Midjourney Prompts for Architecture (V8 Parameters)

TL;DR: A Midjourney architecture prompt is won or lost in the parameter tail. On V8.2, --raw plus a low --stylize gives buildings instead of illustrations, --ar and --hd set the pixels, and --quality, --cref and --oref are marked unsupported. Thirty templates below, by drawing type.

What makes a Midjourney architecture prompt different from a generic image prompt?

The scene description is the part everyone writes and the part that matters least. What separates a usable archviz image from a glossy nothing is the parameter tail, and that tail is version-specific in a way most published prompt lists have stopped tracking.

Midjourney's Parameter List is blunt about the mechanics: parameters go after the prompt text, with a space before the dashes, and "Don't use commas, periods, or other punctuation marks in your parameters" (Parameter List, updated 11 June 2026, read 28 August 2026). A trailing comma after --ar 16:9 breaks the tail, silently.

The version matters more. Midjourney's Version article states that "The current default Midjourney version is V8.2", which took the slot on 24 July 2026 from V8.1 (Version, updated 24 July 2026, read 28 August 2026). Prompt lists written against V7 carry parameters the current model does not accept. Post 159 covers the version flags, and 30 AI prompts for architectural rendering is the model-agnostic sibling to this page.

Which V8.2 parameters are verified, and which are stranded on an older version?

Everything below was read on Midjourney's help centre on 28 August 2026. The right-hand column kills most copied prompt lists.

--quality is the big one. The compatibility chart shows a no-symbol in the merged V8.1 and V8.2 column for the Quality Parameter row. The Quality article itself, updated 27 July 2026, three days after the chart, discusses only version 7 and never mentions the parameter is gone (Quality). Both pages are live and disagree. Leave --q off, and see post 219 on quality and speed modes.

Character consistency is stranded. The Character Reference article says "Character Reference can be used with Midjourney and Niji version 6." The Omni Reference article says "Omni Reference can only be used with Midjourney version 7." Both carry a V8.2 banner, and neither documents a path for V8.1 or V8.2.

One oddity: --exp is a supported chart row with no article, no range and no default anywhere in the 105 articles. Any number you have seen for it did not come from Midjourney.

What aspect ratio and resolution should an architectural image use?

Set the ratio to the layout you will place the image into. The Aspect Ratio article gives the constraints: "The default aspect ratio is 1:1", "--ar cannot contain decimals. Use 139:100 instead of 1.39:1", and "Extremely wide and tall aspect ratios are experimental and may produce unpredictable results" (Aspect Ratio, updated 27 July 2026, read 28 August 2026). Post 147 is the full --ar reference.

These are the only pixel figures Midjourney publishes.

Four practical consequences for architectural work.

HD is a switch, not a default. The Version article describes it as something you turn on in the settings panel or with --hd, at 1.3 GPU minutes against 0.8 for SD. Your settings panel is the authority on what your account does.

HD caps at 4:1. The chart gives the maximum for V8.1 and V8.2 as "14:1 (4:1 for HD)", so a panoramic streetscape at 6:1 is SD only.

HD cannot be upscaled further, and Pan, Zoom Out or Edit and Vary Region on an HD image "will downscale the resulting images to SD". Extend first, then upscale.

Print size is fixed by pixels. By Midjourney's own arithmetic, 2048 pixels at 300 DPI gives a 6.8 inch print. An A1 pin-up means third-party upscaling.

How much --stylize does architectural work actually want?

Less than you think. This is the single change that moves an image from AI render to photograph of a building.

Midjourney describes --stylize as a slider between literal and artistic: "With a low stylize setting, it's like asking for an image that follows your prompt very closely", while higher values give it "more freedom to interpret your idea". Default 100, range 0 to 1000 (Stylize, updated 27 July 2026, read 28 August 2026). Pair it with --raw, which the Raw article frames as switching off an auto-pilot: with Raw on, "you’ll often get more realistic, photo-like images" (Raw, updated 27 July 2026).

