Enhancing a Photo by Comparison

The preference chapters taught their methods on designs small enough to see whole: a hue on a color wheel in Chapter 19, a flat poster landscape with two knobs in Section 20.2. Real design tasks are not like that. A photograph has a dozen adjustments that interact, nobody can write down the function that scores them, and the only instrument that can measure how good a version looks is a person, who gets tired, changes their mind, and sees a different picture on every screen.

In this chapter the problem is real because you are part of it. The figures show a real photograph, and you enhance it by comparing versions: either by choosing the better of two, or by sliding along a line through the space of adjustments, the two query forms that Koyama et al. (2017) and Koyama et al. (2020) studied for photo color enhancement. The algorithm is the one built in Chapter 18 and Chapter 19, unchanged, running on six parameters instead of one or two. We first describe the problem and what a single judgment costs, then the decisions that have to be made before the first question, then the two query forms on the photograph, and finally what a session reveals about the person in it, set against what the published studies found.

25.1 The problem #

25.1.1 A photograph and six adjustments #

The photograph was taken by Aleš Krivec on 19 October 2014 at the Zelenci nature reserve in Slovenia and dedicated to the public domain under CC0 1.0 (Wikimedia Commons, "Zelenci 2014 (2).jpg", first published on Unsplash). It was chosen because its look depends on every adjustment we will offer. It has a bright sky and a lit mountain face that clip if pushed, a dark forest whose detail lives in the shadows, golden grass that turns orange or olive with the white balance, and turquoise water that a warm setting dulls. Most changes that improve one region hurt another, so there is no setting that every viewer will call best.

A design is a setting of six adjustments, each a number $u_i \in [0, 1]$, with $u_i = 0.5$ the untouched photograph. Table 25.1 lists what each one does and the range it spans. Mapping every adjustment to the unit interval is the usual convention, also used by the photo enhancer in the study of Ou et al. (2023); it lets one kernel lengthscale mean roughly the same thing in every direction.

Before any optimization, try the problem the way a photo editor presents it: six sliders.

Spend a minute with it. The sliders are easy to move and the result is hard to settle: improving the water can make the grass worse, more exposure calls for less shadow lift, and soon it is unclear whether the last change helped. Koyama and Igarashi (2018) describe the same experience: sliders affect the picture "in complex ways", so a designer explores by trial and error while building a mental model of the design space, and as parameters are added the exploration becomes "very tedious". That is the difficulty the rest of the chapter hands to an optimizer.

Three of the adjustments are multiplicative, so their ranges are symmetric on a logarithmic scale: halving the saturation is as large a change as doubling it. The hyperparameters of the classifier in Section 22.1.3 are searched on a log scale for the same reason. The adjustments also interact. Exposure and shadow lift both brighten the picture, so too much of one can be undone by less of the other. Saturation and contrast both add punch, and temperature and tint both move the white balance. The utility of a design is therefore not a sum of six separate preferences, which is one reason a Gaussian process over all six together is a sensible model.

The figures apply the adjustments in the browser, with an SVG filter on the image: a lookup table for the tone curve (exposure, shadows, contrast), then a color matrix (saturation, temperature, tint). The same computation on an array of pixels takes a dozen lines.

import numpy as np

def enhance(img, u):
    """img: H x W x 3 sRGB values in [0, 1]; u: six numbers in [0, 1]."""
    ev, c, s, t, m, a = (np.asarray(u) - 0.5) * 2   # each in [-1, 1]
    k = 2.0 ** ev                                    # exposure in stops
    y = k * img / (1 + (k - 1) * img)                # scale dark tones, keep 0 and 1
    y = y + 0.14 * a * y * (1 - y) ** 2 / (4 / 27)   # lift or deepen the shadows
    y = np.clip(0.5 + 2.0 ** (0.6 * c) * (y - 0.5), 0, 1)   # contrast
    luma = (y @ np.array([0.2126, 0.7152, 0.0722]))[..., None]
    y = np.clip(luma + 2.0 ** s * (y - luma), 0, None)      # saturation
    gain = np.array([1 + 0.11 * t + 0.03 * m,        # temperature: red up, blue down
                     1 + 0.015 * t - 0.07 * m,       # tint: green down
                     1 - 0.13 * t + 0.03 * m])
    return np.clip(y * gain, 0, 1)

The filter is deliberately simple. A full photo editor offers far more, for instance separate color balance for shadows, midtones, and highlights; Sequential Gallery's photo enhancer had 12 parameters, including that three-way color balance (Koyama et al., 2020). Six is the number the 2017 sequential line search used, with brightness, contrast, saturation, and a red, green, and blue color balance (Koyama and Igarashi, 2018).

