IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
1
Principal Uncertainty Quantification with Spatial
Correlation for Image Restoration Problems
Omer Belhasin, Yaniv Romano, Daniel Freedman, Ehud Rivlin, Michael Elad
Abstract—Uncertainty quantification for inverse problems in
imaging has drawn much attention lately. Existing approaches
towards this task define uncertainty regions based on probable
values per pixel, while ignoring spatial correlations within the
image, resulting in an exaggerated volume of uncertainty. In this
paper, we propose PUQ (Principal Uncertainty Quantification)
– a novel definition and corresponding analysis of uncertainty
regions that takes into account spatial relationships within the
image, thus providing reduced volume regions. Using recent ad-
vancements in generative models, we derive uncertainty intervals
around principal components of the empirical posterior distribu-
tion, forming an ambiguity region that guarantees the inclusion
of true unseen values with a user-defined confidence probability.
To improve computational efficiency and interpretability, we also
guarantee the recovery of true unseen values using only a few
principal directions, resulting in more informative uncertainty
regions. Our approach is verified through experiments on image
colorization, super-resolution, and inpainting; its effectiveness is
shown through comparison to baseline methods, demonstrating
significantly tighter uncertainty regions.
Index Terms—Uncertainty and probabilistic reasoning, Prob-
ability and Statistics, Restoration, Inverse problems, Stochastic
processes, Correlation and regression analysis.
I. INTRODUCTION
R
ESTORATION tasks in imaging are widely encoun-
tered in various disciplines, including cellular cam-
eras, surveillance, experimental physics, and medical imaging.
These inverse problems are broadly defined as the need to
recover an unknown image given corrupted measurements
of it. Such problems, e.g., colorization, super-resolution, and
inpainting, are typically ill-posed, implying that multiple so-
lutions can explain the unknown target image. In this context,
uncertainty quantification aims to characterize the range of
possible solutions, their spread, and variability. This has an
especially important role in applications such as astronomy
and medical diagnosis, where it is necessary to establish
statistical boundaries for possible gray-value deviations. The
ability to characterize the range of permissible solutions with
accompanying statistical guarantees has thus become an im-
portant and useful challenge, addressed in this paper.
Prior work on this topic [1], [2] has addressed the uncer-
tainty assessment by constructing intervals of possible values
for each pixel via quantile regression [3], or other heuristics
such as estimations of per-pixel residuals. While this line
of thinking is appealing due to its simplicity, it disregards
O.
Belhasin
(omerbe@verily.com),
D.
Freedman
(danielfreed-
man@verily.com),
Ehud
Rivlin
(ehud@verily.com)
and
M.
Elad
(melad@verily.com) are with Verily Life Sciences, Israel.
O.
Belhasin
(omer.be@cs.technion.ac.il)
and
Y.
Romano
(yromano@technion.ac.il) are with the Department of Computer Science,
Technion - Israel Institute of Technology, Haifa, Israel.
Fig. 1.
Comparison of PUQ’s performance on the CelebA-HQ dataset in
image colorization, super-resolution, and inpainting tasks using the E-PUQ
procedure (Section IV-B1) applied on RGB image patches of varying size.
As seen, our method provides tighter uncertainty regions with significantly
smaller uncertainty volumes (×10 in super-res. and inpainting, and ×100 in
colorization). The compared methods are im2im-uq [1] and Conffusion [2].
spatial correlations within the image, and thus provides an
exaggerated uncertainty range. The study in [4] has improved
the above by quantifying the uncertainty in a latent space, thus
taking spatial dependencies into account. However, by relying
on a non-linear, non-invertible and uncertainty-oblivious trans-
formation, this method suffers from interpretability limitations
– See Section II for further discussion.
In this paper, we propose Principal Uncertainty Quan-
tification (PUQ) – a novel approach that accounts for spa-
tial relationships while operating in the image domain, thus
enabling a full and clear interpretation of the quantified
uncertainty region. PUQ uses the principal components of
the empirical posterior probability density function, which
describe the spread of possible solutions. PCA essentially ap-
proximates this posterior by a Gaussian distribution that tightly
encapsulates it. Thus, this approach reduces the uncertainty
volume1, as demonstrated in Figure 1. This figure presents
a comparison between our proposed Exact PUQ procedure
(see Section IV-B1) and previous work [1], [2], showing a
much desired trend of reduced uncertainty volume that further
decreases as the size of the patch under consideration grows.
Our work aims to improve the quantification of the uncer-
tainty volume by leveraging recent advancements in generative
models serving as stochastic solvers for inverse problems.
While our proposed approach is applicable using any such
solver (e.g., conditional GAN [5]), we focus in this work
on diffusion-based techniques, which have recently emerged
as the leading image synthesis approach, surpassing GANs
and other alternative generators [6]. Diffusion models offer a
systematic and well-motivated algorithmic path towards the
task of sampling from a prior probability density function
1The definition of this volume, which plays a critical part in this work, is
further discussed in later sections and given in Equation (3).
arXiv:2305.10124v3  [cs.CV]  20 Jan 2024

