The AI architecture space in 2026 has its own vocabulary — part machine learning, part image generation, part traditional architectural practice — and the overlap between these three worlds causes most of the confusion professionals encounter when evaluating tools. This glossary defines the 50 terms that come up most often in tool documentation, product marketing, and professional conversation, grouped by category.
Each term is defined in one or two sentences, with a note on how it’s used in practice where relevant.
Generation and Core Capabilities
1. Text-to-image
The core capability of most generative AI image tools: producing an image from a written description. In architecture, the baseline feature for any concept-generation tool.
2. Image-to-image
Generating a new image using an existing image as a visual anchor, often with a text instruction to guide what changes. The workflow behind most iterative editing in AI architecture tools.
3. Text-to-design
A narrower category than text-to-image: producing architectural design outputs — exteriors, plans, interiors — from a written brief. Implies that the tool understands architectural conventions, not just any image.
4. Text-to-plan
Generating a floor plan from a written description — typology, area, room count, adjacencies. Some tools also accept structured parametric input alongside text.
5. Sketch-to-render
Turning a hand sketch or a rough 3D model into a finished architectural render. Common in tools like Gendo, mnml.ai, and the render modules of ArchiVinci.
6. Prompt
The written instruction that guides the AI model. Prompt quality is the single largest determinant of output quality in practice.
7. Prompt engineering
The practice of crafting prompts to produce specific, useful results. In architecture, involves balancing typology, style, materials, scale, and atmosphere in a structured way.
8. Negative prompt
Text describing what the image should not contain. Less common in architectural tools than in general image generators but used to exclude unwanted elements (people, cars, specific colors).
9. Seed
A numerical value that determines the initial random state of a generation. Using the same seed with the same prompt produces the same image in most tools — useful for reproducibility.
10. Aspect ratio
The width-to-height proportion of the output image. Architectural visualization commonly uses 3:2, 16:9, or 2:1 for exteriors; square (1:1) for plans and close-up interiors.
AI Model Categories
11. Diffusion model
A type of generative AI model that produces images by gradually refining random noise into a coherent picture. The dominant architecture for image generation in 2026, underlying Stable Diffusion, Midjourney, DALL-E 3, and most professional tools.
12. Transformer
A neural-network architecture originally developed for language but now central to modern image generation when combined with diffusion. Allows models to understand complex relationships between prompt elements.
13. Foundation model
A large, general-purpose AI model trained on broad data, usable across many tasks. Most architecture-specific tools are built on top of foundation models rather than from scratch.
14. Fine-tuning
Additional training of a pre-trained model on a narrower dataset — for example, architectural images — to specialize its output. Some tools fine-tune for architecture; others use prompt engineering on general models.
15. LoRA (Low-Rank Adaptation)
A lightweight fine-tuning method that adapts a model to a specific style, subject, or type of output without retraining the whole model. Common in Stable Diffusion workflows.
16. Checkpoint
A snapshot of a trained model, used as a starting point for generation or further fine-tuning. Stable Diffusion tools give users access to different checkpoints; commercial tools generally don’t expose this.
17. Inference
The process of running a trained model to produce output. “Inference cost” refers to the compute cost per generation, which tool pricing reflects.
18. Token
A chunk of text that the model processes — typically a word or piece of a word. Prompt length limits are usually expressed in tokens.
Conditioning and Control
19. ControlNet
A family of neural network add-ons for Stable Diffusion that guide generation using additional inputs like depth maps, edge detection, or pose information. Used heavily in sketch-to-render workflows.
20. Reference image
An image provided alongside a text prompt to guide the style, composition, or subject of the output. Different tools use reference images differently — Midjourney’s --sref is a widely-known example.
21. Inpainting
Editing a specific region of an image while leaving the rest unchanged. The precise selection-plus-prompt workflow behind targeted edits like changing a facade material.
22. Outpainting
Extending an image beyond its original boundaries — generating what would continue outside the frame. Used to expand exterior views or composition-adjust renders.
23. Mask
A selection of part of an image that an inpainting or editing operation should affect. Often drawn by the user with a brush tool.
24. Style transfer
Applying the visual style of one image to the content of another. The precursor to modern reference-image workflows; less common as a standalone feature in 2026.
