Most AI tools in architecture in 2026 solve the wrong problem. Midjourney renders beautifully. Gendo and mnml.ai turn sketches into rendered output. Maket generates floor plans. But none of them solve the actual bottleneck of early design work — the messy space between “client wrote a brief” and “architect has a direction worth developing.” Nuit exists because that space was underserved.
This is the story of why we built it, what we got wrong early, and what the product is actually for.
The Problem We Noticed
Early-stage concept work has a specific shape that most AI tools miss.
An architect gets a brief. The brief is always partial — the client knows some things, is unsure about others, contradicts themselves on a few points. The architect’s job in the first week or two is to produce enough visual material that the client can react to something concrete. Not final renderings. Not construction drawings. Concept directions — enough form, palette, and mood to answer “do you want this, this, or this?”
This phase used to involve a lot of rendering staff, a lot of sketching, and a lot of time. It was the phase where small practices lost to large ones — because producing three or five directions on a residential concept just cost too much in junior-staff hours.
The first wave of AI tools compressed single-image production. Midjourney can make a beautiful exterior in minutes. Gendo can turn a sketch into a high-quality render. That was progress.
But single-image production isn’t the bottleneck of concept work. The bottleneck is exploration. Trying ten directions. Branching from a promising one into five variants. Carrying a palette across exterior, plan, and interior so the project feels like one project. Iterating on a chosen direction without losing what made it good.
None of the existing tools handled that part.
What We Tried First
We started with a simple assumption: give architects a faster way to render from a text brief and they’ll figure out the rest.
That was wrong.
Architects who used the early versions came back with the same complaint: every render was an island. Generating ten variants meant ten disconnected images. Branching from one to explore further meant starting from scratch and re-describing the direction in text. There was no continuity, no project memory, no way to work on a concept as a single evolving thing.
We had built a slightly faster Midjourney. That wasn’t the product.
The Shift: Concept Work Is a Tree, Not a Gallery
The insight that reshaped the product came from watching how architects actually worked.
A concept doesn’t come out of one prompt. It comes out of a tree. The architect starts with a broad direction. Generates three or four options. Picks the one that clicks. Branches from that one to try variations — different palette, different roof form, different material. Picks again. Goes deeper. Eventually a specific direction emerges, rendered across exterior, plan, and interior, with enough consistency that it reads as one project.
The tree isn’t an artifact of the workflow. It is the workflow.
So we rebuilt the product around it. An infinite canvas where every generation is a node. Every image can be branched from. Every branch carries forward the project’s style context so the child generations stay coherent with the parent. The architect doesn’t have to re-describe the project on every prompt — the project describes itself as it grows.
This is why Nuit is a canvas with branches, not a form with a “generate” button. The form-and-gallery shape fits single-image tools. It does not fit concept exploration.
Image-to-Image Iteration, Not Text-to-Image Restarts
The second insight was about how edits work.
When a client says “I like this concept but make the façade warmer,” the architect shouldn’t have to rewrite the entire brief to try that change. The previous render is the anchor — the new render should start from the image, not from a blank text prompt.
This is how professional iteration actually works. You don’t re-describe the whole thing; you point at what you want to change.
Nuit works this way. When you edit a generation, the prior image is handed forward as a visual anchor. The text instruction describes what to change. The output keeps what should stay and changes what shouldn’t.
This matches how architects talk about their work. It also produces dramatically better iteration output than text-only tools, because the model has the target composition already in view.
Nano Banana, which many architects use separately for precision editing, works on a similar principle and is rightly respected for it. What Nuit adds is integration — the same iteration model, but inside the tree, carrying the project’s full context rather than floating loose.
End-to-End, Not Just Exteriors
The third insight was about scope.
A villa is not an exterior. It’s an exterior, a plan, and a set of interior rooms that belong to each other. Most AI concept tools do one of those. Architects who need all three either use multiple tools (and lose coherence between them) or commission the parts AI can’t do.
Nuit was built to cover exterior, plan, and interior as one flow. The same project context — palette, materials, style notes, site relationship — carries into all three modes. The interiors match the exterior. The plan reflects the form. The whole concept reads as one decision.
This doesn’t mean Nuit replaces specialist tools for every use case. Architects doing a hero rendering of a single façade might still use Midjourney. Architects working on a parametric plan problem might still use Maket. Architects editing a single photograph might still use Nano Banana.
What Nuit offers that those tools don’t is the through-line — the project as one object from brief to full concept package.
What We Explicitly Decided Not to Build
A few things we chose not to do, and the reasoning.
BIM or construction documents. The technical phase is a different problem. Revit, ArchiCAD, and similar tools solve it well. Adding drafting to a concept tool would have diluted the concept-phase focus.
Photorealistic final hero renderings. There’s a ceiling on how realistic concept imagery should be. Too realistic, and clients expect the concept to match the finished space exactly. A slightly stylized concept rendering communicates “this is a direction” better than a hyperreal image does. Hero renderings for marketing can be produced with other tools once the concept is approved.
Automatic code compliance. Zoning, setbacks, energy code, accessibility — all site-specific, jurisdiction-specific, and weight-bearing. AI cannot be trusted with these without human oversight, and pretending otherwise would be dangerous.
BYOK (bring your own API key). Some AI tools let users plug in their own model API keys. This seems flexible but leaks complexity onto the user. Nuit uses a subscription model so the user gets a single predictable price and we handle provider selection and capacity planning behind the scenes.