A working ladder, our practice rather than documented behaviour:

• 50 to 120 for massing and material studies, where accuracy beats mood.
• 120 to 250 for exterior hero shots and interiors that need polish without invention.
• 300 to 500 for illustrative sections, diagrams and competition boards.
• 600 and up for mood boards, where you want the model to run.

--chaos is a different dial. It "lets you add more variety to the image results you get from each prompt", 0 to 100, default 0, and higher values make the four images in a grid diverge. Useful for option studies. Keep it at 0 for a hero shot.

How do architectural photographers actually frame a building?

This is craft, not a model feature. Midjourney documents none of it. The model responds to the vocabulary because that vocabulary sits in the captions of the photographs it learned from.

Verticals stay vertical. Point a normal wide lens up at a building and the verticals converge, so it appears to fall away from you. Architectural photographers fix that with a shift lens, moving the lens element parallel to the sensor rather than tilting it. Naming a tilt-shift lens and asking for parallel verticals is the prompt-side equivalent, and the fastest way to make an exterior read as architectural rather than as a snapshot. That is also why two-point perspective, verticals parallel with horizontals converging left and right, is the archviz default, while one-point reads diagrammatic and three-point reads journalistic.

Focal length is not chosen the way a portrait is. A portrait photographer reaches for 85mm to flatter a face. An architectural photographer reaches for 17mm to 24mm shift lenses for exteriors and interiors, 50mm for an approach view, and 100mm to 200mm to compress a streetscape into layered planes. Post 226 has the full camera and lens vocabulary.

Golden hour and blue hour are different jobs. Golden hour gives warm raking light that models facade relief. Blue hour, after sunset while the sky is deep blue and still luminous, is the classic dusk archviz shot, because interior lighting balances the ambient level instead of blowing out. Post 206 lists the lighting vocabulary in full.

What do exterior, streetscape and approach prompts look like?

Copy, swap the building, keep the tail.

architectural photograph of a four-storey civic library, board-formed concrete and oxidised bronze fins, two-point perspective from the opposite pavement, verticals parallel, 24mm tilt-shift lens, late afternoon raking light, three people at the entrance for scale --ar 4:3 --raw --stylize 150 --hd

exterior hero shot of a timber-framed rural visitor centre, larch rainscreen weathered silver-grey, low horizon line, long shadows across a gravel forecourt, golden hour, deep depth of field --ar 16:9 --raw --stylize 200 --hd

frontal elevation view of a nine-storey residential block, white precast concrete frame with deep window reveals, straight-on composition with no perspective convergence, overcast diffuse light --ar 3:2 --raw --stylize 100

street-level view of a mixed-use corner building in a dense European city, brick and dark metal shopfronts, adjacent buildings visible on both sides, pedestrians mid-stride, wet pavement after rain, 35mm at eye level, overcast --ar 3:2 --raw --stylize 180

urban infill facade seen along the street, oblique two-point perspective, existing Victorian terrace on the left, new build on the right, matching eaves line, morning light from the east, parked bicycles --ar 16:9 --raw --stylize 150 --hd

compressed streetscape from across a six-lane boulevard, 135mm telephoto flattening four facades into layered planes, hazy afternoon air, traffic motion blur at the base --ar 21:9 --raw --stylize 250

approach view along a curved gravel drive to a low rural house, the building small in the frame at one third height, mature oaks framing both edges, meadow grasses in the foreground, soft overcast light, 50mm --ar 16:9 --raw --stylize 200

arrival sequence photographed from the threshold of a walled forecourt, the entrance canopy visible through a gap in the wall, layered foreground middle ground and background, late spring dappled shade --ar 4:5 --raw --stylize 220

What do interior, detail and sectional prompts look like?

