden," where we designed and built cottages, landscaped, and maintained plots, I faced the need to automate and optimize processes. The issue of speed and accuracy during the conceptual design phase was particularly acute. When, at 61, I began to learn programming myself with the help of AI, it opened new horizons for me. I saw potential in generating ideas, quick visualization, and assistance with routine tasks, but quickly realized that general solutions didn't work. My path to formspace.design was dictated by frustration with universal tools and a desire to create something I personally lacked in practice.

Early Experiments and Disappointments

I started with the most popular and accessible AI tools: text-to-image generators, various CAD assistants, and even attempts to use AI to generate 3D models from 2D plans. Initially, it seemed like a breakthrough. You could get a "beautiful picture" of a plot in a matter of minutes by simply describing the desired style or elements. However, the euphoria quickly turned into disappointment. These tools, whether Midjourney, Stable Diffusion, or others, lacked an understanding of landscape principles; they didn't know about insolation zones, terrain features, drainage, or plant compatibility. They could produce a spectacular but completely unfeasible project, for example, suggesting planting palm trees in a temperate climate or placing a pond on a steep slope without considering engineering solutions.

The Problem of Accuracy and Specificity

The main reason for my rejection of many AI tools is the lack of accuracy and understanding of the specifics of landscape design. A successful project requires considering dozens of factors: from soil composition and groundwater levels to wind patterns and neighboring buildings. General AIs don't understand this. They cannot adequately assess that a 60 cm wide path will be inconvenient for a wheelbarrow or lawnmower, or that a 30 cm distance between bushes will lead to them becoming leggy and diseased in a couple of years. I needed a tool that doesn't just draw, but designs, taking into account building codes, biological characteristics of plants, and functional zones of the plot. This is the very "expertise" that AI cannot provide without specialized training. I wrote about how I try to teach AI to understand these nuances in the article "How I Correct AI When It Doesn't Understand a Landscape Task".

Lack of Context and Data

Another big problem is the lack of context and quality specialized data. Many AI models are trained on vast arrays of general images and texts from the internet. But what is a "beautiful garden" for AI? It can be anything from a Japanese Zen garden to an English landscape park, without reference to a specific 10-acre plot, its budget, or the client's wishes. AI doesn't know that in a certain region, winter temperatures drop to -25°C, and therefore the exotic plants it suggests simply won't survive. It doesn't see that the sunny side of the house is ideal for a fruit garden, while the shady side is for a relaxation area with shade-tolerant plants. Such AI "hallucinations," where it produces unrealistic or impractical solutions, became a frequent occurrence for me.

Complexity of Integration and Workflow

Even if an AI tool showed potential, the next problem arose: how to integrate it into the existing workflow? Most of them worked as separate "black boxes": you uploaded a request, you got a result. But for a landscape project, this doesn't work. You need the ability for iterations, corrections, making changes at each stage, and interaction with other programs (CAD, 3D modeling). Attempts to link several disparate AI services into a single chain proved too laborious and inefficient. It's like assembling a car from parts from different manufacturers that don't fit together without significant modification. The working time spent on "tweaking" AI results often exceeded the time needed to create a project manually.

Cost and Return on Investment

One cannot forget the financial side. Many advanced AI tools work on a subscription basis or require payment for each generation. While this seems acceptable at first, with active testing and numerous iterations, costs quickly rise, while the actual benefit remained questionable. For an individual landscape designer or a small studio, this can become a significant expense that doesn't pay off due to the low quality or irrelevance of the solutions provided. I was looking for a tool that would not just be a "toy," but would genuinely reduce work time and improve project quality, thereby increasing profitability. I thought a lot about this, by the way, when I pondered "What AI Can Assemble Itself, and What Still Requires a Developer".

How to Plan in formspace.design

These very frustrations became the impetus for creating formspace.design. I realized that what was needed was not just AI, but specialized AI that understands landscape logic, is trained on relevant data, and is integrated into a convenient workflow. My product formspace.design is designed to solve precisely these problems.

