The image keeps learning — Parallax Sketchbook

SKETCHBOOK / 07 The Lantern Room · 2-week chapter

Page 1 of 3: A sample count belongs to an image state

Accumulation / invalidation

A sample count belongs to an image state

A progressive pixel stores a linear radiance sum or mean and the number of contributing samples. Its state also depends on the camera, scene, material, integrator settings, and sample policy. When one of those changes, old samples no longer estimate the same image.

Use a clear generation or revision boundary to reset accumulation. Changing exposure can often re-display existing linear values; changing the camera cannot. Record which settings alter the estimator and which alter only presentation before adding interactive controls.

The companion Moments class demonstrates bounded online scalar statistics, not a full progressive image scheduler. Your learner renderer owns its accumulation buffer and defines reset behavior. Test that a camera change clears the sample count instead of blending two unrelated views.

LOOK CLOSER

#include "lantern.hpp"
#include <expected>

std::expected<double, lantern::Error> linear_mean() {
    lantern::Moments estimates;
    if (auto a = estimates.add(1); !a) return std::unexpected(a.error());
    if (auto a = estimates.add(3); !a) return std::unexpected(a.error());
    return estimates.mean(); // 2, before exposure or display encoding.
}

Actual scalar companion. The diffuse bridge has a batch image buffer; a persistent progressive buffer and reset policy are your extension.

From the page to your own work

Your studio assignment.

Give the Lantern Room a reliable progressive checkpoint.

A finish line you can see

An accumulation/reset test, a linear-color test, and an honestly labeled convergence sheet.

Keep count and image revision together.
Test camera reset and presentation-only exposure change.
Reject non-finite accumulation input without changing prior statistics.
Document seed/work/exposure for every comparison.

Keep the evidence in your weekly notebook ↗

Explain it like an engineer

Try answering aloud before opening the review notes. Use your own implementation as evidence.

Why can a nearly noise-free image still be the wrong solution?

Estimator bias or missing transport may remain.
Display processing can hide variation.
A variance statistic cannot replace an independent correctness test.

These prompts support practice; they are not an automatic assessment of your work.

A tip for this exercise

Keep a tiny diagnostic crop and its linear statistics alongside every convergence image. Record image revision, count, seed, and exposure so you can explain a reset or brightness change without guessing.

Hints, when you need them.

Try a first step, then reveal as much help as you need. Each hint gets a little more specific.

1Hint 1 · Find a starting point

Sort your controls by whether they change the light estimate or only its presentation. Decide which changes make previously accumulated samples belong to a different image.

2Hint 2 · Choose an approach

Store image revision and sample count with the linear accumulator. Validate a new contribution before committing its updates, and put exposure and display conversion after the mean is formed.

3Hint 3 · Check your reasoning

Samples 0 and 4 should have linear mean 2 at count 2. Re-display at another exposure without changing that count; a camera revision resets it, and rejected non-finite input leaves prior statistics unchanged.

A small creative variation ↗

Compare the same accumulated linear buffer at two exposures without retracing it.

Optional. Save it for later if this chapter already feels full.

Coming back after a break? ↩

Read the current sample count and render one small crop with a fixed seed.

Books & principles behind this study +

Original Parallax explanations and assignments, with these works for further reading.

YOUR MARGIN NOTES

Leave yourself a good question.

What changed your mind? What would you try next?

Notes for The image keeps learning

Notes are private to this browser and saved as you write.

A CHECKPOINT, NOT A SCORE

Can you explain the idea and try the exercise yourself?

Completion records your own assessment. It does not compile or test your C++.

Companion source 10 files