Get the perspective right by building it

The problem a sketch cannot solve

Draw a room. Unless you know how to construct perspective, your walls will not converge correctly, and the model will faithfully reproduce your error. The result is a space that feels wrong in a way viewers notice without being able to name, and clients notice it precisely when the image is on a large screen in a meeting.

The fix is to stop drawing perspective and start building it. Software does perspective perfectly, for free, and you need about an hour of Blender to use it.

What you actually build

Not a model. A blockout: grey boxes standing in for things.

1. Open Blender. Delete the default cube or keep it as your first object.
2. Add planes for floor and walls. Add cubes for furniture, boxes, a counter. A cylinder for a bottle, a rough figure for a person: a capsule is enough.
3. Position the camera. This is the part that matters: set the focal length to the value you would have written in the prompt, and set the height to a realistic eye or table level.
4. Render a depth pass: in the Compositor, enable the Z pass, normalise it, and output it. Or use the Mist pass, which is easier and often good enough.

You now have a greyscale image where near is light and far is dark. That is a depth map, and it is exactly what depth conditioning consumes.

Why this is worth the hour

Perspective is correct by construction. The vanishing points are right because the software computed them.

The camera is a real camera. A 35mm lens at 1.2 metres produces the actual geometry of that setup, not an approximation of the word.

You can move it. Four angles of the same scene, all geometrically consistent, in two minutes. Try that with a sketch.

Scale is honest. Objects are the right size relative to each other, which is something generated images get subtly wrong all the time: the mug that is slightly too large, the chair that could not seat anyone.

It is reusable. The blockout is a file. Next month’s variation on the same scene starts from it.

Running it

Feed the depth map as a depth control at strength 0.5 to 0.8, ending around 0.7 of the steps. Then prompt the scene normally: materials, light, mood, finish.

Depth control preserves spatial arrangement and form and discards everything else. Your grey boxes become timber, marble, upholstery (whatever the prompt says) while staying exactly where and how big you placed them.

A practical detail: keep the depth range tight. If your scene spans two metres but the camera can see a distant wall fifty metres away, normalisation squashes everything of interest into a narrow band of grey and the control becomes weak. Either move the far clipping plane in, or use Mist with a manually set start and depth.

Where else this applies

• Product staging. A cylinder where the bottle goes, correct height, correct camera. Generate the environment around it, then composite the real product into the space you reserved.
• Architecture and interiors. The whole point. A rough massing model renders into a photographic visual.
• Consistent sets. Twelve images of the same room from different angles, geometrically consistent, which is a problem the consistency block otherwise cannot solve.
• Repeatable client revisions. “Move the sofa” is a thirty-second change in the blockout and a re-render, rather than a new hunt for a seed.

The honest costs

Blender is a serious application and its interface is not friendly on day one. What you need here is small (add a primitive, move it, set the camera, render a pass) and that is perhaps an hour of a tutorial. But an hour is an hour, and if you need one image this week it is not the right tool.

Depth is ambiguous about material. A depth map says nothing about what anything is made of, so if the prompt is vague the model will decide, and it will decide differently on every seed. Depth controls space, the prompt controls substance, and both have to be specified.

Depth discards edges. Two objects at the same distance blend into one grey region and may merge in the output. When that matters, stack a second control (the last lesson of this block) or separate them in depth by moving one slightly.

There is also a free alternative worth knowing about if Blender is too much: run a depth estimator over an existing photograph. A monocular depth model will produce a usable depth map from any picture, and it is one click in most interfaces. It gives you the perspective of a scene you photographed rather than one you designed, which covers a surprising number of briefs.