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
2
(PDF), Py, through the repeated application of a trained image-
denoiser [7], [8]. An important extension of these models
allows the sampler to become conditional, drawing samples
from the posterior PDF, Py|x, where x represents the observed
measurements. This approach has recently gained significant
attention [6], [9], [10], [11], yielding a fascinating viewpoint
to inverse problems, in which a variety of candidate high
perceptual quality solutions to such problems are obtained.
In this work, we generalize the pixelwise uncertainty as-
sessment, as developed in [1], [2], so as to incorporate spatial
correlations between pixels. This generalization is obtained
by considering an image-adaptive basis for a linear space
that replaces the standard basis in the pixelwise approach.
To optimize the volume of the output uncertainty region, we
propose a statistical analysis of the posterior obtained from a
diffusion-based sampler (e.g., [10], [11]), considering a series
of candidate restorations. Our method may be applied both
globally (on the entire image) or locally (on selected portions
or patches), yielding a tighter and more accurate encapsulation
of statistically valid uncertainty regions. For the purpose of
adapting the basis, we compute and leverage the principal
components of the candidate restorations. As illustrated in
Figure 2 for a simple 2-dimensional PDFs, the pixelwise
regions are less efficient and may contain vast empty areas, and
especially so in cases where pixels exhibit strong correlation.
Clearly, as the dimension increases, the gap between the
standard and the adapted uncertainty quantifications is further
amplified.
Our proposed method offers two conformal prediction [12],
[13], [14] based calibration options (specifically, using the
Learn then Test [15] scheme) for users to choose from, with
a trade-off between precision and complexity. These include
(i) using the entire set of principal components, (ii) using
a predetermined subset of them2. The proposed calibration
procedures ensure the validity of the uncertainty region to
contain the unknown true values with a user-specified con-
fidence probability, while also ensuring the recovery of the
unknown true values using the selected principal components
when only a subset is used. Applying these approaches allows
for efficient navigation within the uncertainty region of highly
probable solutions.
We conduct various local and global experiments to ver-
ify our method, considering three challenging tasks: image
colorization, super-resolution, and inpainting, all described
in Section V, and all demonstrating the advantages of the
proposed approach. For example, when applied locally on
8×8×3 patches, our experiments show a reduction in the guar-
anteed uncertainty volume by a factor of ∼10-100 compared to
previous approaches, as demonstrated in Figure 1. Moreover,
this local approach can have a substantially reduced compu-
tational complexity while retaining the statistical guarantees,
by drawing far fewer posterior samples and using a small
subset of the principal components. As another example, the
global tests on the colorization task provide an unprecedented
tightness in uncertainty volumes. This is accessible via a
2We also propose a reduced complexity variation of this option that controls
the number of necessary principal components to be used.
Fig. 2.
An illustration of uncertainty regions (in red) of 2d posterior
distributions and considering three different PDF behaviors, shown in blue,
orange, and green. The uncertainty regions are formed from intervals, as
defined in Equation (1), where ˆl(x) and ˆu(x) represent the 0.05 and 0.95
quantiles over the dashed black axes. The top row presents the uncertainty
region in the pixel domain using standard basis vectors that ignores the spatial
correlations, while the lower row presents the regions using the principal
components as the basis. The uncertainty volume, defined in Equation (3), is
indicated in the top left corner of each plot. The 90% coverage guarantee,
outlined in Equation (2) with wi := 1/2, is satisfied by all. As can be seen, the
lower row regions take spatial dependencies into account and are significantly
smaller than the pixelwise corresponding regions in the upper row.
reduced set of drawn samples, while also allowing for efficient
navigation within the solution set.
In summary, our contributions are the following:
1) We introduce a novel generalized definition of uncer-
tainty region that leverages an adapted linear-space basis
for better posterior coverage.
2) We propose a new method for quantifying the uncer-
tainty of inverse problems that considers spatial corre-
lation, thus providing tight uncertainty regions.
3) We present two novel calibration procedures for the un-
certainty quantification that provide statistical guarantees
for unknown data to be included in the uncertainty re-
gion with a desired coverage ratio while being recovered
with a small error by the selected linear axes.
4) We provide a comprehensive empirical study of three
challenging image-to-image translation tasks: coloriza-
tion, super-resolution, and inpainting, demonstrating the
effectiveness of the proposed approach in all modes.
II. RELATED WORK
Inverse problems in imaging have been extensively studied
over the years; this domain has been deeply influenced by
the AI revolution [16], [17], [18], [19], [5]. A promising
recent approach towards image-to-image translation problems
relies on the massive progress made on learned generative
techniques. These new tools enable to model the conditional
distribution of the output images given the input, offering a
fair sampling from this PDF. Generative-based solvers of this
sort create a new and exciting opportunity for getting high
perceptual quality solutions for the problem in hand, while
also accessing a diverse set of such candidate solutions.
Recently,
Denoising
Diffusion
Probabilistic
Models
(DDPM) [7], [8] have emerged as a new paradigm for image
generation, surpassing the state-of-the-art results achieved by