25. Guidance scale
A parameter controlling how closely the output follows the prompt. Higher values produce images more literal to the prompt; lower values give the model more freedom.
Architectural Concepts in AI Tools
26. Massing
The overall volume and shape of a building, without surface detail. AI tools produce recognizable massing studies but don’t validate that the massing is structurally feasible.
27. Elevation
A two-dimensional drawing of one face of a building. AI tools can produce elevation-style images but don’t strictly distinguish elevations from perspective views unless prompted carefully.
28. Plan (floor plan)
A top-down view of a building at a specific level. AI-generated plans are schematic — they communicate program and adjacencies but lack construction-document precision.
29. Section
A cut through a building showing heights, levels, and interior relationships. AI tools are weak at sections — they rarely produce usable ones without specialized prompting or model support.
30. Axonometric
A 3D drawing projected without perspective distortion, useful for showing massing and spatial relationships. AI tools produce axonometric-style images on request but with variable accuracy.
31. Program
The list of spaces and functions required in a building — rooms, sizes, adjacencies. The program is the input that drives plan generation.
32. Typology
The building type — villa, hotel, office, restaurant, multifamily. A core element of any architectural brief and the first parameter AI tools use to shape output.
33. Schematic design
The first detailed design phase after the concept, where the building’s organization and major design decisions become clearer. AI tools operate at or just before this phase.
34. Concept phase
The earliest design phase, where direction and intent are established before detailed design begins. This is the phase AI tools have transformed most.
35. FF&E
Furniture, fixtures, and equipment — the movable contents of a space. AI renders typically show FF&E but don’t specify products; specification remains a human task using supplier databases.
Workflow and Practice Terms
36. Moodboard
A visual reference collection used to convey style and atmosphere before design work begins. AI tools increasingly replace manual moodboarding with generated mood images.
37. Render
A produced image of a designed space, traditionally from a 3D model. AI tools generate renders without an underlying 3D model, blurring the distinction between visualization and illustration.
38. Photoreal render
A rendering that looks like a photograph of the finished space. Traditional photoreal rendering is labor-intensive; AI tools produce images that range from stylized to near-photoreal depending on the prompt and tool.
39. Hero image
A single, high-quality image used as the primary visual representation of a project — for marketing, pitch decks, website covers. Often commissioned separately at higher quality than the project’s other visuals.
40. Presentation deck
A set of slides or images assembled for client, investor, or stakeholder review. The deliverable that the concept phase aims to produce.
41. Design brief
The written statement of a project’s requirements, style direction, and constraints. The quality of the brief determines the quality of AI output.
42. Iteration
Producing multiple versions of a design to explore options or refine toward a final direction. AI dramatically reduces the cost of iteration, changing how many directions can be explored.
43. Branching (in AI workflows)
Generating a new design option from an existing one, preserving its context while allowing changes. The mechanism behind maintaining project coherence across variations.
44. Project context
The accumulated state of a project in an AI tool — previously approved images, style choices, material decisions — that informs subsequent generations. The structural feature that distinguishes concept-design tools from general image generators.
45. Variant
An alternative version of a design, usually differing in one dimension (palette, material, massing move). AI makes variant generation nearly free, changing how variants are used in client conversation.
Specific Tools and Model Names
46. Midjourney
A leading general-purpose image generator known for single-image aesthetic quality. Widely used in architecture for hero and mood images.
47. Nuit
A concept-design tool built around architectural project context, producing exteriors, plans, and interiors with shared visual identity across a single project.
48. Nano Banana
An image-editing AI tool that architects and designers favor for precise image-to-image edits. Widely used to refine selected renders without losing composition.
49. Gendo, mnml.ai, ArchiVinci, Maket, InteriorAI, Rendair AI
A family of architecture-specific AI tools each with distinctive specializations — sketch-to-render, modular workflows, parametric plans, room restyling, and visualization. Part of the active 2026 toolscape alongside Nuit and general image generators.
50. Stable Diffusion, DALL-E
General-purpose image-generation models used in architecture primarily for exploration and one-off images. Stable Diffusion is openly available and customizable; DALL-E is offered by OpenAI as part of broader AI products.
Which additional terms are worth knowing?
Some terms come up in adjacent conversations but don’t strictly belong to the AI architecture vocabulary:
- CAD (computer-aided drafting): the traditional software category for architectural documentation — AutoCAD, VectorWorks, ArchiCAD.
