Sampling Lab / Samples & Noise
One scene. More estimates. A more stable image.
REAL SAMPLES · NO DENOISING
samples per pixel
Same scene. Different grain.
Pin the image at 4 samples, then choose New random samples. Repeat at 256 samples and compare how much the two estimates differ.
Move the light, restart the estimate IPR experiment
Moving the key light changes the scene. This lab clears the old average and starts at 1 sample. The materials, camera, fill light, and exposure stay fixed.
What is being calculated? A compact 3D ray tracer supplies real sample contributions. Play and +1 compute samples live; checkpoint buttons can recall embedded or previously calculated averages. Stored checkpoints are not a render-time benchmark. All views keep the same 480 × 336 resolution. This is an educational model, not Arnold.
Questions
Click a question to reveal the answer.Why does the image look grainy?
The renderer estimates how much light reaches each pixel. Different random samples can return different values. With only a few estimates, neighbouring pixels may disagree strongly. Averaging more samples generally makes those estimates more stable.
The lab does not add artificial noise on top of a clean image. The grain comes from the rendering calculation itself.
Are more samples the same as more pixels?
No. This image always contains 480 × 336 pixels. More samples give each of those pixels more estimates to average. Increasing resolution would create more pixels instead.
The magnified patch deliberately shows individual pixels without smoothing. It does not change the render resolution.
Are samples and bounces the same thing?
No. One sample can involve a path with several ray segments and light-visibility tests. Samples count estimates; a bounce limit limits how far paths can continue through surface interactions. This lab keeps the path settings fixed while the sample count changes.
Why does the last bit of noise take so long?
For independent Monte Carlo sampling with finite variance, typical error falls roughly as 1 ÷ √N. About four times as many samples are needed to halve the noise amplitude—not just twice as many.
This is a statistical rule of thumb, not a guarantee that every pixel improves after every sample. Sampling strategies and denoising can also affect a production renderer’s results.
Should more samples make the image brighter?
Not simply because there are more of them: the renderer uses an average, not an unscaled sum. At a low sample count the brightness estimate can fluctuate as bright or dark contributions arrive.
The lab averages linear-light values before applying its fixed filmic display curve. Averaging the displayed swatch colours would not be the same calculation.
Does 64 here mean Camera (AA) 64 in Arnold?
No. This lab counts estimates per pixel directly. Arnold’s sampling controls have their own meanings: the documented Camera (AA) value of 3 corresponds to 3 × 3 = 9 camera samples. Other settings affect additional sampling work.
Use the lab to understand the idea, not to copy its numbers into Arnold. See the Autodesk reference below.
Is this live rendering or a stored sequence?
Both are clearly labelled. Two random sets include genuine stored checkpoints at 1, 4, 16, 64, and 256 samples for instant comparisons. They were calculated using the same scene and sampler, with no denoising.
Play and +1 sample trace additional samples live. New light positions and other random sets also render live. The inspector regenerates actual contributions for the selected pixel; it is not displaying invented numbers.
Cached images and stored checkpoints should not be used to compare rendering speed. The status bar says which type you are viewing.
Why does moving the light reset the image?
The old samples estimate the old lighting. In this lab, changing the light discards that accumulated result and starts a new estimate. Mixing the old and new samples would mix two different scenes.
During playback, refinement resumes after the change. While paused, the new image waits at 1 sample until you continue.