Midjourney for Architecture

Midjourney is a hosted, closed diffusion model - you cannot see its weights, run it offline, or inspect how it was trained. What you can see is that its makers tuned it relentlessly toward one thing: images people find beautiful. Where an open model gives you a neutral, literal reading of your prompt, Midjourney applies a strong built-in aesthetic - flattering light, confident composition, cinematic depth - before you ask for any of it.

For architecture this is a genuine gift at the concept stage. A half-formed idea - a stepped riverside library in weathered brick, monsoon sky - comes back looking like a competition-winning render in thirty seconds. You are not documenting a design; you are fishing for mood, massing feelings and material atmosphere faster than any other tool allows. The cost of that gift is that the beauty is partly the model's, not yours: it will make almost anything look resolved, including ideas that are structurally nonsense. Treat its polish as a mood, not a verdict.

This is also why Midjourney rewards a different working posture than a general model. With an open tool you often fight to make the image look good; with Midjourney you fight to make it look like your idea specifically rather than a generically gorgeous building. The effort moves from 'add beauty' to 'add intent'. Understanding that inversion up front stops you wasting words on adjectives the model does not need and spending them instead on the specifics only you know.

Midjourney is driven by plain text plus a handful of parameters appended with a double dash. Four are worth learning cold.

--ar (aspect ratio) sets the frame. Architecture wants deliberate framing: --ar 16:9 or --ar 3:2 for a cinematic exterior, --ar 2:3 for a tall facade or a stairwell, --ar 1:1 for a material study. The ratio quietly changes the composition the model reaches for, so choose it first.

--stylize (roughly 0 to 1000) sets how hard Midjourney applies its own taste over your words. Low stylize obeys you more literally and is what you want when a client brief is specific; high stylize is prettier but drifts from your instructions. --chaos (0 to 100) sets how different the four returned options are from each other: low chaos gives four near-siblings for refining one idea, high chaos gives four divergent directions when you are still exploring.

Two more are worth knowing. --no is Midjourney's version of a negative prompt - --no people, text, watermark tells it what to leave out, useful for keeping entourage or captions out of a clean architectural frame. And the --seed value fixes the random starting point, exactly as you learned in Module 0: lock it and a prompt returns a near-identical image, so you can change one word and watch only that aspect move rather than getting a whole new picture. Midjourney also offers version selection - newer model versions render architecture and materials more convincingly - so knowing which version you are on matters when you compare results over time.

The discipline is to move one parameter at a time. If you crank stylize, chaos and aspect ratio together you learn nothing about which lever did what. Change one, read the four results, change the next - the same one-variable method you met in Module 1, now with tool-specific knobs. A designer who knows five parameters cold and moves them deliberately will get more from Midjourney than one who writes elaborate prose and leaves every setting at its default.

The hardest problem in AI imagery for a real project is consistency: making ten views that clearly belong to the same building or the same interior scheme. Midjourney's two reference parameters are its best answer.

--sref (style reference) takes an image URL and carries its look - palette, texture, quality of light, mood - into new generations without copying its subject. Point it at your own approved concept image and every new frame inherits that atmosphere, so a board of interiors reads as one coherent scheme rather than ten unrelated pictures. You can also blend several references or lock a numeric style code to reuse a look for weeks.

--cref (character reference) does the same for a recurring subject - originally a person, but useful for holding a distinctive motif or figure steady across scenes. Neither is true geometric control; you still cannot force an exact plan or elevation, and that is what Module 3's ControlNet exists for. But for the everyday job of keeping a project's images visually of-a-piece, style references are the single most useful habit Midjourney gives you.

Two limits follow from Midjourney being a closed, commercial product, and both matter for practice.

First, control. You cannot install it, cannot run ControlNet or custom LoRAs inside it, cannot force it to obey your line drawing. Its precise conditioning options are far thinner than the open Stable Diffusion world you will meet next lesson. If your task is explore and communicate, that is fine. If your task is make the AI reproduce my actual geometry, Midjourney is the wrong tool and no amount of prompt-craft fixes it.

Second, rights. Because it is a subscription service, what you may do with the images is governed by Midjourney's terms, not by an open licence. Broadly, paid subscribers own the assets they create subject to the terms, but the specifics - and the treatment of images that echo living artists or trademarks - change and deserve a read before you put an image in a paid deliverable. Add to this the unsettled, jurisdiction-dependent question of whether AI images attract copyright at all, and the honest professional stance is: check the current terms, disclose AI use to clients, and never assume an image is cleanly yours because a tool produced it. We treat rights properly in Module 9; here the rule is simply to look before you bill.

Because Midjourney flatters everything, the danger is not making a bad image - it is being seduced by a good-looking one that says nothing true about your project. A small set of habits keeps it honest.

Prompt for substance, not adjectives. The model already supplies beauty; your words should supply specifics it cannot guess - the programme (a riverside library, not 'a building'), the material (weathered brick, board-formed concrete), the site condition (monsoon light, west-facing), the vantage (worm's-eye, across the water). Piling on 'stunning, award-winning, masterpiece' adds nothing the aesthetic engine was not already doing; naming a real material or a real light shifts the image somewhere useful.

Generate in fours, read like a critic. Every prompt returns a grid of four. Treat it as four proposals, not one answer: note what all four share (that is the strong signal in your words) and where they diverge (that is where the model is guessing and you must decide). Upscale the one closest to your intent, not merely the prettiest.

Refine with variations and one-word edits, seed held. Once a frame is close, use the model's variation tools and change a single term at a time - swap 'brick' for 'lime plaster', 'evening' for 'overcast' - so you learn what moved. This is the Module 1 iteration method wearing Midjourney's clothes.

Separate the mood from the decision. When the board is done, write one line per image on what it is evidence for - a material direction, a massing feeling, an entrance idea - and one line on what it is not evidence for. That discipline turns a pretty set into design thinking, and it is the difference between using Midjourney and being used by it.