A model built at the wrong size renders exactly the same as one built at the right size. It only goes wrong later, when something has to fall through it, blur behind it or be lit from four metres up, and by then the repair costs a week rather than a minute.
The scene on this page is one still: a roadside bus shelter outside Kaunas at night, under a single sodium lamp, made over a fortnight for a book cover. Every number below came off that file.
The scene has a size, and four things downstream read it
Type the real dimensions in at the start. The shelter's opening is 2.05 m and the bench sits 450 mm off the ground, because those are the sizes such things are, and once they are right the rest of the scene has something to be wrong against. A file built to no particular size is not neutral, it is committed to a wrong one that nobody has written down yet.
Four things read that size and none of them announces it. Depth of field is computed from real distances, so a shelter modelled ten times too large photographs like an aircraft hangar with a bench in it. Physics runs on gravity in metres per second squared, which is why rain in an oversized scene falls like syrup and no amount of adjusting the drag will fix it. Light obeys the inverse square, so moving the lamp from 4 m to 8 m costs two stops, exactly as it would on a real pole. And every material with a real thickness, the perspex of the timetable case included, is guessing about that thickness in millimetres it was never given.
Noise costs four times what it looks like it should
Noise falls with the square root of the sample count, and that single relationship decides most of what an evening of rendering feels like. Four times the samples for half the noise. Not twice, which is what everybody assumes on the first long render, and the assumption is why people sit through a night of it and get a picture that is still slightly dirty.
On this machine, one frame at 64 samples takes 40 s. Doubling the samples doubles the wait: 1 min 20, then 2 min 40, then 5 min 20, and 1024 samples takes 10 min 40. Sixteen times the wait, between the first and the last, for a quarter of the noise. On a still that is an evening. On four hundred frames of animation it is the reason the piece does not exist.
The square root law
What each doubling of samples buys, and what it costs
The wait doubles with every column. The noise only halves every second column, and that mismatch is the whole economics of rendering.
Minutes for one 1920x1080 frame on one machine, path traced, with no denoiser. Samples double left to right.
- 40 s64
- 1 min 20128
- 2 min 40256
- 5 min 20512
- 10 min 401024
One machine and one scene, so the absolute numbers are this file's rather than anybody's. The shape of the curve is not: it is the square root law, and it holds on every renderer that traces paths.
The stall is the grey render
Almost every stalled 3D project looks the same: a competent model, a default studio environment, a mid grey background, and a maker who has decided the model needs more work. The default lighting is doing something specific and unhelpful. It lights everything evenly from everywhere, which flatters bad geometry and hides good geometry, so it makes an unfinished model look the same amount of finished as a finished one.
The consequence is a diagnosis error that costs months. The picture is disappointing, the only thing on screen is the model, so the model gets blamed, and the next fortnight goes into geometry nobody will ever see. One area light off to the side, a floor for it to bounce off, and a dark background will say more about what is wrong in ten minutes than a fortnight of modelling: where the light comes from, and what it says is the cheapest page in this topic to act on and the one people reach last.
What is holding what up
The order here is a dependency rather than a schedule, which is the part that catches people who have learned other crafts. Scale is underneath the whole stack. Topology sits on it, and quads everywhere it bends is the whole rule, because a triangle in a face that deforms pinches and a subdivision surface built on triangles ripples along the seams. UVs sit on topology, textures sit on UVs, lighting sits on all of it, and the render sits on top and simply reports whatever the layers below agreed.
Taken out of order the failure is always the same shape: work gets redone rather than corrected. Sculpting before topology is the classic one, which is what makes working as if it were clay and making the geometry two different jobs rather than two names for one, and it is why retopology exists as a stage people resent. Texturing before UVs is the same mistake one layer up, and giving a surface its material assumes that layout is settled. Making a model move reaches back down past all of it to the topology, which is the one thing on the list a beginner never expects.
One still, 38 hours
The part of the fortnight that ends up in the picture
The render and the lighting are the two stages whose result a viewer can point at, and together they are 7 hours of the 38. Everything under the line leaves no trace of itself in the picture.
- The render2 hours, and the only stage anybody watches
- The lighting5 hours, and the one that changed the picture most
- Texturing9 hours. The largest single number and nobody guesses it
- Look development7 hours of small material changes between test renders
- Retopology6 hours of rebuilding a surface that already existed
- Blocking the shelter4 hours, and the only part that feels like making something
- UVs3 hours of flattening, which is not creative and cannot be skipped
- Fixing the scale2 hours, and it would have been two minutes on the first day
This is one still by one person, so the split is a sample and not a benchmark. What generalises is which end the surprise is at: the stage people budget for is the one that took the least.
Where the fortnight actually went
The finished still took 38 hours and the render was 2 hours of the 38. Lighting was 5, and those two are the only stages whose result a viewer can point at: 7 hours of the 38, with everything else invisible in the picture it produced. Blocking the shelter took 4. Retopology took 6. UVs took 3, texturing 9, lighting 5, and look development, which is the loop of small material changes between test renders, took 7. Two more went on fixing the scale, because the shelter had been built to no particular size and the rain arrived before anybody noticed.
That distribution is normal rather than a sign of inefficiency, and it is worth seeing before the first project instead of after it. It is also the honest answer to why physics doing the animating has a reputation for eating time: the simulation itself was twenty minutes, and it was the two hours of scale correction underneath it that were expensive.
What to open in 2026
The denoiser is the change that matters, and it repeals the arithmetic above rather than softening it. OpenImageDenoise has been inside Blender's renderer since 2019, it is free, and it means a frame rendered at 64 samples and denoised is usable where it used to take 1024 samples to get there. It replaced rendering four times the samples for half the noise, which was the only method available before it, and it moved turning the scene into an image from an overnight job to a coffee.
Substance 3D Painter is the standard for the texturing layer and it is a subscription, with no way to buy it once, which is the honest statement of the cost. It replaced painting a texture in a flat image editor while guessing where the seams fell. Poly Haven is the third and it is the one nobody mentions: several thousand HDRIs, materials and models, released with no rights reserved, which replaced photographing a chrome ball in a car park to get a sky. The lamp lighting the shelter above is an HDRI from it, and that is a legitimate way to start a scene rather than a shortcut.
What none of this answers is when a scene is finished, and 3D is unusually bad at telling you. A photograph of the same shelter would be finished when the light went. Here every material can be adjusted again tomorrow, look development has no natural end, and the file will still open in a year, which is how a fortnight becomes a season. The trade being made is the same one light, a lens, and deciding what to leave out settles in a second at the shutter, and the discipline that answers it is on the timeline rather than in the scene, in graphics and characters that move, where a deadline is a frame count. For anyone modelling things that get made rather than rendered, the constraint arrives from outside instead, in drawing things that get built or manufactured.