Multi-tenant team features too early. The product is designed around individual architect workflow first. Team features (shared projects, comments, review flows) are on the path but not at the expense of the individual-architect core.
Why does the tree matter to users?
The tree isn’t a UI gimmick. It changes what the tool is for.
Lineage is visible. You can see how a concept evolved. Which image begat which variant. Why the chosen direction looks the way it does. This is useful for presentations, for team review, and for the architect’s own memory weeks later when the project picks up again.
Abandoned branches stay available. A direction that didn’t win the client review isn’t lost — it’s still in the tree. If the project pivots later, the architect can pull a dormant branch back to the top.
Exploration isn’t wasted. In a linear tool, generating ten variants to pick one means nine are throwaway. In a tree, the nine are context — they’re the reason the tenth is the right answer. This changes how it feels to explore.
Context carries. The architect doesn’t have to keep re-describing the project. Branch from an image and the style carries forward. This removes the “prompt fatigue” that makes text-to-image tools feel exhausting on longer projects.
What Surprised Us About Early Users
A few things we didn’t expect.
Architects use the tree more for client presentations than for exploration. We built the branching to help architects explore. In practice, many use it to show clients the exploration — “we considered these three directions, chose this one, developed these variants, picked the best.” The tree becomes a presentation artifact, not just a workflow artifact.
Plan mode gets used more than interior mode. We expected interiors to be the heaviest use, because interior design has a long tradition of mood-board visualization. In practice, plans get used more, because plan variants are the hardest thing to generate manually and architects feel the time savings most sharply there.
Collaboration with clients happens inside the canvas, not around it. Some architects open the canvas during client calls and generate in real time while the client watches. This changed how we think about the product’s pacing — things need to happen fast enough that live collaboration feels natural.
Small practices are the strongest early adopters. Solo architects and two-to-three person studios. Large firms are slower — procurement, IT review, workflow inertia. Small practices have no such friction and move fast.
Where is the product going?
A few directions, not commitments.
Better site context. Concept work is always site-specific. Giving the tool more site information — topography, orientation, surrounding buildings — would let the concept respond to the site more precisely.
Deeper material detail. The gap between “this is a mood direction” and “this is a specification” is big. Narrowing it — even in concept-phase terms — would save architects more handoff work.
Plan-to-rendering coherence. The plan and the rendering should read as the same building. Today they mostly do, but not always. This is a consistency problem worth solving well.
Team workflows. Shared projects. Comments. Review flows. On the path, not yet in hand.
Why did we name it Nuit?
Nuit is French for night. Concept work happens in a kind of twilight — the brief is half-clear, the direction is half-formed, the client is still deciding what they want. You’re working with shadows, feeling for form. The tool should fit that mood, not fight it.
The domain is nuit.archi — architecture-first, contemplative, quiet, focused on the part of the work that happens before things get fully lit.
Related reading
- What is Nuit? Complete 2026 Overview — Nuit is an AI workflow for architectural concept design — exteriors, floor plans, and…
- AI Doesn’t Replace Architects — It Moves the Bottleneck — The question ‘will AI replace architects’ is asked often and answered badly.
- The Concept Phase Is Broken in 2026 — The concept phase of an architectural project is the most underpriced, most overworked,…
- Why AI Design Needs Phase Separation — An architectural concept has phases — exterior, plan, interior, masterplan.
- Branching as a Design Exploration Technique — Branching is the structural mechanism that makes AI concept exploration cheap, comparison…
Frequently Asked Questions
Who is Nuit for?
Architects, interior designers, and small studios doing early-stage concept work. The sweet spot is solo architects and two-to-five person practices who need to produce multi-direction concept packages quickly and coherently. Larger firms use it too, typically on specific concept-heavy projects rather than as an enterprise standard.
How is Nuit different from Midjourney?
Midjourney is a single-image tool; Nuit is a project tool. Midjourney produces beautiful individual renderings with no project memory. Nuit carries project context across every generation — exterior, plan, interior all cohere as one project. The tree-based canvas lets you explore and branch in a way Midjourney’s gallery doesn’t support.
How is Nuit different from Gendo or mnml.ai?
Gendo and mnml.ai turn sketches into rendered output — they need a drawing as input. Nuit works from text briefs and iterates visually, with or without a sketch. The workflows are complementary: architects often use Gendo for sketch-to-render and Nuit for text-to-concept exploration.
Does Nuit replace Nano Banana?
No. Nano Banana is excellent at precision editing of a single image — swapping elements, adjusting lighting, fine-tuning materials. Nuit does its own image-to-image iteration inside the project tree, but Nano Banana remains a respected tool for freestanding single-image edits. Many architects use both.
Why subscription instead of pay-per-image?
Subscription is predictable. Pay-per-image creates pressure to skip exploration to save pennies, which defeats the purpose of a tool built for exploration. A flat subscription lets the architect branch freely without counting every generation.
Does Nuit generate construction documents?
No. Nuit is a concept-phase tool. Construction documents are a different problem solved well by Revit, ArchiCAD, and similar tools. Nuit covers exterior, plan, and interior concept visualization; the technical phase handoff is unchanged.
Can I try it before subscribing?
Yes. There’s a free tier with 10 generations, no card required. Enough to run a full concept project through the tool and decide whether it fits your workflow.
Try Nuit free — 10 generations, no card required. Build a concept as a tree, not a gallery — explore, branch, iterate, and keep your project coherent from brief to direction. Start your project →