← Back to blog

Why AI Design Needs Phase Separation

Nuit Team · · 12 min read
aiarchitectureconcept-designworkflowphasesinteriorfloor-plans

An architectural concept is not one image. It has phases — exterior, floor plan, interior, master plan — and each phase has different rules, different references, and a different relationship to the project’s design identity. A general-purpose image model treats all four as the same text-to-image task and produces results that are individually plausible but collectively incoherent. A purpose-built workflow models each phase as a separate mode that shares the project’s brief and saved references, but runs its own prompt strategy. The result is that the four phases compose into a single design rather than four loosely related images. Phase separation is not a UI choice. It is what makes the output usable.

This is a satellite of the topic cluster anchored at Nano Banana for Architecture: Where It Works, Where It Falls Short. It assumes you have read the pillar — if you have not, the four-gap framing there is the right context for what follows.

The Four Phases of a Concept Package

A concept-stage architectural project produces four classes of deliverable, in roughly this order:

  1. Exterior concept. The massing, materials, and stance of the building. The exterior establishes the project’s design identity — its “language.”
  2. Floor plan. The layout of internal space. Rooms, sizes, adjacencies, circulation. The plan makes the building inhabitable.
  3. Interior visualizations. Photorealistic views of the rooms defined in the floor plan, in the style of the exterior. Interiors are how the project becomes vivid for clients.
  4. Master plan or site plan. The relationship between the building and its larger context — site, neighbors, landscape, road. Not all projects need this; residential renovations might skip it; developments and resorts depend on it.

These are not interchangeable. They serve different purposes, are evaluated against different criteria, and require different kinds of accuracy. An exterior is judged on visual impact and material clarity. A floor plan is judged on dimensional logic and circulation. An interior is judged on atmosphere, light, and how well it expresses the project’s character. A master plan is judged on the relationship between the building’s footprint and its site.

Trying to produce all four from a single image-generation prompt is structurally wrong, no matter how good the model is.

What happens when you try single-prompt generation?

A common first attempt with a general image model goes like this:

“Generate a 200 m² modernist villa in Bali — exterior view with the floor plan visible below, and the interior of the living room and kitchen also shown. Two stories, indoor-outdoor flow, warm wood and stone.”

The model produces an image. It is often beautiful. It contains an exterior in the upper portion, an attempt at a floor plan in the lower portion, and small thumbnails of interior moments. Looked at as a single image, it can be impressive.

Looked at as a project, it is wrong in three ways:

The floor plan is not a floor plan. It is a stylized illustration of a floor plan — boxes that look approximately like rooms, with labels that look approximately like room names. The dimensions are nonsense. The wall thicknesses are inconsistent. The doors are decorative. Trying to use it as an actual floor plan means redrawing it from scratch.

The interiors are unrelated to the plan. The little kitchen thumbnail in the corner is a kitchen the model imagined, not the kitchen of the plan above it. The wall behind the dining table does not exist in the plan. The window placement in the interior does not match the window placement on the exterior.

The exterior is the only piece that works. And it works because it is the only piece the model treated as the primary subject of the image. The rest is decoration around it.

A general image model is one tool with one mode. It cannot be four tools at once. The output reflects this.

What Phase Separation Actually Looks Like

The correct shape is a workflow where each phase has its own mode, with the appropriate generation strategy, and the phases share data through the project rather than through a single mega-prompt.

In Nuit specifically:

  • Exterior phase. Text-to-image generation with optional reference images. The output is one or more 16:9 (or other aspect-ratio) photorealistic views of the building. The project brief is appended automatically. The user iterates with branching until the right exterior emerges and saves it.

  • Plans phase. A different mode entirely. The user first authors a floor-plan brief — a structured list of rooms with target areas, types, and adjacencies. The brief can be written manually or generated by AI from the project context and then edited. The generation step takes the floor-plan brief, the project context, and the saved exterior as a reference, and produces a top-down floor plan in a style that matches the project. The saved plan defines the room list that the next phase reads.

  • Interiors phase. A room-at-a-time mode. The room list comes from the saved floor plan. The user picks a room — Kitchen, Master Bedroom, Living Room — and generates interior views. The prompt template includes the room type, the saved floor plan as a structural reference, the saved exterior as a stylistic reference, and the relevant moodboard section if it exists. The output is a photorealistic interior consistent with the rest of the project.

  • Master Plan phase. A site-context mode. The saved exterior is placed in a larger site image — the building and its surroundings, from an aerial or elevated perspective. Useful for developments, resorts, campuses, and any project where the relationship to context is part of the brief.

Each mode looks different, runs different prompt templates, and has different inputs. What they share is the project’s brief, the project’s saved references, and the project’s accumulated visual memory.

What Each Phase Specifically Needs

A useful way to understand why one prompt cannot cover all four: each phase has its own kind of accuracy.

Exterior accuracy is about visual identity. The building should look the way the brief describes — typology, materials, scale, stance. A general model is good at this if the prompt is good.

Floor-plan accuracy is dimensional and topological. The rooms should be at roughly the right size relative to each other. The connections between rooms should be sensible (kitchen near dining, bedrooms grouped, bathrooms accessible). The circulation should make sense. This is closer to a diagram than to an image. General models are weak at it without a dedicated mode and the right template.