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
3
GANs [20], [6]. Consequently, several conditional diffusion
methods have been explored [6], [9], [10], [11], including SR3
[10] – a diffusion-based method for image super-resolution,
Palette [11] – a diffusion-based unified framework for
image-to-image translation tasks, and more (e.g. [21], [22],
[23], [24], [25], [26], [27]). Note that current conditional
algorithms for inverse problems do not offer statistical
guarantees against model deviations and hallucinations.
Moving to uncertainty quantification, the field of machine
learning has been seeing rich work on the derivation of
statistically rigorous confidence intervals for predictions [28],
[29], [30], [31], [32]. One key paradigm in this context is
conformal prediction (CP) [12], [13], [14] and risk-controlling
methods [33], [15], [34], which allow to rigorously quantify
the prediction uncertainty of a machine learning model with
a user-specified probability guarantee. Despite many proposed
methods, only a few have focused on mitigating uncertainty
assessment in image restoration problems, including im2im-
uq [1] and Conffusion [2]. The work reported in [35] is
closely related as it introduced a generalized, and thus im-
proved, calibration scheme for Conffusion [2]. All these works
have employed a risk-controlling paradigm [33] to provide
statistically valid prediction intervals over the pixel domain,
ensuring the inclusion of ground-truth solutions in the output
intervals. However, these approaches share the same limitation
of operating in the pixel domain while disregarding spatial
correlations within the image or the color layers. This leads
to an unnecessarily exaggerated volume of uncertainty.
An exception to the above is [4], which quantifies uncer-
tainty in the latent space of GANs. Their migration from
the image domain to the latent space is a rigid, global, non-
linear, non-invertible and uncertainty-oblivious transformation.
Therefore, quantification of the uncertainty in this domain
is quite limited. More specifically, rigidity implies that this
approach cannot adapt to the complexity of the problem by
adjusting the latent space dimension; Globality suggests that
it cannot be operated locally on patches in order to better
localize the uncertainty assessments; Being non-linear implies
that an evaluation of the uncertainty volume (see Section III)
in the image domain is hard and next to impossible; Non-
invertability of means that some energy is lost from the image
in the analysis and not accounted for, thus hampering the
validity of the statistical guarantees; Finally, note that the
latent space is associated with the image content, but does
not represent the prime axes of the uncertainty behavior. Note
that due to the above, and especially the inability to provide
certified volumes of uncertainty, an experimental comparison
of our method to [4] is impossible.
Inspired by the above contributions, we propose a novel
alternative uncertainty quantification approach that takes spa-
tial relationships into account. Our work provides tight un-
certainty regions, compared to prior work, with user-defined
statistical guarantees through the use of a CP-based paradigm.
Specifically, we adopted the Learn then Test [15] that provides
statistical guarantees for controlling multiple risks.
III. PROBLEM FORMULATION
Let Px,y be a probability distribution over X × Y, where X
and Y represent the input and the output space, respectively,
for the inverse problem at hand. E.g., for the task of image
colorization, Y could represent full-color high-quality images,
while X represents their colorless versions to operate on. We
assume that X, Y ⊂[0, 1]d ⊂Rd, where, without loss of
generality, d is assumed to be the dimension of both spaces.
In the context of examining patches within output images,
we define Ypatch as the patch space of the output images. For
simplicity, we use the same notation, d, for Y and Ypatch, while
it is clear that the dimension of Ypatch is smaller and controlled
by the user through the patch size to work on.
Given an input measurement x ∈Rd, we aim to quantify the
uncertainty of the possible solutions to the inverse problem, as
manifested by the estimated d-dimensional posterior distribu-
tion, ˆPy|x. The idea is to enhance the definition of pixelwise
uncertainty intervals by integrating the spatial correlations
between pixels to yield a better structured uncertainty region.
To achieve this, we propose to construct uncertainty intervals
using a designated collection of orthonormal basis vectors
for Rd instead of intervals over individual pixels. We denote
this collection by ˆB(x) = {ˆv1(x), ˆv2(x) . . . ˆvd(x)}, where
ˆvi(x) ∈Rd. These vectors are instance-dependent, thus best
adapted to their task. An intuitive example of such a basis is
the standard one, ˆB(x) = {e1, e2 . . . ed}, where ei ∈Rd is
the one-hot vector with value 1 in the ith entry. In our work,
we use a set of principal components of ˆPy|x, which will be
discussed in detail in Section IV.
Similar to [1], [2], we use an interval-based method cen-
tered around the conditional mean image, i.e., an estimate of
E[y|x] ∈Rd, denoted by ˆµ(x). Formally, we utilize the follow-
ing interval-valued function that constructs prediction intervals
along each basis vector around the estimated conditional mean:
T (x; ˆB(x))i :=
h
ˆvi(x)T ˆµ(x) −ˆl(x)i, ˆvi(x)T ˆµ(x) + ˆu(x)i
i
. (1)
In the above, i ∈{1, 2 . . . d} is a basis vector index, and
ˆl(x)i ∈R+ and ˆu(x)i ∈R+ are the lower and upper interval
boundaries for the projected values of candidate solutions
emerging from ˆPy|x. That is, if ˆy ∼ˆPy|x is such a solution,
ˆvi(x)T ˆy is its i-th projection, and this value should fall within
T (x; ˆB(x))i with high probability. Returning to the example
of the standard basis, the above equation is nothing but pixel-
wise prediction intervals, which is precisely the approach taken
in [1], [2]. By leveraging this generalization, the uncertainty
intervals using these basis vectors form a d-dimensional hyper-
rectangle, referred to as the uncertainty region.
Importantly, we propose that the interval-valued function,
T , should produce valid intervals that contain a user-specified
fraction of the projected ground-truth values within a risk level
of α ∈(0, 1). In other words, more than 1−α of the projected
ground-truth values should be contained within the intervals,
similar to the approach taken in previous work in the pixel
domain. To achieve this, we propose a holistic expression that
aggregates the effect of all the intervals, T (x; ˆB(x)). This
expression leads to the following condition:
E
" d
X
i=1
ˆwi(x) · 1
n
ˆvi(x)T y ∈T (x; B(x))i
o#
> 1 −α,
(2)

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
4
where y ∈Rd is the unknown ground-truth and ˆwi(x) ∈[0, 1]
s.t. Pd
i=1 ˆwi(x) = 1 are the weight factors that set the
importance of covering the projected ground-truth values along
each interval. In Section IV we discuss the proposed holistic
expression and a specific choice of these weights. As an
example, we could set α = 0.1 and ˆwi(x) := 1/d, indicating
that more than 90% of the projected ground-truth values onto
the basis vectors are contained in the intervals, as illustrated
in a 2d example in Figure 2 for different kinds of ˆPy|x.
As discussed above and demonstrated in Figure 2, if the
orthonormal basis in Equation (1) is chosen to be the standard
one, we get the pixel-based intervals that disregard spatial
correlations within the image, thus leading to an exaggerated
uncertainty region. In this work, we address this limitation by
transitioning to an instance-adapted orthonormal basis of Rd
that allows the description of uncertainty using axes that are
not necessarily pixel-independent, thereby providing tighter
uncertainty regions. While such a basis could have been
defined analytically using, for example, orthonormal wavelets
[36], we suggest a learned and thus a better-tuned one. The
choice to use a linear and orthonormal representation for the
uncertainty quantification comes as a natural extension of
the pixelwise approach, retaining much of the simplicity and
efficiency of treating each axis separately. Note that the or-
thogonality enables the decomposition of y around ˆµ(x) via its
projected values, y = ˆµ(x)+Pd
i=1

ˆvi(x)T (y −ˆµ(x))

