Elevations & Facades from Line Drawings

A line elevation is, in machine terms, almost an edge map already. The whole point of ControlNet's Canny and MLSD conditioners is to reduce an image to its significant edges and force the diffusion process to respect them; an elevation arrives pre-reduced. The floor lines, the window heads and sills, the parapet, the column rhythm - these are exactly the structural edges the conditioner wants, and because you drew them deliberately they carry real design intent rather than photographic accident.

This is why an elevation gives you far more predictable results than a plan. A plan has to be pushed through geometry and camera before the model can render a view; an elevation is the view, orthographic and honest. Feed a clean line elevation through an edge conditioner and the model treats your openings as fixed constraints, filling only the surfaces between them. The rhythm you laboured over stays put - which is precisely the part of a facade that a naive prompt would have scrambled. The lesson from Module 3 lands cleanly here: conditioning turns the model from a gambler into a collaborator that colours inside your lines.

There is a caveat that keeps this honest: an orthographic elevation is a flattened projection, so a straight conditioned render reads as an elevation-on, head-on view - useful for material study, less so for drama. If you want a three-quarter, eye-level hero shot you either draw the elevation already in slight perspective, or move to the massing workflow of the next lesson where the camera is free. Knowing which you want before you start - a flat, honest material study or a perspectival hero - decides whether an elevation is the right input at all.

Not every edge conditioner suits every elevation. Canny detects all significant edges and is the general workhorse - good when your elevation has texture, mullions and detail you want honoured. MLSD finds straight lines only, which makes it superb for crisp orthogonal facades - curtain walls, grids, regular fenestration - because it ignores hand-drawn wobble and locks onto the rectilinear order. Lineart and scribble conditioners are gentler, tracking your strokes more loosely, which is what you want for a looser concept elevation where you would rather the model interpret than obey.

Match the conditioner to the drawing and the intent. A hard-line CAD elevation of a glass office wants MLSD; a textured hand elevation of a stone haveli wants Canny or lineart. The control weight then sets how tightly the model obeys - high weight for 'render exactly this facade in travertine', lower weight for 'use this as a guide but surprise me with the materiality'. Learning to reach for the right conditioner and dial is most of the craft; the prompt is almost secondary once the lines are held.

One practical detail saves a lot of grief: preprocess deliberately. Most tools run your drawing through the detector automatically, but the preview of that detected edge map is the single most useful thing to inspect, because it shows you exactly what the model will actually obey. If the Canny preview is a mess of doubled lines from a hatched wall, or MLSD has dropped a subtle set-back you cared about, the render will inherit that error no matter how good your prompt is. Clean the drawing until the edge preview reads the way you want the facade to read, and half of your conditioning problems disappear before you have generated a single image.

With the composition locked by conditioning, your prompt stops fighting for the layout and does the job it is good at: specifying surface and situation. This is where a facade comes alive. Name the materials precisely - 'weathered Corten', 'fair-faced concrete with board marks', 'Jaisalmer sandstone', 'anodised bronze mullions', 'fluted glass' - because the model has rich associations for named materials and vague ones for adjectives like 'nice' or 'modern'. Name the glazing behaviour: 'reflective', 'transparent showing interior', 'spandrel'. Name the weather and time: 'golden hour raking light', 'overcast soft light', 'monsoon-wet, reflective pavement'.

The light angle is not cosmetic - it is how a facade reveals its depth. Raking, low-angle light throws the shadows that make reveals, fins and projections read; flat frontal light flattens everything to a drawing again. So your material-and-light prompt is really a depth-and-texture instruction, and pairing it with the locked lines is what separates a rendered facade from a coloured-in elevation. Keep the prompt disciplined - material, glazing, light, weather, context - and let conditioning carry the shape.

The classic failure of elevation-to-render is the cardboard cut-out - a facade that has your openings and your material but sits dead flat, as if printed on a sheet, because an orthographic elevation encodes no depth and the model defaulted to none. Real facades are layered: a back wall, mid-plane fins or jaali, projecting balconies and cornices, deep window reveals. If the drawing and prompt do not signal those layers, the render will not invent them convincingly.

There are two fixes, and you often use both. In the drawing, add depth cues the conditioner can see - shadow lines under overhangs, a hint of the returns on reveals, tonal difference between planes; or better, condition on a shallow 3D pull of the facade so the model gets an actual depth map. In the prompt, use the vocabulary of relief: 'deep window reveals', 'projecting balconies casting shadow', 'recessed entrance', 'self-shadowing brise-soleil', 'strong raking light'. Together these tell the model the facade is a set of stacked planes, not a single surface, and the shadows it renders will finally read as architecture rather than wallpaper.

Once the pipeline is smooth, its real gift appears: the material study becomes instant. Hold the elevation and seed fixed and swap only the material clause, and you generate the same facade in stone, then timber, then render, then perforated metal, in the time it once took to redraw a single option. This is a genuinely new capability for practice - a client conversation about materiality that used to need a week of visualiser time now happens live, with the composition guaranteed constant because the lines never move.

The discipline is the same one running through this whole module. Audit that the model has not silently added a floor of glazing you did not draw or invented a canopy; check that the material reads at the right scale (brick courses, panel joints, mullion spacing); and remember that a beautiful AI facade is a concept of the elevation, not a fabrication drawing.

Think about where this sits in a real project, too. A materiality conversation with a client used to be a bottleneck - each option cost visualiser hours, so you brought two or three and defended them. With a locked-line, seed-fixed workflow you can bring the client into the exploration itself, auditioning a dozen directions live and letting them react. That changes the relationship: the client stops choosing from a menu you pre-cooked and starts co-selecting from a space you steer, while you keep quiet control of the composition the whole time. Used this way - lines locked, materials swapped, depth coaxed, output audited - elevation-to-render is one of the most immediately useful things generative AI does for the built environment, and Studio Matrx's DesignAI is built to make exactly this loop fast.