25.1.2 Why the eye is the only objective #

An engineer's first instinct is to find a formula. Image-quality metrics and automatic enhancement exist, and every photo editor has an auto button. But an automatic enhancement encodes somebody's taste, learned from somebody's data, and the question here is what this viewer wants. When Koyama and Igarashi (2018) asked crowd workers which of four versions of a photograph looked best (the original, the result of their crowd-driven optimization, and the automatic enhancements of Adobe Photoshop and Lightroom), the optimized versions of three photographs received 32, 26, and 29 votes; the originals received 0, 2, and 0, Photoshop's versions 0, 2, and 1, and Lightroom's 1, 3, and 3.

The objective also depends on the purpose, which is set by the instruction, and the published studies set it differently. The crowd workers of sequential line search were told simply to adjust the slider until the image looked best (Koyama et al., 2017). In a pairwise-comparison task from the same line of work, workers chose the photograph "that would be better to use in a magazine or product advertisement" (Koyama and Igarashi, 2018). The participants of Sequential Gallery imagined uploading the photographs to their Facebook or Instagram accounts and making them appealing to their friends (Koyama et al., 2020). These instructions need not share an optimum (inference). The figures below ask only which version you prefer; decide for yourself what the picture is for, and keep that fixed.

Asking a person for a number would not work well either. A score on a scale from 1 to 10 requires knowing what the rest of the design space looks like, which a newcomer does not, while a comparison "is easy to answer even for non-experts" (Koyama and Igarashi, 2018); Section 16.1.1 collects the evidence. So the measurement is a comparison, the person is the instrument, and every property of that instrument (its noise, its drift, its patience) becomes a property of the optimization problem.

25.1.3 What a judgment costs #

A judgment costs seconds, not hours. In the closest published analogue, 11 professional designers compared color variations of banner ads, 50 pairs per task on spaces of 12 and 6 parameters, and took 4.2 seconds per choice on average, ranging from 1.0 to 23.1 seconds (Iwai et al., 2025). A richer query takes longer: one plane of Sequential Gallery took 14.8 seconds on average (Koyama et al., 2020). For a single slider, the nearest measurement comes from a different task: when three participants steered a generative image model toward a reference image, five reference images each, one iteration took 17.2 seconds on average with one slider and 53.4 seconds with four (Chong et al., 2021). Neither the 4.2 seconds nor the 17.2 seconds was measured on photo enhancement. For a slider against a pair on a photo task we found only a relative measurement: in crowdsourced color enhancement toward a reference image, a slider task took longer than a task of choosing between two versions, but less than twice as long, with the time for reading the instructions and for repeated quality-control questions included (Koyama et al., 2017).

The binding constraint is attention. Participants in the Sequential Gallery study pressed a "satisfied" button after 5.36 iterations on average (standard deviation 2.69) (Koyama et al., 2020), and in the three-month field deployment described in Section 32.3 most evaluation sequences never reached the optimizer at all (Ou et al., 2022). The crowd version of sequential line search bought its patience: 15 iterations of seven slider tasks each, paid 0.05 USD per task (Section 20.2.3). A personal session has perhaps ten to thirty answers to spend, the same order as the comparisons in Figure 19.1, but now in six dimensions.

Two further costs never appear in a test function. Every answer is noisy: the same person, shown the same pair twice, will sometimes answer differently, which Section 16.3 models as noise on each option's perceived utility. And the instrument includes the display. Sequential Gallery's study fixed it, a 13-inch MacBook Pro with a mouse (Koyama et al., 2020); you are reading this on a screen of unknown brightness and color, in light we cannot see, and a version that looks best on a phone in daylight may look garish on a monitor at night (inference).

25.2 Before the first question #

Every choice in this section is made before any data arrives, and each one changes what the optimizer can find. Table 25.2 puts the choices of three published photo-enhancement systems next to those of the figures in this chapter.

The sources for the table are Koyama et al. (2017) and Koyama and Igarashi (2018) for the first column, Koyama et al. (2020) for the second, and Ou et al. (2023) for the third. The rest of this section explains the choices in the last column, and where they differ from the published ones, why.

