Render Sampling Budget

What it is for

Use Render Sampling Budget before starting a local concept render. It compares two sample counts at the same image dimensions, helping you understand the extra work involved in a cleaner preview. The calculation describes Kerfwise Render Desk and takes inspiration from MoonRay's sampling guidance.

This tool does not measure your processor or predict elapsed time. Scene geometry, occlusion, finish, browser and device speed all affect the actual render. Treat its figures as a workload comparison and try a draft preview on your device before committing to the larger image.

Worked example

Enter width 640 pixels, height 480 pixels, eight samples per pixel and sixteen comparison samples. The first image uses 307200 pixels and 2457600 pixel samples. The comparison doubles that work to 4915200 pixel samples.

Under the independent-sampling model, the estimated noise ratio is approximately 0.707. That means about 70.7 percent of the original sampling noise, rather than half the noise. To halve the standard error in that model you need four times as many samples, which may exceed the local rendering budget at your chosen dimensions.

For a quick draft, use 320 by 240 at two samples and compare it with eight. This makes the effect of changing samples clear without also changing image dimensions. The preset names in Render Desk offer convenient starting points, but custom values remain available.

How to read the result

Pixel samples equal width multiplied by height and samples per pixel. Every sample traces one primary camera ray. The conservative secondary-ray bound reserves at most four additional rays per sample in this concept renderer; actual paths often use fewer.

The relative sampling-noise estimate follows standard error proportional to one divided by the square root of sample count, assuming independent samples and finite variance. It excludes bias, correlated samples and geometry aliasing. It is not a perceptual quality score.

The conservative image-buffer estimate reserves six floating-point RGBA accumulators and six byte RGBA outputs. Render Desk currently computes one pass at a time and allocates fewer image buffers. Geometry, browser, worker and PNG-export memory are additional.

Tips and common mistakes

Keep dimensions fixed when comparing samples, so you can attribute the work change to sampling alone. Increasing width and height also increases pixel count; doubling both dimensions multiplies pixel work by four before adding more samples.

Both choices must fit local limits: dimensions 32–800 pixels, at most 384000 pixels and six million pixel samples. Fractional sample counts are rejected. If a comparison exceeds the limit, reduce image dimensions or the target count.

Inspect the beauty, normals and light-visibility passes when diagnosing an image. More samples cannot fix incorrect geometry, an unsuitable camera or a poorly placed light.

FAQ

Is the noise ratio a guarantee?

No. It is an independent-sample statistical estimate, not a promised improvement for every surface, image or sampling pattern. Biased lighting approximations remain biased as sample count rises.

Does the buffer figure include everything?

No. It covers conservative image arrays for six passes. The actual worker handles one pass at a time, while geometry and browser overhead require extra memory.

Is this available to agents?

Yes. The JSON API and MCP tool kerfwise_render_budget accept direct fields width, height, samples and target. They use the same budget function as the local workspace.