Floor Plan Best Practices — what to feed the AI for a perfect read

If you've ever uploaded a floor plan and felt the result almost landed — a door drifted onto the wrong wall, a window went missing, a room came back labelled Room 3 — the fix is usually upstream of the AI. It's in the plan you sent.

This is the field guide we wish every visitor had open before they hit Conjure. Spend two minutes on these, and hours of iteration disappear.

1. Pick the right file format

Our pipeline reads five inputs natively. Pick the one you already have, and we handle the rest:

• JPG, PNG, WebP — the everyday choice. Phone snaps of a printed sheet, screenshots from a planning app, exports from any CAD tool. Large captures get gently re-encoded so the upload travels fast; no manual resize needed.
• PDF — drag the whole document in. Up to 10 MB, up to 100 pages. Our reader walks every page like a draftsman with a clipboard. You don't need to flatten, crop, or convert it first.
• HEIC, HEIF — the iPhone default. Transcoded transparently the moment it lands. No more "your browser can't read this image" errors.
• DXF — the precision path. CAD files run through a deterministic, millimeter-accurate reader. No estimation, no drift. If you have a DXF, send it.
• DWG — the one to avoid. It's a proprietary binary; even CAD tools struggle to read each other's DWG files reliably. Export to DXF first.

2. Resolution is the silent variable

The single highest-impact thing you can do is upload a sharp image.

• At least 1500 pixels on the shortest edge. Smaller and per-room labels start to dissolve into wall ink — and that's how Primary Bedroom turns into Room 3.
• Sharp, in focus, well-lit. A phone photo of a flat printed sheet works beautifully. Keep the camera parallel to the page, avoid a desk-lamp glare, and don't crop with your thumb in the frame.
• No perspective tilt. A skewed capture stretches the geometry on one side, and the AI dutifully reads it as this room is a trapezoid. Shoot straight-on, or use your phone's document-scan mode — it flattens perspective for free.
• Low compression is enough. A clean 500 KB JPEG carries every detail the pipeline needs; a 50 KB one starts losing labels at the edges.

3. Walls — bold, black, unbroken

The single most common failure mode we see is a plan rendered in pale gray on white. Our wall detector keys on ink density; faint walls read as background, and the rooms melt into one open space.

• Black walls beat gray walls. When you export from CAD, pick the bolder line weight if you have a choice.
• Solid beats dashed. Dashed lines are perfect for future addition annotations, but on the main outline they trip the detector.
• Closed perimeters matter. A four-pixel gap at a corner is enough for our flood-fill to mistake the room for the outside world. Walk the envelope with your eye before uploading.

4. Doors and windows the AI can read

Our pipeline classifies openings two ways — by symbol (door swing arcs, window double-lines) and by label. Labels win every time.

• Label them when you can. A small WIN 1.5m or DOOR 0.9m annotation next to the opening lets the AI lift the dimension verbatim instead of guessing from pixel width.
• Door swings face into the room they belong to. The quarter-circle arc tells us which side of the wall owns the door — without it, a door on a shared wall can attach to the wrong neighbour.
• Don't merge windows. Two windows drawn as one wide rectangle get read as one big window. Leave a sliver of wall between them and both survive the read.

5. Dimensions — at least one, ideally a chain

Every plan needs a scale anchor. Without one, the AI can describe the rooms beautifully and still get the absolute sizes wrong — which cascades into the budget, the furniture sizing, the whole brief.

• One labelled overall dimension is enough. A 12.40 m on the bottom edge, or a single room captioned Living · 24 m², gives us a fixed point.
• A perimeter chain is the gold standard. Most architectural plans have one along at least one side. If yours does, leave it in — don't crop it out to "tidy up" the upload.
• Imperial works too. Feet-and-inches labels are read alongside metric without any setup; the brief comes back in the units you sent.

6. Label your rooms

The difference between an analysis that says Bedroom 3 · 12 m² and one that says Children's Bedroom — south-facing · 12 m² is one word printed on the plan.

• Even a single word helps. Kitchen, Living, Bath, Bedroom, Office, Garage — write them inside the rectangle, not in a margin legend the eye has to chase.
• Open-plan zones can carry multiple labels. A combined kitchen-dining-living reads cleanly when each function has its own caption inside the shared rectangle. We join them with slashes on our side — Living / Dining / Kitchen.
• When a label is unreadable, the AI guesses. Sometimes it guesses well; sometimes it lands on Room 3. When that happens we surface a one-click What is this room? suggester on the results page — but a labelled plan saves you the step.

7. One floor per upload

Multi-floor buildings get one upload per floor — separate JPGs, separate PDFs, or separate pages within one PDF.

• Don't stack two floors side by side in one image. The envelope detector assumes one outline per upload; two floors crammed in get treated as one strangely-shaped house.
• Multi-page PDFs are fine. Each page is read as its own plan, and the brief threads them together.
• No floor plan at all? Switch modes. If you're starting from an empty lot and want the AI to design the layout from scratch, the land-plot upload is the right entry point. Authoritative measurements (length, width, story count) carry straight into a buildable brief — no plan required.

8. The DXF shortcut

If you have access to the CAD file behind your plan, send the DXF and skip every estimation step. The DXF reader is deterministic — same input always returns the same output, millimeter-precise, with no vision pass in the loop. Room areas land exact. Wall thicknesses land exact. The brief that comes back is buildable, not interpreted.

Most architects, draftspeople, and design-build firms will hand you a DXF on request. It's a one-line ask that saves the AI an entire layer of guesswork.

9. What you don't need to worry about

A few things people clean up before upload that they don't actually need to — we handle them on our side:

• File size up to 10 MB. A high-resolution phone capture is well within bounds.
• iPhone HEIC format. Auto-transcoded the moment it lands; drop it raw.
• Colour plans. Black-and-white reads cleanest, but a coloured architectural sheet works too — our wall detector keys on ink density, not hue.
• Furniture symbols on the source plan. Beds, sofas, tables drawn into the plan are useful context for the AI, not a distraction.
• Title blocks, north arrows, scale bars, firm stamps. All fine. We focus on the plan envelope and ignore the surround.

10. When the AI still gets a wall wrong — Reweave openings

Even with a flawless upload, a vision pass occasionally misreads a closet door for a wall opening, or paints a window where the plan shows a bare wall. Every analysis ships with a Reweave openings editor on the Arch. Plans tab — a per-room view of every door, window, and opening, with drag-to-reposition, rename, add, and delete all two clicks away.

If your renders keep surfacing the same misread opening, this is the recovery path. The editor saves back to the same plan; every downstream render — interior hero, walkthrough scene, PDF report — reads the corrected geometry on the next pass.

The short version

A clean upload makes the AI's job almost mechanical. Bold black walls, labelled rooms, at least one dimension, one plan per file — and the rest of the design lands the way it does in the demos.

Upload a floor plan and see what comes back. If it's not quite right, the Reweave overlay is one click away — and every note that lands in our inbox shapes the next iteration.