The box. The ranges in Table 25.1 decide what the optimizer can reach. A box that is too narrow excludes the look the person wants; a box that is too wide fills most of its volume with absurd pictures, and in six dimensions volume is everything. A design drawn uniformly from our box is almost always worse than the original, because the original sits at the center and most random settings push at least one adjustment to an extreme. Figure 25.4 below shows the consequence: twenty random pairs leave a simulated viewer with a recommendation no better than the untouched photo. The ranges here were set by eye to span from "clearly too little" to "clearly too much" for this photograph.

The kernel. The utility model is the Gaussian process of Chapter 18 with a Matérn 5/2 kernel, a lengthscale of 0.4 in every direction, prior variance 1, and probit noise $\sigma = 0.5$, fitted with the Laplace approximation (Section 18.2). The values are fixed, not learned from the answers. Learning six lengthscales (Section 9.2) from one-bit answers at a budget of twenty is fragile, and Section 30.3.3 reports that no study has checked whether it is even possible at such budgets. Koyama and colleagues reached the same practical answer by another route: their log-normal priors on the kernel hyperparameters, with $\sigma^2 = 0.01$, leave the lengthscales almost no room to move (Koyama et al., 2020). We chose 0.4 and 0.5 by running simulated sessions, which is the honest way to set such values before a real person sits down.

A visible difference. The acquisition function of Section 19.4 knows nothing about eyes. Left alone, EUBO often proposes two versions so close that a person cannot tell them apart, a symptom of the collapse toward the current best described in Section 19.6. The authors of a study of vibrotactile preferences name the same problem for touch among their limitations: very similar signals "may be perceptually indistinguishable", and comparisons between nearby points "are dominated by response noise" (Zhang et al., 2026b). The figures therefore only consider pairs whose two settings are at least 0.2 apart in Euclidean distance, a crude stand-in for a just-noticeable difference in this space. Answers of "about the same" (Section 20.4.1) are the principled alternative; we kept the interface to two buttons.

The candidate pool. In one dimension, Figure 19.1 scored every pair of 36 hues. In six dimensions there is no grid to enumerate, so before each question the figure draws 160 settings uniformly from the box, 60 small random perturbations of the current best guess, and every setting compared so far, then scores every pair in this pool of up to about 250 settings with the closed form Equation (19.3), roughly 30,000 pairs per question. The local perturbations matter: a uniform point lands within 0.1 of the best guess in every coordinate with probability $0.2^6 \approx 6 \times 10^{-5}$, so without them the pool would almost never contain the small refinements that the end of a session needs.

The recommendation. The best guess is the compared setting with the highest posterior mean, the rule sequential line search uses for its $\vx^+$ (Koyama and Igarashi, 2018). Restricting it to settings the person has actually seen means the recommendation is never a picture nobody looked at.

The start. The first question shows the original against a random edit. The original stays in the data, so every later answer is connected to it through the comparison graph (Section 18.5), and the figure always shows the original beside the best guess. One participant in the Sequential Gallery study asked for exactly that, to "have the original photo alongside to compare with during the enhancement", and the same participant found the initial photograph of one task already satisfactory and was unsure whether to press the "satisfied" button at the very start (Koyama et al., 2020). A system that cannot recommend the original is biased toward change (inference).

25.3 Comparing versions #

Here is the loop with you as the oracle. Pick the version you prefer, a dozen times or more. Take each choice at face value; there is no right answer.

The two large pictures are the question. Below them, the figure shows the model's current best guess beside the original, so you can always check whether the session is improving on doing nothing. The six small panels show what the model has learned: each is a slice of the posterior utility (Chapter 18) along one adjustment, with the other five held at the best guess, so a peaked curve says "this adjustment matters to you, and here is where you like it", and a flat curve with a wide band says "no information yet". The letters A and B mark where the two versions of the current question sit on each adjustment, which shows what the question is probing.

Some things to try. Answer ten or fifteen pairs and watch which panels sharpen first: they are the adjustments your answers have been most decisive about. Watch the letters A and B in the panels. The first question changes every adjustment at once; later questions tend to hold most adjustments nearly equal and vary a few, so each answer probes a smaller part of the space (in simulated sessions, the two versions differ by more than 0.1 in all six adjustments at the first question and in two or three by the tenth). After a dozen answers, compare the best guess with the original and ask yourself honestly which you prefer. Then press Ask an earlier pair again a few times and answer as you normally would; we will come back to what it measures in Section 25.5.1.

Two cautions about reading the panels. A slice is not the whole story: it shows how the utility changes along one adjustment at the best guess, and with interactions between adjustments the best exposure at a different shadow setting may differ. And the vertical axis has no units. Comparisons identify utility differences only up to a shift and a scale (Section 18.4), so what can be read off is the shape of each curve and the relative heights of the six, not a number.