ˆvi(x),
which we refer to as the exact reconstruction property.
To evaluate the uncertainty across different uncertainty
regions, we introduce a new metric called the uncertainty
volume, V(x; T (x; ˆB(x))), which represents the dth root of
the uncertainty volume with respect to intervals T (x; ˆB(x)),
defined in the following equation:
V(x; T (x; ˆB(x))) :=
d
v
u
u
t
d
Y
i=1
h
ˆu(x)i + ˆl(x)i
i
(3)
≈exp
 
1
d
d
X
i=1
log

ˆu(x)i + ˆl(x)i + ϵ
!
−ϵ ,
where ϵ > 0 is a small hyperparameter used for numerical
stability. In Section V we demonstrate that our approach results
in a significantly reduction in these uncertainty volumes when
compared to previous methods.
When operating in high dimensions (e.g. on the full image),
providing uncertainty intervals for all the d-dimensions poses
severe challenges, both in complexity and interpretability.
In this case, constructing and maintaining the basis vectors
becomes infeasible. Moreover, the uncertainty quantification
using these intervals may be less intuitive compared to the con-
ventional pixelwise approach because of the pixel-dependency
between the basis vectors, which makes it difficult to commu-
nicate the uncertainty to the user. To mitigate these challenges,
we propose an option of using K ≪d basis vectors that
capture the essence of the uncertainty. In Section IV, we
discuss how to dynamically adjust K to provide fewer axes.
While reducing the number of basis vectors benefits in
interpretability and complexity, this option does not fulfill
the exact reconstruction property. Therefore, we propose an
extension to the conventional coverage validity of Equation (2)
y
x
ˆyi ∼ˆPy|x
Fig. 3. The sampling procedure for two image restoration problems using a
conditional stochastic generator. The top row corresponds to super-resolution
in local mode with patches, while the bottom row shows colorization in global
mode. The implementation details are described in Section IV-A.
Pixels Axes
PCs Axes
Approximation
Calibration
Fig. 4. Illustration of our PUQ procedure in 2D (d = 2) for a single instance
x ∈X. The top row corresponds to the case when K = d = 2 (as in
E-PUQ), while the bottom row depicts the case when K = 1 < d = 2
(as in DA-PUQ and RDA-PUQ). The procedure begins by drawing samples
ˆyi ∼ˆPy|x. Next, these samples are projected onto the PCs domain: ˆV T ˆyi,
where ˆV
:= [ˆv1, . . . , ˆvK] ∈Rd×K. Then, we compute bounds along
the PCs to contain the samples at the correct ratio, forming the intervals
specified in Equation (1). Finally, the intervals are scaled to statistically
guarantee Equation (2) and contain the correct ratio over solutions for unseen
input instances. In the bottom row, the procedure also statistically guarantees
Equation (4) by ensuring a small recovery error of solutions to unseen input
instances, as demonstrated by the small variance around the single PC.
that takes into account the reconstruction error of the decom-
posed ground-truth images. Specifically, the user sets a ratio
of pixels, q ∈R, and a maximum acceptable reconstruction
error over this ratio, β ∈(0, 1). This approximation allows
us to reduce the number of basis vectors used to formulate
ˆB(x), such that the reconstruction will be valid according to
the following condition:
E
"
ˆQq
 



K
X
j=1
ˆvj(x)T ycˆvj(x) −yc

i



d
i=1
!#
≤β ,
(4)
where yc := y −ˆµ(x) is the ground-truth image centered
around ˆµ(x), and ˆQq(·) is the empirical quantile function de-
fined by the smallest z satisfying 1
d
Pd
i=1 1{zi ≤ˆQq(z)} ≥q.
In Section IV, we discuss this expression for assessing the
validity of the basis vectors. As an example, setting q = 0.9
and β = 0.05 would mean that the maximal reconstruction
error of 90% of the ground-truth pixels is no more than 5%
of the [0, 1] dynamic range.

