A prompt is a drawing made of words. It carries the same intentions as a sketch — light, mass, material, mood — but speaks them in a language the machine reads. This guide teaches the grammar of that language, the rhythm of its phrases, and the discipline that turns description into design.
Before you write a single token, understand how the model interprets language. The model is not a search engine, not an architect, and not a translator. It is a pattern-matching engine that has seen millions of labelled images.
Each word you write activates a cluster of visual associations. Write "brutalist", and the model summons concrete, raw form-work, monolithic mass, deep shadows, often grey weather. Write "villa", and it conjures terraces, pools, warm light, sloped roofs. Your prompt is less a sentence and more a chord — multiple notes struck at once.
Think of yourself less as a writer and more as a director of a film crew that cannot ask clarifying questions. Every assumption you leave implicit, the model will fill in — usually with the most generic version it knows.
A strong architectural prompt has a skeleton. Once you internalise it, you can write prompts the way a draftsman lays out a sheet — by ordered convention, with room for invention.
Use this seven-part formula as your default. You can drop or rearrange parts, but if a prompt feels weak, check which slot is missing.
Write one prompt with all seven slots, then strip out one part at a time. Render each version. You will see exactly how much each ingredient contributes.
The model recognises building typologies far better than abstract descriptions. "Single-family home" is weaker than "ranch house" or "Edwardian terrace". Reach for the specific noun.
After typology, give the model a hint about form. These words carry enormous weight in architecture.
Style words are the most efficient ingredient in any architectural prompt. They carry geometry, material, and proportion in a single token. But they must be used with care.
Architect names are powerful but slippery. The model has stronger associations with some than others, and outputs tend to flatten everyone into a single visual stereotype.
Materials are where most prompts fall short. "Wooden" is vague — the model averages it across plywood, oak, cedar, and IKEA. Be specific: "weathered cedar cladding" tells a story.
Surface comes alive when you describe what time and weather have done to it.
Architectural images live or die by their light. The same building at golden hour and at noon are two different buildings. Treat light as a first-class part of the prompt, not an afterthought.
The phrase "long shadows" implicitly tells the model "low-angle sun", which means dawn or dusk, which means warmth, which means atmosphere. One phrase, four implications.
Telling the model what kind of photograph you want is half the job. The vocabulary of architectural photography is well represented in training data — use it.
You can ask for photographs, but you can also ask for other media. These produce dramatically different results.
A building without a site is a model on a turntable. Tell the model where the building stands — geographically, climatically, culturally. The context will pull the design language with it.
ComfyUI inherits a small grammar of attention-shifting tools. Master these and you will spend less time fighting the model and more time directing it.
Most ComfyUI text encoders (SD 1.5, SDXL with A1111-style parsing) accept emphasis syntax. Use it to push or pull a concept.
In some samplers, the keyword BREAK resets the attention chunk, useful when you want two themes not to bleed into each other (e.g., describing exterior, then a separate clause for foreground figures).
Advanced users can schedule prompts over diffusion steps. The syntax [A:B:0.5] tells the model to start with A and switch to B halfway. Useful for combining structural language early with material detail later.
A negative prompt is your editor's red pencil. It does not direct the model where to go; it tells the model where not to go. Used well, it cleans your image. Used badly, it adds nothing.
Do not over-stuff the negative prompt. Each token still consumes attention budget. 12–20 terms is plenty for most jobs. If you have written 50 negatives, the problem is probably in the positive prompt.
Flux-family models often work better with no negative prompt at all, or with very minimal negatives. If using Flux, default to leaving it empty and add only when needed.
Copy these, modify them, and make them your own. Each is built on the seven-part formula and shows how the same skeleton produces wildly different buildings depending on the words chosen.
Nobody writes a perfect prompt on the first try. The discipline is not in writing — it is in revising. Here's a workflow that compresses ten random attempts into three intentional ones.
Change one variable at a time. If you swap the architect, the materials, and the time of day all in one go, you cannot learn which change did what. Treat the prompt as a controlled experiment.
The best prompts read almost like a passage from a good architectural monograph — they describe a building so vividly that the reader can already see it before the model does. Write toward that ideal, and the renders will follow.
Now that you have the theory, use the visual generator to build prompts faster — pick from dropdowns, see all three model formats side by side, save your favourites.