Press 3. Watch the tapered tower you spent the morning on turn into a melted candle — the slab edges rounding off into nothing, one corner at level 27 pinching inward like someone's pulled a thread, the crown ballooning where you wanted a crisp arris.
Nothing is wrong with Maya. Everything is wrong with your topology. Smooth preview isn't a render setting, it's a calculation that runs off the cage you built, and it does exactly what the cage tells it to. A four-sided face surrounded by four-sided faces smooths predictably. A five-sided face next to a triangle does not. The fix isn't a smoothing slider — it's going back and rebuilding the handful of faces that are lying to the algorithm.
This is the bit that trips up architects coming from Rhino or SketchUp, where geometry is geometry and a surface is a surface. In poly modelling, the mesh is an instruction set. You're not drawing the form. You're drawing the rules the form will be derived from. Once that clicks, everything about Maya gets easier.
Maya's smooth mesh preview runs Catmull–Clark subdivision, and Catmull–Clark was designed around quads. Feed it a grid of four-sided faces and it subdivides evenly, each face becoming four, curvature flowing smoothly across the surface. Feed it a triangle or an n-gon and it still subdivides — but it creates a vertex where a number of edges other than four meet. Modellers call that a pole, or an extraordinary vertex.
Poles aren't illegal. Every closed form has some, and a sphere cap is nothing but poles. The problem is that curvature behaves differently around them. Light runs across the surface smoothly and then does something slightly odd at that one point, and on a 40-storey tapering tower that odd point gets repeated 40 times up the elevation, which is how you end up with a vertical scar down the corner of your building that you can see in every render and cannot find in the cage.
So the discipline is not "never use a triangle". It's put the poles where nobody will ever look. Under a slab. Behind a mullion. On the underside of a soffit. Inside the core. In practice, we hide them in the thickness of floor plates — the one place in a tower where a bit of curvature weirdness is invisible because the geometry is 200mm deep and in permanent shadow.
The other half of the discipline is edge loops. A loop is a continuous ring of edges that runs all the way round the form without hitting a dead end. On a tower, your loops want to be horizontal — one per slab, one per transfer level, one per setback — so that selecting a floor is one double-click rather than forty shift-clicks. The moment a loop terminates halfway along a facade because you extruded a balcony and left an n-gon behind, you've lost the ability to select, move or bevel that level as a unit. Loop continuity is the difference between a model you can edit at 11pm and a model you rebuild.
The Holding Loop Is the Whole Trick
Here is the single most transferable move in poly modelling for architecture, and it takes about four seconds.
Subdivision pulls geometry toward the average of its neighbours. A sharp edge with nothing near it gets pulled a long way, because its nearest neighbours are far away — that's your melted slab edge. Put another edge loop close to it, and the pull distance collapses. The edge barely moves. It stays sharp.
That's a holding loop, sometimes called a supporting loop. The rule we teach is two of them: one either side of any edge you want to read as crisp after smoothing. And the distance from the loop to the edge is the radius of the resulting arris. Tight loops give you a hard architectural edge. Loops set further back give you a generous rounded one. You're not approximating a fillet — you're controlling it, directly, with a tool you already know how to use.
On a tapered tower this is how you get slab edges that survive. Model the floor plate, bevel the outer edge once to give yourself the chamfer, then run a holding loop just inside each side of that bevel. Press 3. The chamfer holds. Press 1, nudge the loops a millimetre closer, press 3 again. It's a dial.
Maya will also let you crease an edge directly with the Crease tool, which looks like a shortcut and sometimes is. But a crease is an attribute on an edge, not geometry — and what we've found in practice is that creases are unreliable travellers. Export the mesh, open it somewhere else, and the crease has gone, taking your sharp edge with it. Holding loops are actual vertices in actual space. They go everywhere the model goes. Use creases to test an idea quickly; commit the ones you're keeping as loops.
Fix the Topology at One Floor, Not Forty
The reason a tower goes wrong is almost never the tower. It's that the mistake got replicated.
The workflow Nicholas Turki uses in the course's SOM tower case study — the Baltimore Tower at Canary Wharf — is to build a single typical storey and then drive the whole building with Duplicate Special, which can apply an incremental rotation and translation to every copy. One floor plate, a small rotation per instance, and the tower twists. It's an elegant bit of economics: the entire form is parametrically controlled by four numbers in a dialogue box.