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
5
IV. PUQ: PRINCIPAL UNCERTAINTY QUANTIFICATION
In this section, we present Principal Uncertainty Quantifica-
tion (PUQ), our method for quantifying the uncertainty in in-
verse problems while taking into account spatial dependencies
between pixels. PUQ uses the principal components (PCs) of
the solutions to the inverse problem for achieving its goal. In
Appendix A, we provide an intuition behind the choice of the
PCs as the basis. Our approach can be used either globally
across the entire image, referred to as the global mode; or
locally within predefined patches or segments of interest,
referred to as the local mode. Local uncertainty quantification
can be applied to any task, where the dimensionality of
the target space is fully controlled by the user. In contrast,
global quantification is particularly advantageous for tasks that
exhibit strong spatial correlations between pixels.
Our proposed method consists of two phases. In the first,
referred to as the approximation phase, a machine learning
system is trained to predict the PCs of possible solutions, de-
noted by ˆB(x) = {ˆv1(x), ˆv2(x), . . . , ˆvK(x)} (where K ≤d),
as well as a set of importance weights, ˆw(x) ∈RK, referring
to the vectors in ˆB(x). In addition, the system estimates the
necessary terms in Equation (1), which include the conditional
mean, ˆµ(x) ∈Rd, and the lower and upper bounds, ˜l(x) ∈RK
and ˜u(x) ∈RK 3, for the spread of projected solutions over
ˆB(x). All these ingredients are obtained by a diffusion-based
conditional sampler as described in Figure 3. More details on
this computational process are brought in Section IV-A.
The above-described approximation phase is merely an
estimation, as the corresponding heuristic intervals of Equa-
tion (1) may not contain the projected ground-truth values
with a desired ratio. Additionally, the basis vectors may not
be able to recover the ground-truth pixel values within an
acceptable threshold when K < d, or the basis set may
contain insignificant axes in terms of variability. Therefore,
in the second, calibration phase, we offer two calibration
procedures on an held-out set of calibration data, denoted by
Scal := {(xi, yi)}n
i=1. These assess the validity of our proposed
uncertainty region over unseen data, which is composed by the
intervals defined in Equation (1). The choice between the two
calibration procedures depends on the user, taking into account
the trade-off between precision and complexity. The steps of
our proposed method are summarized in Algorithm 1, and the
two calibration strategies are as follows:
(1) Exact PUQ (E-PUQ - Section IV-B1): In the setting of an
exact uncertainty assessment, while assuming that d PCs can
be constructed and maintained in full, the exact reconstruction
property is satisfied. Consequently, the calibration procedure
is straightforward, involving only scaling of the intervals
until they contain the user-specified miscoverage preference,
denoted by α ∈(0, 1), of the projected ground-truth values
falling outside the uncertainty region. This is similar to the
approach taken in previous work over the pixel domain.
(2) Dimension-Adaptive PUQ (DA-PUQ - Section IV-B2,
RDA-PUQ - Appendix D): In the setting of an approximate
3Note that these bounds are meant for ˆPy|x and not for Py|x, and thus
marked with tilde. The ˆl(x), ˆu(x) bounds that are related to Equation (1) are
defined later in the calibration schemes (Section IV-B).
uncertainty assessment, while allowing for a small recovery
error of projected ground-truth instances to full-dimensional
instances, either due to complexity or interpretability reasons
(see Section III), the exact reconstruction property is no longer
satisfied. Hence, in addition to the scaling procedure outlined
above, we must verify that the K PCs can decompose the
ground-truth pixel values with a small error. In this calibration
process, we also control the minimum number of the first ˆk(x)
PCs out of the K PCs, such that a small reconstruction error
can be guaranteed for unseen data. This number is dynamically
determined per input image, so that instances with greater
pixel correlations are assigned more PCs than those with
weaker correlations. As manually determining K might be
challenging, we introduce the Reduced Dimension-Adaptive
PUQ (RDA-PUQ) procedure that also controls that value as
part of the calibration - see Appendix D.
Algorithm 1 Generating PUQ Axes and Intervals
Input: Training set. Calibration set. Number of PCs K ∈N. An
unseen input instance x ∈Rd.
Output: Statistically valid uncertainty axes and intervals for x.
▷Approximation phase
1: Train a machine learning system (e.g., Section IV-A) to estimate
the following:
K PCs of ˆPy|x
Importance weights of PCs
The conditional mean
Lower and upper bounds on the PCs
▷Calibration phase
2: if Exact uncertainty (accurate) then
3:
Apply E-PUQ using the calibration data
4: else if Approximate uncertainty (reduced complexity) then
5:
Apply DA-PUQ / RDA-PUQ using the calibration data
6: end if
▷Inference
7: Provide statistically valid uncertainty axes and intervals in terms
of Equation (2) and Equation (4), applied to an unseen input
instance x
In Section V we demonstrate a significant decrease in
the uncertainty volume, as defined in Equation (3) for each
procedure, whether applied globally or locally, compared to
prior work. On the one hand, the E-PUQ procedure is the
simplest and can be applied locally to any task, and globally
to certain tasks where the computation of d PCs is feasible.
On the other hand, the DA-PUQ and RDA-PUQ procedures
are more involved and can be applied both globally or locally
to any task, while these are particularly effective in cases
in which pixels exhibit strong correlations, such as in the
image colorization task. Our method is visually illustrated in
Figure 4, showing a sampling methodology and a calibration
scheme using the full PCs or only a subset of them.
A. Diffusion Models for the Approximation Phase
The approximation phase, summarized in Algorithm 1 in RED,
can be achieved in various ways. In this section, we describe
the implementation we used to obtain the results in Section V.
While we aim to construct the uncertainty axes and intervals

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
6
in the most straightforward way, further exploration of more
advanced methods to achieve the PCs is left for future work.
In our implementation, we leverage the recent advances in
stochastic regression solvers for inverse problems based on
diffusion models, which enable to train a machine learning
model to generate high-quality samples from ˆPy|x. Formally,
we define fθ : X × Z →Y as a stochastic regression
solver for an inverse problem in global mode, where Z is
the noise seed space. Similarly, in local mode, we consider
fθ : X × Z →Ypatch. Given an input instance x ∈Rd, we
propose to generate K samples, denoted by {fθ(x, zi)}K
i=1,
where, fθ(x, zi) ∼ˆPy|x. These samples are used to estimate
the PCs of possible solutions and their importance weights
using the SVD decomposition of the generated samples. The
importance weights assign high values to axes with large
variance among projected samples, and low ones to those with
small variance. In Section IV-B, we elaborate on how these
weights are used in the calibration phase. Additionally, the
samples are utilized to estimate the conditional mean, ˆµ(x),
and the lower and upper bounds, ˜l(x) and ˜u(x), necessary
for Equation (1). ˜l(x) and ˜u(x) are obtained by calculating
quantiles of the projected samples onto each PC, with a user-
specified miss-coverage ratio α ∈(0, 1).
To capture the full spread and variability of ˆPy|x, it is
necessary to generate at least K = d samples to feed to
the SVD procedure, which is computationally challenging for
high-dimensional data. As a way out, we suggest working
locally on patches, where d is small and fully controlled by
the user by specifying the patch size to work on. However, for
tasks with strong pixel correlation, such as image colorization,
a few PCs can describe the variability of ˆPy|x with a very small
error. Therefore, only a few samples (i.e., K ≪d) are required
for the SVD procedure to construct meaningful PCs for the
entire image, while capturing most of the richness in ˆPy|x.
We formally summarize our sampling-based methodology, in
either global or local modes, in Algorithm 2.
B. Calibration Phase
In order to refine the approximation phase and obtain valid
uncertainty axes and intervals that satisfy the guarantees of
Equation (2) and Equation (4), it is necessary to apply a
calibration phase, as summarized in Algorithm 1 in BLUE.
This phase includes two different options based on particular
conditions on the number of PCs to be constructed and main-
tained during the calibration procedure or during inference,
when applied either globally or locally. Below we outline each
of these options in more details.
1) Exact PUQ: The Exact PUQ (E-PUQ) procedure pro-
vides the complete uncertainty of the d-dimensional posterior
distribution, Py|x. In this case, the exact reconstruction prop-
erty discussed in Section III is satisfied, and Equation (4) is
fulfilled with 0% error (β = 0) across 100% (q = 1.0) of
the pixels. Therefore, the calibration is simple, involving only
a scaling of intervals to ensure Equation (2) is satisfied with
high probability, similar to previous work [1], [2].
Formally, for each input instance x and its corresponding
ground-truth value y ∈Rd in the calibration data, we use the
estimators obtained in the approximation phase to get d PCs
Algorithm 2 Approximation Phase via Sampling
Input: Instance x ∈X. Conditional stochastic generative model fθ :
X →Y or fθ : X →Ypatch. Maximal PCs / samples number
K ≤d. Misscoverage ratio α ∈(0, 1).
▷Generate samples drawn from ˆPy|x
1: for i = 1 to K do
2:
ˆyi(x) ←fθ(x, zi)
3: end for
▷Compute conditional mean
4: ˆµ(x) ←
1
K
PK
i=1 ˆyi(x)
▷Apply SVD decomposition and extract the PCs and weights
5: ˆY (x) ←[ˆy1(x), ˆy2(x) . . . ˆyK(x)] ∈Rd×K
6: ˆY (x) −ˆµ(x) · 1T
K = ˆV (x)ˆΣ(x) ˆU(x)T
7: ˆB(x) ←{ˆv1(x), ˆv2(x) . . . ˆvK(x)}, where ˆvi(x) = [ ˆV (x)]i
8: ˆw(x) ←

