What Is Generative Architecture? Generative Design in Architecture Explained

Generative architecture is architecture designed with generative design: the architect sets the goals, constraints and parameters of a project, and software generates many design options, scores each one against those goals, evolves the best ones and ranks the results for the architect to explore and choose from. Autodesk's Generative Design Primer calls it "a collaborative design process between humans and computers".

The words are also used in two other ways. "Generative AI in architecture" usually means image and text models that make pictures from a prompt or a sketch: they generate options too, but they do not score them, and the comparison below shows the difference. "Generative AI architecture" can also mean how AI software systems are built, a computing topic this guide does not cover.

Below: how the generate, evaluate, evolve and explore loop works, generative design vs parametric design vs generative AI, the tools architects use as of October 2026, documented projects, where the method does not help, and how a chosen option becomes images for a client.

How Generative Design Works

Generative design turns a design question into a search. Instead of drawing one solution and testing it, you describe what a good solution has to achieve and let the computer try many. Autodesk's primer splits the process into four stages:

1. Generate: the computer creates design options by changing the parameters of a model within the limits you set.

2. Evaluate: each option is measured against the goals, such as daylight, views, usable area or walking distance, by evaluators that return a score.

3. Evolve: the scores steer the next round, so new options build on the better earlier ones.

4. Explore: the architect looks through the ranked results, compares the trade-offs and picks the options worth developing.

Goals, Constraints and Parameters

Every study needs three inputs. Goals are what you want more or less of, written so they can be measured: daylight, views to the outside, net-to-gross area, structural weight, the distance between teams that work together. Constraints are the limits every option must respect: the site boundary, height limits, setbacks, the building code. Parameters are the variables the computer may change: orientation, the number of floors, the unit mix, the position of a core. Underneath sits a parametric model in which every combination of parameters still produces a valid design, and building that model is most of the work. Goals about movement and views draw on the methods of spatial analysis.

How the Computer Evolves Options

The most common search method is the genetic algorithm. A population of options is generated and scored by a fitness function that stands for the goals. The stronger options are kept, their parameters are mixed (crossover) and slightly changed (mutation), and the next generation is scored again. When the scores stop improving, the run ends.

Say the brief asks for more usable area in an apartment building. The algorithm can test core positions, corridor widths and unit ratios on the parametric model and keep the options that raise the net-to-gross ratio without breaking the constraints.

Several Goals at Once: the Pareto Set

Real briefs have goals that pull against each other: more glass brings more daylight and more heat gain. A multi-objective search therefore does not return one winner. It returns a Pareto set, the group of options in which no goal can be improved without making another one worse. The architect chooses among them, and that is where judgement, the client's priorities and everything a score cannot measure come in.

Generative Design vs Parametric Design vs Generative AI

The three terms overlap, but they name different ways of working:

Parametric design is the base layer: generative design runs on a parametric model and adds the goals and the search. Our guide to parametric design covers it with examples. Procedural modeling, which grows geometry from rules such as shape grammars or L-systems, is another way to produce the families of options a search can score.

What Is Generative AI in Architecture?

Generative AI in architecture means models trained on large sets of images or text that produce new ones on request. Diffusion models start from noise and refine it step by step into an image that matches a prompt; when the input is an image, they start from a noised version of it. GANs, an older family, train a generator against a discriminator; see GANs in architecture. A review of the research by Li and co-authors (arXiv 2404.01335, revised October 2024) covers GANs, VAEs and diffusion models and finds that most published work is on plan design. For the wider picture, see how AI is used in architecture.

These models produce many options quickly, but they do not measure daylight, area or cost, so a picture can look right and still miss the brief. ArchiVinci, an AI architecture app, sits on this side of the table: it generates images from a sketch, a model view or a sentence, and it does not generate or score design options.

Generative Design Tools for Architects

The tools below generate and evaluate options. The facts come from each vendor's own pages, read in October 2026; prices are left out because they depend on the plan and the region.

Wallacei was developed by Mohammed Makki, Milad Showkatbakhsh and Yutao Song from Makki's doctoral research at the Architectural Association. The Revit tool runs the same generate, evaluate, evolve and explore stages as the primer. For questions at the scale of a city, see AI in urban planning.

Layouts Generated from a Short Brief

Not every generator scores its options; some work from rules alone. ArchiVinci's Floor Plan Maker has a "Generate automatically" option: you fill in a short brief (property type, total area, floors, bedrooms, bathrooms, kitchens, living rooms and whether the kitchen is open-plan) and it draws a complete furnished layout. It is rule-based, runs without AI and costs nothing, and "Try another layout" gives a new variation. It does not score or optimise the layouts, so you are the evaluator: keep the one that works, then move walls and rooms with wall lengths and furniture dimensions shown as you edit, like in a CAD program, check the room areas, look at it in 3D and export DXF, PDF or GLB. It runs in desktop browsers.

The Floor Plan Maker editor, where a layout is edited after it is drawn or generated (screenshot from the Floor Plan Maker page).

Generative Architecture Examples

Autodesk's Office at MaRS in Toronto

The best documented project is Autodesk's own office and research space in the MaRS Innovation District in Toronto, designed with The Living. The designers built a geometric system of work neighborhoods, amenity spaces and circulation zones, and the computer generated, evaluated and evolved thousands of design alternatives. Six measurable goals were defined from what employees and managers said about their work styles and location preferences: work-style preference, adjacency preference, low distraction, interconnectivity, daylight and views to the outside. Evaluators scored each plan against them, and the team explored the results to pick the design that best met the goals overall. The case is described step by step in Autodesk's primer.