It's also the moment of maximum exposure. Every flaw in that one floor gets built forty times. An n-gon at the lift core corner becomes a forty-storey column of bad shading. A holding loop that doesn't close becomes a chamfer that disappears at every level. And because the copies are generated, not drawn, you can't fix them individually in any sane amount of time — you go back to the original and run the whole thing again, which means redoing anything you did after.
So the sequence matters more than the tools. Build one floor. Clean it properly — every face a quad, every loop closed, holding loops placed on the edges you care about, poles buried in the slab thickness. Press 3 and interrogate that single plate as if it were the whole building, because it effectively is. Then duplicate. Ten minutes of topology cleanup on one floor buys you a tower that behaves, and skipping it buys you an afternoon.
Delete History, Freeze Transforms, Then Argue About It
Two housekeeping habits that feel like admin and are actually structural.
Delete history. Every bevel, extrude and boolean hangs around in the node graph as live construction history. It's useful while you're modelling — you can go back and change a bevel's width. It's a liability afterwards, because a stack of fifty history nodes on a duplicated tower is both slow and fragile, and a boolean that recalculates when you move something is how geometry quietly disappears overnight. Delete history on a floor plate once you're happy with its topology and before you replicate it.
Freeze transforms. If you've scaled an object non-uniformly, its local axes no longer match the world. Bevels and extrudes operate in local space, so a bevel on a non-uniformly scaled object comes out thicker on one axis than another, for no reason you can see. (And if you ever scale negatively to mirror something, you've flipped the normals — that one shows up as black faces in render and nowhere else.) Freezing resets the transform values to zero without moving the geometry. Do it before you export, always.
Getting It Out for Drawings
Here's the reframe that saves the most time, and it's counterintuitive: don't export the smoothed mesh.
Smoothed subdivision geometry is dense, triangulating on export, and useless for drawing. Section it in Rhino and you get a polyline made of hundreds of tiny segments, each one a face intersection — unusable for dimensioning, horrible to hatch, and it'll choke whatever you take it into next.
The cage is a better drawing than the smooth mesh is. It's low-poly, it's all quads, its edge loops sit exactly on your slab levels, and sectioning it gives you clean polylines with vertices where you put them. The smooth version is for the render. The cage is for the drawings. Export both, from the same model, and keep them honest by never editing one without the other.
On the way in, the same logic applies in reverse. Bring your setting-out from Rhino as curves, not as a mesh — in the poly modelling lesson Nicholas exports a plan out of Rhino as IGES specifically because that's the format Maya will take NURBS curves in, which is mildly surprising if you assumed 2D meant DWG. Import the curves, group them, lock the layer, and model to them.
When Rhino Still Wins — And It Often Does
I'd use Maya for the form and Rhino for the fabric, and I don't think that's a compromise.
Maya is unbeatable when the form is still an argument. Pushing a tower profile around, testing twelve crown options before lunch, driving a perforated envelope with blend shapes — a technique borrowed wholesale from facial animation, and the basis of the Tom Wiscombe hotel case study — these are things that are genuinely hard anywhere else. Subdivision is a form-finding medium. It rewards nudging.
But a subdivision surface has no radius. It has no centreline, no degree, no rational panel. The second someone needs to dimension it, panelise it, check planarity on a quad, or schedule 900 unique cladding units, you want NURBS and you want them in Rhino. Rebuilding a smoothed Maya mesh as a NURBS surface after the fact is a tax. Decide early which one the deliverable needs.
The rule of thumb I'd give: if the next person to touch it is a visualiser, stay in Maya. If the next person is a facade engineer, get it into Rhino before the form hardens.
All of this — poly modelling foundations, the holding-loop discipline, Duplicate Special on the SOM tower, blend-shape facades — is what our Autodesk Maya for Architecture course walks through properly, with the files, taught by someone who used Maya at Zaha Hadid Architects for five years.
Go back to your tower, press 1, and look at the cage instead of the form. Count the faces that aren't quads. That's your afternoon, and it's the shortest afternoon you'll spend on this building.