ˆσ1(x)2, . . . , ˆσK(x)2
/c ∈RK, where ˆσi(x) =
[ˆΣ(x)]i and c = PK
j=1 ˆσj(x)2.
▷Compute α/2 and 1 −α/2 empirical quantiles of projected
samples onto each PC
9: for i = 1 to K do
10:
˜l(x)i ←ˆQα/2({ˆvi(x)T (ˆyj(x) −ˆµ(x))}K
j=1)
11:
˜u(x)i ←ˆQ1−α/2({ˆvi(x)T (ˆyj(x) −ˆµ(x))}K
j=1)
12: end for
Output: K PCs ˆB(x), importance weights ˆw(x), conditional mean
ˆµ(x), lower and upper bounds ˜l(x) and ˜u(x).
of possible solutions ˆB(x), their corresponding importance
weights ˆw(x), the conditional mean ˆµ(x), and the lower and
upper bounds, denoted by ˜l(x) and ˜u(x). We then define
the scaled intervals to be those specified in Equation (1),
with the upper and lower bounds defined as ˆu(x) := λ˜u(x)
and ˆl(x) := λ˜l(x), where λ ∈R+ is a tunable parameter
that controls the scaling. Notably, the size of the uncertainty
intervals decreases as λ decreases. We denote the scaled
uncertainty intervals by Tλ(x; ˆB(x)). The following weighted
coverage loss function is used to guide our design of λ:
L(x, y; λ) :=
d
X
i=1
ˆwi(x) · 1
n
ˆvi(x)T y ̸∈Tλ(x; ˆB(x))i
o
.
(5)
This loss is closely related to the expression in Equation (2),
and while it may seem arbitrary at first, this choice is a direct
extension to the one practiced in [1], [2]. In Appendix B we
provide an additional justification for it, more tuned to the
realm discussed in this paper.
Our goal is to ensure that the expectation of L(x, y; λ)
is below a pre-specified threshold, α, with high probability
over the calibration data. This is accomplished by a conformal
prediction based calibration scheme, and in our paper we use
the LTT [15] procedure, which guarantees the following:
P

E[L(x, y; ˆλ)] ≤α

≥1 −δ ,
(6)
for a set of candidate values of λ, given as the set ˆΛ. δ ∈(0, 1)
is an error level on the calibration set and ˆλ is the smallest
value within ˆΛ satisfying the above condition, so as to provide
the smallest uncertainty volume over the scaled intervals, as
defined in Equation (3), which we denote by Vˆλ.
Put simply, the above guarantees that more than 1 −α of
the ground-truth values projected onto the full d PCs of ˆPy|x
are contained in the uncertainty intervals with probability at
least 1−δ, where the latter probability is over the randomness
of the calibration set. The scaling factor takes into account the
weights to ensure that uncertainty intervals with high variabil-
ity contain a higher proportion of projected ground-truth values