interior photograph of a double-height reading room, exposed cross-laminated timber soffit, north-facing clerestory glazing, oak floor, felt acoustic panels, two people seated reading, natural light only, 17mm shift lens, verticals parallel --ar 4:3 --raw --stylize 150 --hd

domestic kitchen and dining space, pale terrazzo floor, matt olive joinery, brushed stainless worktop, one wall fully glazed to a garden, mixed daylight and warm pendant light at dusk, 24mm corner composition --ar 3:2 --raw --stylize 200

open-plan office floorplate seen down its long axis, exposed services painted black, raised access floor, full-height glazing on the right, people working at low desks, midday daylight with blinds half drawn --ar 16:9 --raw --stylize 120

close-up material study of a board-formed concrete wall meeting a bronze window reveal, raking side light across the board marks, shallow depth of field, 100mm macro, neutral colour --ar 1:1 --raw --stylize 100 --hd

junction detail of larch cladding meeting a zinc parapet, weathering and staining visible, grazing morning light revealing the profile depth, 85mm, sharp throughout --ar 4:3 --raw --stylize 80

seamless texture of hand-thrown glazed terracotta tiles in a mottled sage glaze, flat even lighting, top-down, no cast shadows --tile --raw --stylize 100

cutaway perspective of a three-storey house with the front wall removed, all three floors visible at once, furniture and figures in each room, warm interior light against cool exterior, clean white cut planes, illustrative not photographic --ar 4:3 --stylize 350

sectional perspective through an atrium, the cut face rendered as flat dark grey poché, the space beyond in full perspective, staircase and bridges legible, diagram-like clarity --ar 3:4 --stylize 400 --chaos 10

The cutaway prompts carry a caveat. The model is not cutting geometry, it is drawing something with the visual signature of a cut, learned from published section perspectives. Poché will be inconsistent, floor-to-floor heights will not agree, and the stair will not resolve. That is a concept image of a section, not a section.

What do massing, masterplan and mood prompts look like?

massing study of a five-block campus scheme, plain untextured white volumes on a light grey ground plane, soft studio lighting with a single shadow direction, three-quarter aerial view, no windows and no materials --ar 16:9 --raw --stylize 50

white card model photographed on a plain background, four stepped volumes on a chipboard base, raking light casting long shadows, shallow depth of field, physical model aesthetic --ar 3:2 --raw --stylize 150

massing option study of a corner site, simple extruded volumes in matte white, identical camera position and identical light, three-quarter view --ar 4:3 --raw --stylize 50 --seed 1234567

oblique aerial view of a masterplan, twelve building volumes arranged around a central green, street pattern and planting legible, low sun at 45 degrees, physical model aesthetic in white card and pale green --ar 16:9 --raw --stylize 100

top-down aerial photograph of an urban block from directly overhead, roofs courtyards and street pattern visible, midday sun with minimal shadow, drone altitude --ar 1:1 --raw --stylize 100 --hd

illustrative site diagram in the manner of a hand-drawn masterplan, muted watercolour washes for landscape zones, building footprints in warm grey, loose linework, textured paper --ar 3:2 --stylize 300

dusk exterior of a glazed pavilion at blue hour, interior lighting balanced against a deep blue sky, warm pools of light spilling onto the terrace, long exposure feel on a tripod, 24mm tilt-shift, verticals parallel --ar 16:9 --raw --stylize 200 --hd

night streetscape of a shopfront and canopy, illuminated signage in the window reading "ARCHIVE", wet asphalt reflecting the light, cool ambient mixed with warm interior, 35mm --ar 3:2 --raw --stylize 180

tower at night seen from a distance, floor plates reading as horizontal bands of light, surrounding city dark, slight atmospheric haze, 200mm telephoto compression --ar 4:5 --raw --stylize 220

mood board image for a coastal house scheme, weathered timber, salt-bleached concrete, sea mist, muted greens and greys, texture and atmosphere only with no building visible --ar 1:1 --stylize 600 --chaos 25

atmospheric concept image of a threshold moment, deep shadow opening to a bright courtyard, heavy masonry, dust suspended in a light beam, emotion over information --ar 4:5 --stylize 700 --weird 250

palette and mood study, four material swatches arranged flat, brushed brass, pale oak, off-white lime plaster, dark green marble, soft even north light --ar 3:2 --raw --stylize 120

The signage prompt uses double quotation marks deliberately. Midjourney's Text Generation article says text works "When you use Midjourney versions 6 and later" inside double quotes, that "single quotes or apostrophes (' ') won't work for this", and that if text misbehaves you should "consider using Raw or lowering the Stylize value" (Text Generation, updated 27 July 2026). Keep the string short.