- BIM (building information modeling): the data-rich 3D-modeling category for buildings — Revit, ArchiCAD, Vectorworks BIM features.
- Parametric design: computational design where geometry responds to inputs or rules — Grasshopper, Dynamo. Conceptually related to parametric AI plan tools but distinct in practice.
- Photogrammetry: reconstructing 3D geometry from photographs. Not a generative AI category but increasingly combined with generative tools for real-world site capture.
- Generative design: a term used in traditional CAD/BIM for algorithmically produced geometry options based on constraints. Not the same as “generative AI” in the image sense, though the vocabulary overlaps.
How do these terms show up in tool documentation?
Most documentation falls into three categories in how it uses these terms:
Consumer-friendly tools (Planner 5D, InteriorAI) use minimal jargon. You see “style,” “room,” “generate” rather than “diffusion” or “inference.”
Professional tools (Nuit, ArchiVinci, Maket) mix terms. Architectural vocabulary (typology, program, massing) appears alongside AI-specific terms (prompt, iteration, reference).
Developer-facing tools (Stable Diffusion, Replicate, Hugging Face-hosted models) use the full machine-learning vocabulary — LoRA, checkpoint, ControlNet, inference. Relevant if you’re building your own workflows; less relevant if you’re a working architect.
Knowing which category of documentation you’re reading helps calibrate how much translation work is involved in evaluating a tool.
Related reading
- AI Design Vocabulary: 25 Terms for Clients — As AI tools have become routine in design practice, clients have begun using AI…
- What Is AI Concept Design? Definition — AI concept design is the use of generative AI tools to produce early-stage architectural…
- What End-to-End AI Design Means in 2026 — End-to-end AI design is the workflow where one tool carries a project from brief to…
- AI Architecture Design: The Complete Guide for 2026 — AI architecture design is the use of generative AI tools to produce architectural…
- Best AI Tools for Architectural Concept Design in 2026 — The best AI tools for architectural concept design in 2026 are Nuit, Midjourney,…
- AI for Architecture Students in 2026 — How students should approach AI tools: portfolio, studio projects, and thesis work…
Frequently Asked Questions
What’s the difference between generative AI and traditional CAD?
Generative AI produces design options from a description; traditional CAD draws what you tell it to draw. Generative AI is good at exploration and concept work; traditional CAD is necessary for documentation and construction. They’re complementary, not competing, categories.
Do I need to know the technical AI terms to use these tools?
For using the tools, no. Most professional tools are designed for architects and designers, not machine learning engineers. Knowing terms like “prompt” and “reference image” is enough. The deeper technical vocabulary (LoRA, ControlNet, checkpoint) matters mainly if you’re customizing open-source tools like Stable Diffusion.
What’s the difference between a render and a generation?
A traditional render is produced from a 3D model with lighting, materials, and camera settings computed explicitly. An AI generation is produced from a text prompt or reference image, with no underlying 3D model. The outputs look similar; the process and level of control differ.
Is “AI render” the same as “photoreal render”?
Not necessarily. AI tools produce images on a spectrum from stylized illustration to near-photoreal quality. The degree of realism depends on the tool, the prompt, and the underlying model. A photoreal AI render aims to look like a photograph; a stylized AI render looks more illustrative.
What does “project context” actually mean?
It’s the memory an AI tool keeps about the specific project you’re working on — the approved exterior, the style choices, the material decisions — which the tool uses when generating subsequent images. Tools with strong project context produce coherent sets of images for one project; tools without it produce independent images that don’t necessarily read as the same building.
What’s a foundation model and why does it matter for architecture tools?
A foundation model is a large general-purpose AI model trained on broad data. Most architecture-specific tools build on top of foundation models rather than training from scratch — because training a large model from zero is prohibitively expensive. The foundation model determines much of the quality ceiling of the tool; the architecture-specific layer determines the workflow and project-context features.
Which terms are most important to understand when evaluating AI tools?
Prompt, reference image, iteration, project context, and the distinction between text-to-image and image-to-image. These five cover the essential capabilities that differentiate tools in real use.
Try Nuit free — 10 generations, no card required. See how project context works in practice — generate an exterior, plan, and interior that read as one design. Start your project →