You need to show six material options for a client's floor: concrete, light oak, dark walnut, polished marble, white microcement, and reclaimed brick. The client needs to see them side by side and understand what each one does to the space.
You open Midjourney. You type your prompt. It generates a beautiful image — warm morning light, plausible proportions, a contemporary open-plan layout. You change the material term from "concrete" to "light oak" and regenerate.
The second image is a different room. A different scale. A different window configuration. A different camera angle. The comparison is meaningless.
You try again. Same problem. You add "same perspective, same camera angle, same ceiling height" to your prompt text. The AI ignores these instructions almost entirely.
This is not a prompting problem. This is a fundamental architecture problem — and no amount of better prompting will solve it.
Why Standard Prompting Cannot Maintain Spatial Continuity
Midjourney's standard prompting architecture generates each image independently. It reads your prompt text, draws on its training data, and produces the image it thinks best matches the description. There is no persistent spatial memory between generations.
The phrase "same space" in a prompt is not a constraint the AI can enforce. It is just text to be interpreted alongside everything else. Every generation is a fresh draw from the model's probability distribution. Small changes in prompt wording produce unpredictable changes in the output.
This is fine for creative exploration. It is completely unusable for professional architectural presentation work, where the client needs to compare materials, not compare rooms.
The Parametric Solution
The solution is not a better prompt. The solution is a different approach entirely: parametric consistency.
Midjourney has four advanced parameters that provide explicit control over the generation process:
- `--sref` (Style Reference): Anchors the aesthetic DNA — lighting quality, colour temperature, material feel, tonal range.
- `--cref` (Character Reference): Anchors the spatial geometry — perspective type, camera height, proportional relationships, spatial scale.
- `--seed`: Provides a deterministic generation anchor. Same seed plus same prompt equals the same spatial base.
- `::weighting`: Gives fine-grained priority control over specific prompt elements — forcing a one-point perspective or a specific camera height.
These four tools, used together in the right sequence, transform Midjourney from a creative lottery into a precision instrument.
The Controlled Iteration Framework
The complete workflow has four steps:
- Step 1 — Hero Image Establishment: Generate a single baseline image that defines the spatial DNA perfectly. The hero image prompt must specify six things precisely: programme, dimensions, camera height (always 1.5 metres for architectural interior work), perspective type (always specify explicitly — "one-point perspective, central vanishing point"), lighting quality and direction, and materiality palette.
- Step 2 — DNA Lock: Extract the `--sref`, `--cref`, and `--seed` from the hero image and apply them to all subsequent generations. Without this step, every new prompt is a fresh generation. With it, every subsequent image inherits the spatial geometry and aesthetic character of the hero image exactly.
- Step 3 — Controlled Variable Iteration: With the DNA locked, systematically alter only the design variable being tested — materiality, programme, or lighting — while everything else remains constant. The only text that changes between prompts is the specific variable term. Everything else — programme, dimensions, camera, perspective, lighting, photography style, DNA lock — stays identical.
- Step 4 — Orthographic Consistency Check: Review every image against a consistency threshold: camera height approximately 1.5 metres, perspective type matches hero image, ceiling height consistent, lighting colour temperature consistent, window configuration similar. Any image that fails two or more checks is regenerated.
The Consistency Threshold
Before including any image in the final deck, apply the consistency threshold check:
| Check | Pass/Fail |
|-------|-----------|
| Camera height approximately 1.5 metres? | |
| Perspective type matches hero image? | |
| Ceiling height consistent with hero dimensions? | |
| Lighting colour temperature consistent? | |
| Window configuration similar to hero? | |
| Material rendering photographic, not stylised? | |
Any image that fails two or more checks should be regenerated. An image that fails one check may be acceptable depending on project context.
Why the DNA Lock Must Be Applied Together
The most common error architects make with these parameters is using `--sref` without `--cref` — or vice versa.
Using `--sref` without `--cref` gives you the aesthetic but not the geometry. The lighting and colour temperature match the hero image, but the camera height shifts and the perspective warps.
Using `--cref` without `--sref` gives you the geometry but not the visual character. The spatial proportions hold, but the rendering style looks different from the hero — different colour temperature, different material rendering approach, different tonal range.
Always apply `--sref` and `--cref` together, pointing to the same hero image URL. The parameters do different things; they are not interchangeable.
The Style Weight Sweet Spot
The `--sw` parameter controls how strongly `--sref` influences the output. Too low and the aesthetic drifts between generations. Too high and the geometry begins to degrade — edges warp, proportions stretch, perspective becomes inconsistent.
For architectural work, `--sw 150` is the recommended starting point. This is strong enough to maintain aesthetic cohesion across a material deck but not so strong that it starts distorting geometry. For very large variation sets where aesthetic cohesion is paramount, try 200 before going higher. Above 300, geometric accuracy begins to degrade.
What This Is and Is Not
The Controlled Iteration Framework produces spatially consistent, client-ready presentation imagery. It does not produce verified design drawings. It does not produce planning compliance evidence. It does not produce structural calculations.
AI-generated imagery is a visualisation and concept exploration tool. It is not a design document. It is not a specification. It does not constitute a professional opinion on structural feasibility, planning compliance, or spatial proportions.
Every AI image distributed to any third party — clients, contractors, planning authorities — must carry the mandatory disclosure:
> AI Imagery Disclosure: These are AI-generated conceptual impressions for discussion purposes only — they do not represent a final design, a planning application, or a structural or material specification. All dimensions, proportions, and materials must be verified against the architect's drawings and specifications.
This is not a legal formality. It is professional accountability. The architect authorises every output.
Related Resources
The Controlled Iteration Framework is the specific solution documented in Architect's Parametric Prompt Engineering: The Controlled Iteration Framework for AI Spatial Prototyping. The guide covers the complete four-step workflow — hero image establishment, DNA lock, controlled variable iteration, and orthographic consistency check — with worked examples for both residential material decks and commercial programme iteration decks. It includes the full Midjourney parameter reference table, the ::weighting syntax guide, and the mandatory AI imagery disclosure requirements for professional practice.
If you are working with AI-generated concept imagery at the earlier design phases and need to translate those images into buildable specifications for your drafting team, From Generative Visuals to Geometric Parameters covers the complementary workflow: using ChatGPT with Advanced Data Analysis to extract estimated spatial proportions from a Midjourney concept image into a reviewable CSV constraint sheet.
Stop Guessing, Start Systemising
The lottery problem is real. Standard prompting cannot solve it. But the solution is not complicated — it is a structured parametric workflow that treats Midjourney as what it can be: a precision instrument for generating spatially consistent imagery, rather than a creative box that produces beautiful but uncontrollable results.
The Architect's Parametric Prompt Engineering guide documents the complete four-step Controlled Iteration Framework with 25 copy-paste-ready prompts, two complete worked examples (residential material deck and hotel lobby programme iteration), and the full Midjourney parameter reference table. Tested with Midjourney v6.1 and v7.
Grab it here.
Practical AI for Professionals