Crystal vs Corona

At equal noise Crystal on an RTX 4080 was faster than Corona on a 24-core i9 48× on grass, 29× on a product shot, 11× on an interior and 4.2× on an exterior. With both denoisers on, the gap grows to 70×, 31×, 18× and 6.4×.

This page is a summary. The full report with noise curves and 1:1 crops for every scene is in Russian; the numbers and charts read the same.

No denoiser

Time to equal noise

scenenoiseCoronaCrystalfaster
Interior1.67 %103.1 s9.7 s×11
Exterior0.14 %358.2 s85.4 s×4.2
Product shot0.66 %210.4 s7.2 s×29
Vegetation2.46 %307.5 s6.5 s×48

Full render call time including scene preparation, at the lowest noise both renderers reached.

Denoiser 1.0

Time to equal cleanliness

sceneerrorCoronaCrystalfaster
Interior1.03 %55.7 s3.0 s×18
Exterior0.23 %169.9 s26.4 s×6.4
Product shot0.98 %108.7 s1.5 s×70
Vegetation3.81 %162.0 s5.3 s×31

Corona High Quality denoiser, Crystal OptiX. Error is the difference from a long denoised render by the same renderer. Corona's reference is noisier than Crystal's, so these gains are more likely overstated than understated.

Method

Equal noise, honest time

  1. Scenes open in their own units with the Corona settings they were made with. Sharpening, blur and Bloom & Glare are off for both: they change the grain being measured.
  2. Corona renders 4–64 passes, Crystal 64–4096 samples, 1024 px on the long side. Time is the full render call after one warm-up render.
  3. Every pass or sample count is rendered twice with different sampling patterns; noise is the mean difference of the pair divided by √2. Each renderer's noise is measured against its own picture, so the difference in look between the engines stays out of the number.
  4. Time to equal noise is read off Crystal's curve between two measured points, on a log scale, with no extrapolation.
CoronaIntel Core i9-13900KF, 24 cores, 32 threads. Scene settings, High Quality denoiser.
Crystal 0.1NVIDIA GeForce RTX 4080, 16 GB, OptiX. Default GI cache, light tree and portals; adaptive sampling off.
3ds Max 2024Windows 10 Pro, one machine, rendered from 3dsmaxbatch with nothing else running.

Scene 1 of 4 · 768 × 1024 px

Interior: Scandinavian kitchen

Daylight through a window: CoronaSun, Corona lights, 1569 objects, textured wood, tiles and fabric, glass and plants. Scene by Vyacheslav Kungurov.

Corona, 9.3 s, denoisedCrystal, 8.9 s, denoised Corona · 9.3 sCrystal · 8.9 s
Denoised: Corona 2 passes, Crystal 512 samples. Drag the divider.

Scene 2 of 4 · 1024 × 819 px

Exterior: residential complex

An evening sky, motion blur in both renderers: CoronaSun and HDRI, 214 Corona lights, decals, Chaos Scatter and Forest Pack, 1864 objects. Scene by Vyacheslav Kungurov.

Corona, 37.4 s, denoisedCrystal, 40.0 s, denoised Corona · 37.4 sCrystal · 40.0 s
Denoised: Corona 8 passes, Crystal 1024 samples. Drag the divider.

The car in front is blue in Corona and brown in Crystal, and Crystal's road is slightly purple: a material translation difference that does not affect the noise measurement.

Scene 3 of 4 · 1024 × 1024 px

Product shot: baskets

A still life on black: wicker baskets, wheat, apples, candles in glass, a maple leaf; 3 lights, 22 objects.

Corona, 3.2 s, denoisedCrystal, 3.3 s, denoised Corona · 3.2 sCrystal · 3.3 s
Denoised: Corona 1 pass, Crystal 128 samples. Drag the divider.

Scene 4 of 4 · 1024 × 1024 px

Vegetation: meadow grass

Dense meadow grass with clover in sunlight, 61 objects: thin blades, light through them.

Corona, 14.5 s, denoisedCrystal, 14.3 s, denoised Corona · 14.5 sCrystal · 14.3 s
Denoised: Corona 1 pass, Crystal 512 samples. Drag the divider.

The pictures differ visibly: Crystal's grass is greener and flatter, without Corona's yellow sunlit blades. Part of the speed gap comes from that.