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
7
than those with low variability. This is particularly important
for tasks with strong pixel correlations, where the first few
PCs capture most of the variability in possible solutions. We
describe in detail the E-PUQ procedure in Algorithm 3.
Algorithm 3 Exact PUQ Procedure
Input: Calibration set Scal
:= {xi, yi}n
i=1. Scanned calibration
parameter values Λ = [1 . . . λmax]. Approximation phase esti-
mations ˆB, ˆw, ˆµ, ˜u, ˜l. Misscoverage ratio α ∈(0, 1). Calibration
error level δ ∈(0, 1).
1: for (x, y) ∈Scal do
2:
ˆB(x), ˆw(x), ˆµ(x), ˜u(x), ˜l(x) ←Apply Algorithm 2 to
x, with the choice of K = d samples
3:
for λ ∈Λ do
▷Scale uncertainty intervals
4:
ˆu(x) ←λ˜u(x) and ˆl(x) ←λ˜l(x)
5:
Tλ(x; ˆB(x)) ←Equation (1) using ˆµ(x), ˆu(x), ˆl(x)
▷Compute weighted coverage loss, Equation (5)
6:
L(x, y; λ) ←
7:
Pd
i=1 ˆwi(x) · 1
n
ˆvi(x)T y ̸∈Tλ(x; ˆB(x))i
o
8:
end for
9: end for
10: ˆΛ ←Extract valid λs from LTT [15] applied on {L(x, y; λ) :
(x, y) ∈Scal, λ ∈Λ} at risk level α and error level δ, referring
to Equation (6).
▷Compute the minimizer for the uncert. volume, Equation (3)
11: ˆλ ←arg minλ∈ˆΛ
n
1
n
Pn
i=1 Vλ(xi; ˆB(x))
o
Output: Given a new instance x ∈X, obtain valid uncertainty
intervals for it, Tˆλ(x; ˆB(x)).
2) Dimension-Adaptive PUQ: The E-PUQ procedure as-
sumes the ability to construct and maintain d PCs, which
can be computationally challenging both locally and globally.
Furthermore, an uncertainty quantification over these axes may
be less intuitive, due to the many axes involved, thus harming
the method’s interpretability (see discussion in Section III).
To address these, we propose the Dimension-Adaptive PUQ
(DA-PUQ) procedure, which describes the uncertainty region
with fewer axes, K ≤d. The use of only a few leading
dimensions, e.g., K = 3, can lead to a more interpretable
uncertainty region, enabling an effective visual navigation
within the obtained uncertainty range.
While this approach does not satisfy the exact reconstruction
property (see Section III), the decomposed ground-truth values
can still be recovered through the K PCs with a small user-
defined error in addition to the coverage guarantee. By doing
so, we can achieve both the guarantees outlined in Equation (2)
and Equation (4) with high probability.
To satisfy both the coverage and reconstruction guarantees
while enhancing interpretability, we use a dynamic function,
ˆk(x) : X →N, and a scaling factor to control the recon-
struction and coverage risks. The function ˆk(x) determines the
number of top PCs (out of K) to include in the uncertainty
region, focusing on the smallest number that can satisfy both
Equation (2) and (4), so as to increase interpretability.
Formally, for each input instance x and its corresponding
ground-truth value y ∈Rd in the calibration data, we use
the estimators obtained in the approximation phase to esti-
mate K ≤d PCs of possible solutions, denoted by ˆB(x),
their corresponding importance weights, denoted by ˆw(x),
the conditional mean denoted by ˆµ(x), and the lower and
upper bounds denoted by ˜l(x) and ˜u(x), respectively. We
then introduce a threshold λ1 ∈(0, 1) for the decay of
the importance weights over the PCs of solutions to x. The
adaptive number of PCs to be used is defined as follows:
ˆk(x; λ1) :=
min
1≤k≤K
(
k
s.t.
k
X
i=1
ˆwi(x) ≥λ1
)
.
(7)
Obviously, the importance weights are arranged in a descend-
ing order, starting from the most significant axis and ending
with the least significant one. Furthermore, let q ∈(0, 1) be
a specified ratio of pixels, and β ∈(0, 1) be a maximum al-
lowable reconstruction error over this ratio. The reconstruction
loss function to be controlled is defined as:
L1(x, y; λ1) := ˆQq







ˆk(x;λ1)
X
j=1
ˆvj(x)T ycˆvj(x) −yc

i



d
i=1


,
(8)
where ˆQq(·) selects the empirical q-quantile of the recon-
struction errors, and yc = y −ˆµ(x) is the ground-truth image
centered around ˆµ(x). In Appendix C, we discuss further this
specific loss function for controlling the capability of the linear
subspace to capture the richness of the complete d-dimensional
posterior distribution.
At the same time, we also control the coverage risk over
the ˆk(x) PCs, with α ∈(0, 1) representing a user-specified
acceptable misscoverage rate and λ2 ∈R+ representing the
calibration factor parameter. To control this coverage risk,
we define the coverage loss function to be the same as in
Equation (5), but limited to the ˆk(x) PCs, that is:
L2(x, y; λ1, λ2) :=
(9)
ˆk(x;λ1)
X
i=1
ˆwi(x) · 1
n
ˆvi(x)T y ̸∈Tλ2(x; ˆB(x))i
o
.
Finally, using the reconstruction loss function of Equa-
tion (8) and the coverage loss function of Equation (9), we seek
to minimize the uncertainty volume, defined in Equation (3),
for the scaled intervals where any unused axes (out of d) are
fixed to zero. We denote this uncertainty volume as Vλ1,λ2.
The minimization of Vλ1,λ2 is achieved by minimizing λ1 and
λ2, while ensuring that the guarantees of Equation (2) and
Equation (4) hold with high probability over the calibration
data. This can be provided, for example, through the LTT [15]
calibration scheme, which guarantees the following:
P

E[L1(x, y; ˆλ1)] ≤β
E[L2(x, y; ˆλ1, ˆλ2)] ≤α

≥1 −δ ,
(10)
where ˆλ1 and ˆλ2 are the minimizers for the uncertainty
volume among valid calibration parameter results, ˆΛ, obtained
through the LTT procedure. In other words, we can reconstruct
a fraction q of the ground-truth pixel values with an error
no greater than β, and a fraction of more than 1 −α of
the projected ground-truth values onto the first ˆk(x; ˆλ1) PCs
of Py|x are contained in the uncertainty intervals, with a
probability of at least 1 −δ. A detailed description of the
DA-PUQ procedure is given in Algorithm 4.
The above-described DA-PUQ procedure reduces the num-
ber of PCs to be constructed to K
≤
d while using
ˆk(x; ˆλ1) ≤K PCs, leading to increased efficiency in both time
and space during inference. However, determining manually
the smallest K value that can guarantee both Equation (2)
and Equation (4) can be challenging. To address this, we

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
8
propose an expansion of the DA-PUQ procedure; the Reduced
Dimension-Adaptive PUQ (RDA-PUQ) procedure that also
controls the maximum number of PCs, K, required for the
uncertainty assessment. This approach is advantageous for
inference as it reduces the number of samples required to
construct the PCs using Algorithm 2, while ensuring both the
coverage and reconstruction guarantees of Equation (2) and
Equation (4) with high probability. The RDA-PUQ procedure
is fully described in Appendix D.
Algorithm 4 Dimension-Adaptive PUQ Procedure
Input: Calibration set Scal
:= {xi, yi}n
i=1. Scanned calibration
parameter values Λ1 ←[1 . . . λ1max] and Λ2 ←[1 . . . λ2max].
Maximal PCs number K ≤d. Approximation phase estimators
ˆB, ˆw, ˆµ, ˜u, ˜l. Recovered pixels ratio q ∈(0, 1). Reconstruction
error β ∈(0, 1). Misscoverage ratio α ∈(0, 1). Calibration error
level δ ∈(0, 1). For an effective calibration, α, β, δ should be
close to 0 while q should be close to 1.
1: for (x, y) ∈Scal do
2:
ˆB(x), ˆw(x), ˆµ(x), ˜u(x), ˜l(x) ←Apply Algorithm 2 to
x, with the choice of K samples
3:
for λ1 ∈Λ1 do
▷Compute adaptive dimensionality, Equation (7)
4:
ˆk(x; λ1) ←mink
n
k : PK
i=1 ˆwi(x) ≥λ1
o
▷Compute reconstruction loss, Equation (8)
5:
yc ←y −ˆµ(x)
6:
L1(x, y; λ1) ←
ˆQq