Switch to Simulated taste to watch the same loop with a hidden utility answering. The simulated taste has a favorite setting and a utility that falls off quadratically from it, with different weights per adjustment (exposure and temperature matter most, tint least) and the exposure and shadow interaction described above. Press Simulate five a few times, then reveal the favorite: its picture replaces the original beside the best guess, the dashed lines mark its settings in the panels, and the status line reports the utility gap of the best guess, the simulated utility at the favorite minus the utility at the best guess, next to the gap of the original. A gap of zero is perfect.

25.4 Searching along a line #

A pair gives the model one bit. A slider gives it a whole one-dimensional search, performed by the person: they move along a line through the design space, watch the picture change, and stop where it looks best. This is sequential line search (Koyama et al., 2017), described in Section 20.2, now on the real photograph.

The figure departs from the published method in two ways, and both are choices a practitioner faces.

Which line. Koyama et al. (2017) run the slider from the best design so far, $\vx^+$, to the design with the highest expected improvement over it, $\vx^{\EI}$ (Section 20.2.1), and mention enlarging the segment by a fixed factor such as 1.25 as an option. We use the same two points but extend the segment all the way to the edges of the box in both directions. The person can then ask for less of the proposed change as well as more, and the slider covers enough range for the differences to be visible. Before any answer, the line runs through the original and the same random edit that opens the pairs figure. Whether the extension helps depends on the noise, as the comparison below shows. One published detail we kept: the slider starts at a random position for every question, which Koyama et al. (2017) did "to reduce cognitive bias", the pull of wherever the knob happens to rest (an anchoring effect, Section 37.2.3).

How a choice is recorded. Koyama and colleagues record a slider answer as a choice from three options, the chosen design against both ends, with the choice likelihood of Equation (20.1) (Koyama et al., 2017; Koyama et al., 2020). Our model has only the pairwise probit likelihood of Chapter 18, so the figure records the choice as up to four separate comparisons: the chosen point beats both ends and the points at one third and two thirds of the slider. This overstates the evidence. The four comparisons come from one act of judgment, so their errors are not independent, yet the model treats them as four independent answers and becomes more confident than it should (Exercise 25.2). Mikkola et al. (2020) take the opposite view, that a slider answer is uncountably many comparisons, and approximate the resulting likelihood with a finite sample; both records discard the fact that the person saw the whole continuum at once.

Some things to try. Drag slowly along the strip and notice how much of the line is clearly wrong; most of the information in your answer is in which region you avoided. After a few answers the line gets shorter in the panels, because the expected-improvement point moves closer to the best guess. Move the knob to the star and choose it, and the answer says that the best design so far beats everything on the line, which is a strong statement.

25.4.1 Pairs or a slider? #

The two forms trade information per answer against time per answer. The published comparison comes from the crowd. Koyama et al. (2017) ran their slider against a choice between two versions and a choice among four, in crowdsourced photo color enhancement toward a reference image, three runs per condition. The workers matched a target picture; they did not state a preference. The slider's error fell faster over the iterations, and a slider task took longer than a comparison task, but less than twice as long. As of September 2026 we found no study that compares the two forms for one person's own preference on a photo task, with the same model and budget. The precomputed simulations below run four query strategies (EUBO pairs, random pairs, the published slider segment, and the full-line slider) sixteen times each against the simulated taste of Figure 25.2.

Per answer, the slider is ahead early and the forms converge later. At the default noise ($\sigma = 0.1$), after five answers the median gap is 0.37 for EUBO pairs and 0.34 and 0.30 for the published and the full-line slider; after twenty answers it is 0.14 for EUBO pairs, 0.08 for the published slider, and 0.14 for the full line, against 0.49 for the untouched photo. At low noise ($\sigma = 0.05$) the three end within 0.02 of each other. At high noise ($\sigma = 0.25$) the full-line slider holds up best, with a median gap of 0.16 after twenty answers against 0.31 for EUBO pairs and 0.34 for the published segment. Random pairs are a warning: after twenty answers at the default noise, the recommendation beat the untouched photograph in only 7 of 16 sessions, so here EUBO pairs are far ahead of random ones. The recorded runs of Section 19.5.1 show the opposite ordering, with EUBO no better than random pairs after about twenty answers, under a different setup: generated designs with a different simulated taste, no minimum distance between the two options of a pair, and a best guess taken from the whole candidate pool rather than from the compared settings; we have not separated which of these differences accounts for the reversal (inference).