Interior accuracy is about atmosphere and consistency with the rest of the project. The room should feel like it belongs to the building you have already designed. The materials should echo the exterior. The light should be consistent. The proportions should match the floor plan. This is the phase where most projects fall apart visually, because consistency across multiple rooms is hard without explicit project memory.

Master-plan accuracy is about context and scale. The building should sit on the site correctly — not floating, not too big or too small, with sensible relationships to neighbors and landscape. The image needs to read as a place, not as an isolated render.

A single prompt cannot meaningfully optimize for all four kinds of accuracy at once. A workflow that splits them into four modes can — each mode applies its own strategy, and the phases compose at the project level rather than at the prompt level.

How do the phases connect through project memory?

The trick that makes phase separation work is that the phases are not independent. They share a common state — the project brief, the saved references, the style summary — that flows automatically between them.

In Nuit:

  • The project brief is set once and travels with every prompt in every phase.
  • The saved exterior becomes a reference for the Plans, Interiors, and Master Plan phases. The user does not have to manually attach it each time.
  • The saved floor plan defines the room list available in the Interiors phase. Choosing “Kitchen” in interiors automatically pulls the kitchen’s footprint from the plan.
  • The saved interiors in turn become references for subsequent rooms — once the living room is settled, the kitchen generation has the living room as a visual anchor to match.
  • The moodboard sections sit alongside all of this. A “Kitchen” section’s references are attached when generating in the kitchen room.

The user’s job is to make the local decisions in each phase. The cross-phase consistency is handled by data flowing between phases, not by the user re-typing context.

This is the meaningful difference from working in a general image tool with manual workflow. In a manual workflow, every reference, every piece of context, every cross-phase relationship is the user’s responsibility. In a phase-modeled workflow, the workflow remembers and the user designs.

The Plans Phase in Detail — The Hardest Phase to Get Right

Floor plans are where most architectural AI tools, including general image models, perform worst. The reason is that a floor plan is not a normal image. It is a structured diagram with specific conventions — wall thicknesses, door swings, dimension lines, room labels — that a model trained mostly on photographic data will approximate but not reliably produce.

A purpose-built Plans phase addresses this with a different pipeline. The starting point is a floor-plan brief — a structured list of rooms, areas, and adjacencies — not a free-text prompt. The brief can be authored by the user manually, or generated from the project context by AI and then edited.

The brief is the most consistent contribution to plan quality. It tells the model not just what the building should look like, but exactly what rooms it should contain and how big they should be. Without it, the model invents a room list. With it, the model lays out the room list you provided.

After generation, the user reviews the plans (count = 3 by default, like exterior). When the right plan is chosen, the Save action does two things: it adds the plan to the project’s references, and it unlocks the Interiors phase by extracting the room list from the saved plan.

This unlock is the bridge between Plans and Interiors. Without a saved plan, Interiors does not know what rooms exist. With one, Interiors becomes a list of rooms the user can generate one at a time.

The Interiors Phase — Why It Needs the Plan

People sometimes ask why Interiors cannot start independently. The answer is that the value of Interiors is not in generating one beautiful room — a general model can do that. The value is in generating multiple rooms of the same project that read as the same project.

For that, the workflow needs to know:

  • Which rooms exist. Otherwise the user has to type the room list manually and remember which rooms have been done and which have not.
  • What the exterior looks like. So the interior’s materials and light feel consistent.
  • What previous interiors look like. So the kitchen and the living room of the same villa share a palette.

A general model has none of this. A phase-modeled workflow has all of it for free — the room list comes from the saved plan, the exterior is the saved-concept reference, the previous interiors are saved-interior references. The user picks a room and clicks Generate. The result fits.

For deeper detail on the consistency mechanisms that underlie all of this, see How to Get AI to Generate Consistent Designs Across a Project.

Why can’t you just add a floor plan mode to a general image tool?

A reasonable objection: why not just add a floor-plan mode to a general tool, and call it solved?

The answer is that the mode is not the hard part. The hard part is the data flow between modes. Adding a separate plan-generation button to a flat image tool gives you a separate plan generator — but the plan does not know about your exterior, the interiors do not know about the plan, and the project as a whole has no memory.

The architectural workflow is the right shape because the phases share state. The phases are an expression of the workflow, not a substitute for it. Tools that try to be “general image with modes added” produce the same flat-list output, just with more buttons.

The Practical Takeaway

If you are doing a concept-phase project, the question to ask is not “can this tool make a floor plan.” Every modern image tool can make something that looks like a floor plan. The question is “can the floor plan it makes connect to the exterior I already saved and feed the room list to the interiors I want to generate next?”

If the answer is no, you are doing four separate jobs instead of one.

The pillar piece — Nano Banana for Architecture — covers the broader frame. The head-to-head with Nano Banana on this specific point is in Nuit vs Nano Banana: When Each Fits. And the related piece on organizing references across phases is Moodboards with Sections for AI Workflows.

Start designing with Nuit

Generate architectural concepts from a simple description. No sketches, no 3D software.

Try it free