Imagine a scenario: you upload a client's plot plan into Module 1.x "Plan/Zones/Objects." Here, you can precisely define boundaries, mark existing structures and utilities. Then, using Module 2.1 "AI Concept," you input the client's wishes regarding style, functional zones (e.g., a 3x4 m BBQ area, a 5x6 m playground), plant preferences, and budget. formspace.design's AI, trained on thousands of real landscape projects and dendrological plans, will suggest several concept options, taking into account zoning, insolation, terrain, and even plant compatibility for your region. You get not just pictures, but functional schemes that can be refined. For example, you can quickly rearrange flowerbeds or change the shape of paths, ensuring that the minimum width of the main path is 1.2 meters and the secondary path is 0.8 meters, as required by convenience standards. Module 3.1 "Design by Photo" allows you to visualize changes directly on a real photo of the plot, which significantly simplifies communication with the client and speeds up project approval.

FAQ

What types of AI tools did you test first?

First and foremost, I tested universal generative models, such as text-to-image AI (e.g., Midjourney, Stable Diffusion), which create images from textual descriptions. I also tried various online planners with AI functions promising automatic object placement, and some experimental systems for converting 2D plans into 3D models. These tools were attractive due to their apparent simplicity and speed.

Why are general AI tools unsuitable for landscape design?

General AI tools do not understand the specifics of landscape design because they are not trained on a sufficient volume of specialized data that takes into account plant biology, engineering aspects, climatic conditions, zoning norms, and rules. They can generate beautiful but absolutely non-functional, unfeasible, or even dangerous solutions, such as placing sun-loving plants in the shade or creating ponds without considering drainage. The lack of context and deep understanding of the subject matter makes them useless for serious design.

Were there any AI tools that almost fit but still didn't meet expectations?

Some AI tools for 3D modeling and visualization showed good potential in terms of rendering, but their integration with real landscape data and the possibility of iterative refinement were extremely limited. They could quickly create an impressive image but did not allow for easy changes to the layout, adaptation to specific site conditions, or consideration of plant biological characteristics. This required manual intervention and rework, which negated all the speed advantages.

How did you determine that a tool was unsuitable for your tasks?

I determined a tool's unsuitability based on several criteria: lack of accuracy and relevance of the proposed solutions, inability to account for specific site conditions (climate, soil, terrain), difficulty of integration into my workflow, and high time costs for refining "raw" results. If the AI generated concepts requiring radical rework or containing errors that could not be quickly corrected, I understood that it was not what I needed.

What specific AI errors did you encounter most often?

Most often, I encountered errors such as: suggesting plants incompatible with the climatic zone or soil type; ignoring functional zoning and logistics of movement around the site; creating unrealistic terrain solutions without considering drainage; failure to observe basic ergonomic dimensions (e.g., too narrow paths or small areas); and "hallucinations" – generating objects or elements that made no sense in the context of landscape design.

Should a landscape designer try AI tools at all?

Absolutely, it's worth it, but with an understanding of their limitations and purpose. General AIs can be useful for inspiration, brainstorming, or creating initial, very general sketches. However, for professional design, specialized tools are necessary that take into account all the nuances of the landscape. It's important to look for AI that doesn't try to replace the designer but complements their expertise and automates routine tasks, providing accurate and feasible solutions. This is precisely the approach I implement in formspace.design.

Does formspace.design's approach differ from other AI tools you tested?

Yes, it fundamentally differs. formspace.design is initially created as a specialized tool for landscape design. Its AI models are trained on relevant data, taking into account biology, engineering, zoning, and climatic features. We do not strive for universality but focus on a deep understanding of one area. This allows us to offer not just beautiful pictures, but functional, implementable concepts that can be easily corrected and refined within a single platform, significantly saving the designer's time and resources.

A Look into the Future

My experience has shown that AI is a powerful tool, but its effectiveness directly depends on specialization and understanding of the subject area. General solutions do not work for complex and multifaceted tasks like landscape design. That is why I focused my efforts on creating formspace.design — a platform designed to become a truly useful assistant for everyone involved in site planning. I believe that the future lies in specialized AIs that do not just generate, but understand and help create unique and functional landscapes. Try formspace.design and see for yourself how AI can work for you, not against you.