Per second, the picture reverses, if the borrowed timings of Section 25.1.3 hold for this task. Switch the axis to Seconds, which charges each pair 4.2 seconds and each slider answer 17.2 seconds: in the 84 seconds that twenty pairs then take, a slider session gets through about five answers, and its median gap is still about 0.3 while the pairs are at 0.14. The simulated slider is also generous to the slider: the simulated person evaluates 41 positions with independent noise and keeps the best, a careful search that a real person may not perform in 17 seconds. The reversal also rests on the assumed ratio of about four between the two times. If a slider answer took twice as long as a pair, the upper bound that Koyama et al. (2017) report for crowd tasks, a slider session would give ten answers in those 84 seconds and reach a median gap of 0.17 to 0.18, close to the pairs' 0.14 (inference). So the simulation does not settle the question. It shows instead which numbers would settle it, the time per answer and the noise of a slider answer compared with a pairwise one, measured on the same people and the same task (inference). A 2026 study that compared interfaces on one task, GimmBO, had 12 participants match a target image by merging image-model adapters: its ranking interface, driven by Bayesian optimization, reached a higher similarity to the target and a higher success rate than a panel of sliders adjusted by hand, at 50.5 against 34.7 seconds of interaction per step (Liu et al., 2026b); Section 32.4 reports it with the other comparisons of forms.

25.5 What a session reveals #

Run a session on yourself and three things become visible that a test function hides: you do not always give the same answer twice, your taste may move while you answer, and the result depends on which questions you happened to be asked. Each has a counterpart in the published studies, and for each the studies say less than one would like.

25.5.1 Consistency #

Ask an earlier pair again shows you a pair from earlier in the session with the sides swapped, and the history then marks your answer same or flipped. Under the model of Section 16.3 the probability of a flip has a simple form. If the person prefers A to B with probability $p = \Phi\big(\Delta / (\sqrt{2}\,\sigma)\big)$, where $\Delta$ is the utility difference and $\sigma$ the noise on each option, then two independent answers agree with probability

Agreement is never below one half, it reaches one only for pairs whose difference is large compared with the noise, and for a pair the person finds nearly equal it is close to a coin flip. With $p = 0.8$, a fairly clear preference, two answers agree only 68% of the time. So a flipped answer is not by itself a sign of carelessness. A high rate of flips on pairs the model considered clear is a sign that the model's noise scale is too small for this person, and that its posterior is overconfident (Exercise 25.1).

Equation (25.1) assumes that the second answer is independent of the first, and memory can break that in either direction: a remembered answer can simply be repeated, and the effort to recall it can add noise of its own. In experiments on repeated beauty ratings, recall memory contributed half of the measured variance when a single stimulus was rated, but added only about 10% to the variance when the set had nine or more stimuli (Pombo et al., 2023). The figure therefore swaps the sides, never announces a repeat, and only repeats pairs that are at least two answers old; a careful study would space the repeats further apart.

Repeating pairs within a session is cheap, yet as of September 2026 we found no study of preference-based design optimization that did it, re-asking early comparisons later in the session to test consistency or drift (Section 32.3). The closest evidence is indirect. When 17 sighted participants personalized simulated retinal-implant vision, only about 50% of their choices agreed with the predictions of a simulated agent (Schoinas et al., 2025), so a simulated person can be a poor stand-in for a real one. In the photo task of Ou et al. (2023), participants with more photo-editing experience expressed indifference, placing versions in the same rating region, significantly more often than novices ($p = 0.015$): the experienced eye declared more ties, not fewer.

25.5.2 Drift #

A taste can move during a session: by adapting to vivid versions, so that the original starts to look dull; by noticing something, such as the color of the water, that was not a criterion at the start; or by fatigue. Koyama et al. (2020) assume that the user's perceptual function "does not change over time", yet one of their six participants reported selecting designs "based on criteria that I didn't have at the beginning". The model of this chapter has no notion of time. It treats an answer given in the first minute and one given in the tenth as two noisy readings of the same utility.

Turn on Taste drifts during the session in Figure 25.4. The simulated favorite now moves steadily over the twenty answers, toward more saturation and warmth and less contrast, by about 0.3 in all. The gap curves flatten, because the target is moving, and the bottom panel shows where the sessions ended relative to where the favorite started and where it ended. At the two lower noise levels they land between the two: projected onto the drift, the final recommendations sit on average 13% to 56% of the way from the starting favorite to the final one, depending on the query strategy. At the highest noise, the EUBO sessions show no movement toward the new favorite at all. A model that assumes a fixed utility averages over the session, so it recommends what the person liked on average, not what they like now (inference, from these simulations). The remedies, a utility that changes with time or a forgetting factor on old answers, are discussed in Section 46.5; none has been tested in a photo task.