Site Studies in Autodesk Forma

Autodesk describes Forma's Site Automation as a generative design feature for site planning: it generates building and circulation layouts within a site boundary and suggests an initial massing from your inputs, such as a grid layout with setback distances and building types like city blocks, line buildings and towers. The options are then checked with Forma's rapid analyses, for example wind and noise.

Where the Idea Comes From

Designing by evolution is older than today's software. An early book on it is John Frazer's An Evolutionary Architecture, published by the Architectural Association in London in 1995.

Generative AI Images: Many Options, No Score

For comparison, here is the image side of the table. The three pictures below come from the same upload, each run in ArchiVinci's Exterior module with a one-sentence style prompt. They show how quickly generative AI gives many looks, and also what it does not do: none of them was scored for daylight, area or cost.

Craftsman: "Craftsman house: low-pitched gables, tapered porch columns on stone piers, wood siding in sage green."

Spanish Colonial Revival: "Spanish Colonial Revival house: white stucco walls, red clay tile roof, arched windows, wrought-iron details."

Modern farmhouse: "Modern farmhouse: white board-and-batten siding, black window frames, metal roof, covered front porch."

From a Chosen Option to Images for the Client

A generative study ends with an option chosen on paper: a massing, a site layout or a facade. The client then wants to see it. Export or screenshot the view of that option from Rhino, Revit or Forma (JPG, PNG or WebP, up to 5 MB) and upload it to ArchiVinci's Exterior module, described under AI exterior design. Describe the materials and the light in a sentence, or pick a lighting button such as Golden hour, and you get a new photorealistic image that follows your view, in seconds. Every run draws a new image, so you can compare facade materials or moods side by side. The default engine, ArchiVinci 4.0, keeps the perspective and the geometry of your view, so every version shows the chosen option. Using the images in client work or marketing needs an active plan.

A site plan or a massing model of several buildings goes to the Masterplan module, described under masterplan design. With no view yet, the Imagine module turns a typed description into concept images; see the AI concept generator.

Left: the massing view as uploaded. Right: the result from the Exterior module with the prompt left empty and no Lighting button picked.

When Generative Design Does Not Help

Generative design pays off when a question has many possible answers and goals you can measure. It helps less in these cases:

• A simple brief: a small house on an open plot has few real alternatives, and sketching two or three by hand is faster than building a parametric model.

• Goals you cannot score: beauty, character and fit with a street are judged by people; a search can only improve what an evaluator measures.

• A tight deadline: setting up the model, the constraints and the evaluators takes time before the first useful result appears.

• Weak inputs: if the goals or the analysis behind a score are wrong, the best-scored option is wrong too.

Every result is a starting option that architects and engineers still develop, check against the code and coordinate. The tools change the work more than the role, as our article on whether AI will replace architects discusses.

Generative Design and Sustainability

Energy and daylight goals suit generative design well because they can be measured early. Orientation, massing, window-to-wall ratio and shading depth can all be parameters, and an energy or daylight analysis can be the evaluator, so the trade-offs show up before the building systems are sized. How much a project gains depends on the site, the brief and the quality of the analysis, so treat any fixed percentage you read with care.

Frequently Asked Questions

Is generative design AI?

Not necessarily. Classic generative design uses search and optimisation, such as genetic algorithms, on a parametric model, and the goals and evaluators are written by people. Some newer platforms add machine learning, and generative AI image models are a different technology altogether.

Is generative AI the same as generative design?

No. Generative design generates options and scores them against measurable goals. Generative AI creates new images or text from patterns in its training data and scores nothing. Architects can use both: generative design to choose a massing or a layout, generative AI to show the chosen option.

What software is used for generative design in architecture?

As of October 2026: Grasshopper for Rhino with the Galapagos solver or the free Wallacei plugin, Generative Design in Revit with Dynamo (AEC Collection, from Revit 2021), the Site Automation feature of Autodesk Forma, and platforms such as TestFit and Finch. The tools table above says what each one does.

Who started generative architecture?

No single person. An early book on the idea is John Frazer's An Evolutionary Architecture, published by the Architectural Association in London in 1995. Today the same evolutionary approach runs in tools such as Galapagos inside Grasshopper.

Can generative design produce buildable designs?

Yes, when the constraints encode the real limits: the site boundary, height, setbacks, structure, the code and the budget. Even then the output is a starting option that architects and engineers develop, check and coordinate like any other design.

How do parametric design and procedural modeling work together?

Parametric design links dimensions and parts so that one change updates the whole model; procedural modeling grows geometry from rules such as shape grammars or L-systems. Both produce the families of options that a generative search then scores.

What do diffusion models add to early design?

They turn a short prompt, a sketch or a reference image into pictures quickly, which helps with moods, materials and first ideas. They do not measure performance, so the pictures say nothing about daylight, area or cost; that stays the job of the analysis and the architect.

Try it in ArchiVinci

Describe a building in a sentence, or upload a sketch or a 3D model view: ArchiVinci generates concept images that follow it. A new account starts with 3 free credits.

• From a sketch: Turn a sketch into a render with AI