Camera Samples Lab
One pixel. Several camera samples. One combined result.
Render view
32 × 32 · magnifiedInside one pixel
1 × 1 pixelFollow one camera sample
PINHOLESample 1 of 8
All these rays belong to the same output pixel. Only their image-plane positions differ.
What is happening?
Try this
The key distinction
Questions — click to reveal answers
Does taking 16 samples create 16 smaller pixels?
No. The image is always 32 × 32 output pixels. Multiple camera evaluations contribute to each pixel’s one stored colour. The enlarged square in the middle is one pixel’s sampling domain, not a higher-resolution output image.
Why does an edge pixel become grey?
In the edge experiment, each sample returns either the dark object value or the bright background value. If a pixel’s samples see both, averaging their values gives an intermediate result. This represents partial pixel coverage; it does not make the object transparent. The lab averages in linear light, then applies a fixed sRGB display conversion.
Is the camera physically sending light into the scene?
No. The diagram follows the renderer’s camera-ray tracing direction. The calculation estimates the light arriving at the camera. Here the objects return fixed values, deliberately avoiding additional lighting and shading noise.
Are camera samples the same as ray bounces?
No. These samples create primary camera rays. In a full renderer, a camera ray can lead to additional visibility checks and indirect-light paths. Bounce limits restrict path depth; they are not a count of camera samples. This lab stops at the first visible object or background and does not simulate indirect lighting.
What does Camera (AA) 3 mean in Arnold?
For the positive, fixed-sampling settings illustrated here, the camera count is squared: AA 3 gives 3 × 3 = 9 camera samples per pixel. AA also multiplies secondary sampling work in Arnold. Adaptive sampling can give different pixels different amounts of work. The drop-down is a count guide, not a replica of Arnold’s sampler or a render-quality recommendation.
Why can adding one sample change the dot positions?
The regular and jittered patterns rebuild their equal-area cells for the requested count. For square counts, those cells form a square grid; other counts may use rectangular grids. Use Random (seeded) to retain the existing sequence when increasing the count. Repeating the centre intentionally repeats exactly the same query. None of these patterns claims to match Arnold’s implementation.
Will more samples remove depth-of-field or motion blur?
No. Aperture and focus determine the intended defocus here; motion and shutter duration determine the intended motion blur. More samples improve the estimate of that blurred image rather than restoring sharpness. A camera sample can specify an image position, a lens position, and a time. These tabs vary just one of those domains at a time so their effects are easier to distinguish.
Does every increase in samples produce a better pixel?
Not necessarily. A particular estimate can move away from the exact value as more samples are used, and structured patterns can miss particular features. Random sampling generally reduces estimation error as more independent samples are averaged, but no individual step guarantees improvement. Compare the exact coverage reference in the edge tab and try different random sets.
About the model, controls and sources
This is a deliberately small, live mathematical camera-sampling demonstration, not an Arnold renderer. There are no downloaded images, noise overlays, external libraries or background network requests. The displayed sample results and rendered pixel averages are calculated from the same scene tests.
Edge: a pinhole camera samples image positions against an opaque diagonal half-plane; each output pixel uses an equal-weight box average. Depth of field: a thin-lens model samples a uniform circular aperture and intersects two flat objects at different depths; image positions are fixed at pixel centres to isolate lens sampling. Aperture changes are brightness-normalized here. Motion: a fixed pinhole ray samples a uniformly open shutter while a flat object moves at constant speed; image positions stay at pixel centres. Shutter changes are also brightness-normalized. These are isolation experiments, not a complete photographic exposure simulation.
The edge scene returns linear values of 0.025 for the dark object and 0.800 for the background. The far object in the depth experiment returns 0.120. These values are averaged before display conversion. A real renderer can also use weighted reconstruction filters, texture filtering, indirect-light sampling, adaptive sampling and denoising. Those are outside this lab.
Controls: click the render to choose a pixel; use the render’s arrow keys to move it. Click a dot or numbered swatch to follow a sample. Space on an ordinary button activates that button. Each tab starts a new experiment. Animation runs only when explicitly started, stops on tab changes, and pauses when the page is hidden. Camera diagrams are explanatory projections; the edge ray bundle is magnified, and depth of field uses a one-dimensional side projection of the aperture.
Sources support the camera and sampling concepts. The interface, simplified scene and sample-pattern implementation are specific to this teaching lab. Source links need internet access; the lab itself does not.