The drift in the figure is invented for illustration. We do not know how fast, or even whether, a viewer's preferred saturation moves in a session of twenty comparisons, and the experiment that would tell, re-asking early pairs at the end with the order of questions randomized, is the one described in Section 47.4.

25.5.3 Order #

Even with a fixed taste, the questions a person happens to be asked change where the session ends. The bottom panel of Figure 25.4 shows the final recommendations of sixteen sessions with the same simulated taste and the same noise; they differ only in their random candidate pools, their first random edit, and the noise in individual answers. At the default noise the final settings scatter with a standard deviation of 0.05 to 0.12 per adjustment, on the unit scale of Table 25.1; 0.07 in exposure is about 0.14 EV. The scatter is smallest for exposure, the adjustment the simulated taste weighs most, and among the widest for tint, which it weighs least. That is the posterior working as intended: where the utility is flat, the answers cannot pin down a setting, and it costs little to be wrong.

The published evidence on order comes from the crowd. Koyama et al. (2017) repeated sequential line search three times on the same photograph from different random starting conditions and measured the differences between the three runs over the iterations, both as a distance between parameter settings and as a mean perceptual color difference per pixel: the differences shrank rapidly in the first four or five iterations, and the three runs approached similar enhancements. The authors took this as support for their assumption of a goodness function shared by the crowd, and added that it may not hold where some people prefer one style and others another. Each iteration's answer, however, was the median of at least five workers' slider positions, which averages away much of any one person's noise. How much a single person's result depends on the order of their questions has not, as far as we found, been measured in a photo-enhancement session. Studies of judgment give reasons to expect an effect. With artistic photographs and faces, a neutral picture was preferred more after a preferable one than after a less preferable one, and a similar trend remained when response bias was controlled (Chang et al., 2017). In seven experiments on beauty ratings, sequential dependence did not affect the variance of the ratings as long as the stimuli were diverse, drawn from different object categories (Pombo et al., 2023). Versions of one photograph are the opposite of diverse, so the reassurance does not carry over to this task (inference); Section 37.3.2 reviews these sequential effects.

25.5.4 Stopping, and whose result it is #

A session ends when the person stops, and the published numbers show how early that is. Sequential Gallery's participants were satisfied after 5.36 planes on average (Koyama et al., 2020). In Ou et al. (2023), with five photo adjustments in $[0, 1]$, four rounds of quasi-random versions before the optimizer took over, and at most 20 iterations, participants with more photo-editing experience ran significantly more iterations than novices ($p = 0.008$), and the authors summarize the study as finding that novices were more easily satisfied and finished sooner, while experts explored further and reported lower satisfaction. Satisfaction is not convergence: Section 46.6 explains why a person's "good enough" and a model's flattened posterior are different stopping signals.

The other question a session raises is whose result it is. In the crowd version, the answer is a population's: the authors note that the individual user's preference "is not reflected in computation when crowdsourcing is the only source of data", and that crowds from different backgrounds can have clearly different preferences (Koyama and Igarashi, 2018). In the figures of this chapter, it is one person's, on one screen, on one day, with the noise and drift that came with them. Section 32.5 follows the line of work that borrows strength from other users' sessions to start a new user's session from a better prior.

25.6 Exercises #

Further reading #

• Koyama et al. (2017) introduce sequential line search and apply it to photo color enhancement with six parameters, answered by crowds.
• Koyama and Igarashi (2018) is a book chapter on computational design with crowds; it gives the crowd pipeline, the vote against Photoshop and Lightroom, and the three runs from different starting conditions.
• Koyama et al. (2020) replace the slider with a zoomable grid on a plane (Sequential Gallery) and report a small user study on photo enhancement with 12 parameters, including the participants' own words.
• Ou et al. (2023) compare novices and experts in three tasks, one of them photo color enhancement with an EUBO-based optimizer and a ranking interface.
• Iwai et al. (2025) run pairwise preferential optimization with professional designers on color variations of banner ads, 50 comparisons per task, with a click-through constraint handled by the machine.
• Mikkola et al. (2020) treat a slider-like answer as infinitely many comparisons and derive its likelihood.