Why is a generated image not a design?

Because it has no dimensions, no structure, no code compliance and no buildability, and a better prompt adds none of them.

Midjourney will give you a beautiful cantilever that cannot stand, because it learned what cantilevers look like in photographs and nothing about moment or deflection. It will give you a stair whose risers meet no building regulation anywhere, because riser rules are numeric and it is not counting. It will render a window schedule that means nothing and a fire escape route that goes nowhere. Those are not defects to be prompted away. They are what the tool is.

Fit for: concept exploration, mood, early client conversation, testing whether a massing idea reads, building a material palette, generating options fast enough to throw most away.

Not fit for: construction, planning submission, tender information, coordination, or anything presented as a design proposal.

Be direct about that last one. Putting AI imagery in front of a client or an authority as if it were your design work, without saying what it is, is a professional problem and not a stylistic one. It misrepresents how much design has happened, and invites decisions on the strength of a picture with nothing behind it. Midjourney's own Community Guidelines put it in one line: "Do not intentionally mislead recipients of generated images about their nature or source" (Community Guidelines, updated 17 August 2026, read 28 August 2026).

We are not going to invent a rule for you, because the rules binding you depend on where and with whom you are registered. Read your own registration board's current code and your institute's current guidance, in the actual text rather than a summary. Across borders, this is a per-jurisdiction question with no single answer.

One thing often mistaken for a solution: Midjourney embeds an identifier in the files it delivers, checkable at midjourney.com/verify. Its Content Authenticity page says that tag "is easily lost: Screenshots, edits, and most social media sites remove it" (updated 17 August 2026, read 28 August 2026). Embedded metadata is not a disclosure mechanism. Saying the words is.

How do you keep a board consistent across ten images?

Style references, because the character features are stranded on older versions. Midjourney's Style Reference article describes --sref as capturing the overall style: "like colors, medium, textures, or lighting", compatible with "versions 6 and later", paired with --sw on a 0 to 1000 scale defaulting to 100 (Style Reference, updated 27 July 2026). Run --sref random once, note the code, reuse it across the set.

exterior hero shot of a chapel in a pine clearing, board-marked concrete, single tall slot window, overcast light --ar 3:2 --raw --stylize 150 --sref 1234567890 --sw 250

interior of the same chapel looking back toward the entrance, timber benches, light falling down the concrete wall --ar 3:2 --raw --stylize 150 --sref 1234567890 --sw 250

Three documented constraints. You "cannot create a style code based on an uploaded image", so upload it as a style reference instead. --sw is not compatible with Moodboards. And --sv, the style-reference version parameter, is documented only under headings for version 7 and version 6, with nothing published for V8.x, so treat it as undocumented rather than guessing. Post 214 has the full --sref breakdown.

Permutations put alternatives in curly braces and produce one prompt per option, parameters included, capped by plan at 4 on Basic, 10 on Standard and 40 on Pro and Mega, in Fast and Turbo only. --repeat reruns a prompt under the same caps.

--seed is the weakest consistency tool and the docs say so. Whole numbers 0 to 4294967295, a header reading "Seeds are 99% identical in V8.X", and an explicit warning that seeds are "Not for Saving Styles" and should not be used in Turbo mode. It is a control variable for A/B testing one word, not a way to lock a building down. Post 236 covers reproducing a result with --seed.

Quick reference: which parameter for which job?

Where do you keep these once they work?

Architects abandon this workflow not because the images are bad, but because the tail that worked in March is buried in a Discord scrollback, and by the time you rebuild it the version has moved and the parameters are stranded. Keep the working tails somewhere versioned, with the date you last verified them and the version they were checked against.

For the wider practice workflow rather than the image craft, ChatGPT for architects covers briefing and reporting, and AI prompts for landscape architecture covers the site side.

Prompt Architects builds and stores the prompt. We do not generate the image, so Midjourney still does the rendering. Image Prompt Generation starts on the Pro plan, $4.99 a month at the time of writing on our pricing page, and the Image Prompt Library sits on Advanced. The Free plan includes 5 prompt enhancements per day, forever, per our FAQ page.