Pˆk(x;λ1)
j=1
ˆvj(x)T ycˆvj(x) −yc

i
d
i=1

7:
for λ2 ∈Λ2 do
▷Scale uncertainty intervals
8:
ˆu(xi) ←λ2˜u(x) and ˆl(x) ←λ2˜l(x)
9:
Tλ2(x; ˆB(x)) ←Eq. (1) using ˆµ(x), ˆu(x), ˆl(x)
▷Compute weighted coverage loss, Equation (5)
10:
L2(x, y; λ1, λ2) ←Pˆk(x;λ1)
i=1
ˆwi(x)·
1
n
ˆvi(x)T y ̸∈Tλ2(x; ˆB(x))i
o
11:
end for
12:
end for
13: end for
14: ˆΛ
←
Extract
valid
λs
from
LTT
[15]
ap-
plied
on
{(L1(x, y; λ1), L2(x, y; λ1, λ2))
:
(x, y) ∈Scal, λ1 ∈Λ1, λ2 ∈Λ2}
at
risk
levels
(β, α)
and
error level δ, referring to Equation (10)
▷Compute the minimizers for the uncer. volume, Equation (3)
15: ˆλ1, ˆλ2 ←arg minλ1,λ2∈ˆΛ
n
1
n
Pn
i=1 Vλ1,λ2(xi; ˆB(xi))
o
Output: Given a new instance x ∈X, obtain valid uncertainty
intervals for it, Tˆλ2(x; ˆB(x)) over ˆk(x; ˆλ1) ≤K PCs.
V. EMPIRICAL STUDY
This section presents a comprehensive empirical study of our
proposed method PUQ, applied to three challenging tasks:
image colorization, super-resolution, and inpainting, over the
CelebA-HQ dataset [37]. Our approximation phase starts with
a sampling from the posterior, applied in our work by the
SR3 conditional diffusion model [10]. Figure 5 presents typical
sampling results for these three tasks, showing the expected
diversity in the images obtained.
The experiments we present herein verify that our method
satisfies both the reconstruction and coverage guarantees and
demonstrate that PUQ provides more confined uncertainty
regions compared to prior work, including im2im-uq [1] and
Conffusion [2]. Through the experiments, we present superi-
ority in uncertainty volume, as defined in Equation (3), and in
interpretability through the use of only a few PCs to assess the
uncertainty of either a patch or a complete image. All the ex-
periments were conducted over 100 calibration-test splits. For
in-depth additional details of our experiments and the settings
used, we refer the reader to Appendix E. Additionally, an ab-
lation study has been conducted, as elaborated in Appendix H.
This study presents an analysis of user-defined parameters: α,
β, q, and δ, aiming to provide a comprehensive insight into
their selection. Furthermore, we have investigated the trade-
off between precision and complexity in Appendix I to offer
a complete understanding of our method’s performance.
A. Evaluation Metrics
Before presenting the results, we discuss the metrics used to
evaluate the performance of the different methods. Although
our approach is proved to guarantee Equation (6) for E-PUQ
and Equation (10) for DA-PUQ, (through LTT [15]), we assess
the validity and tightness of these guarantees as well.
Empirical coverage risk.
We measure the risk associated
with the inclusion of projected unseen ground-truth values in
the uncertainty intervals. In E-PUQ, we report the average
coverage loss, defined in Equation (5). In the case of DA-PUQ
and RDA-PUQ, we report the value defined by Equation (9).
Empirical reconstruction risk.
We measure the risk in
recovering unseen ground-truth pixel values using the selected
PCs. In the case of E-PUQ, this risk is zero by definition.
However, for DA-PUQ and RDA-PUQ, we report the average
reconstruction loss, defined by Equation (8).
Interval-Size.
We report the calibrated uncertainty inter-
vals’ sizes of Equation (1), and compare them with baseline
methods. For E-PUQ, we compare intervals over the full basis
set of PCs with the intervals in the pixel domain used in
previous work. In the DA-PUQ and RDA-PUQ procedures,
we apply dimensionality reduction to K ≪d dimensions. To
validly compare the intervals’ sizes of these methods to those
methods over the full d dimensions, we pad the remaining
d −K dimensions with zeros as we assume that the error
in reconstructing the ground-truth from the dimensionally
reduced samples is negligible.
Uncertainty Volume. We report these volumes, defined
in Equation (3), for the calibrated uncertainty regions and
compare them with previous work. A smaller volume implies
a higher level of certainty in probable solutions to Py|x. In
E-PUQ, we compare volumes over the full basis set of PCs,
whereas for the DA-PUQ and RDA-PUQ procedures, we pad
the remaining dimensions with zeros.
B. Local Experiments on Patches
We apply our proposed methods on RGB patches of increasing
size — 1x1, 2x2, 4x4, and 8x8 — for image colorization,
super-resolution, and inpainting tasks. The obtained results are
illustrated in Figure 6 and Figure 7, where Figure 6 compares
our exact procedure, E-PUQ, to baseline methods, and Figure 7
examines our approximation procedures, DA-PUQ and RDA-
PUQ. In Table I we present a numerical comparison of

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
9
x
y
Samples
Fig. 5. The three image recovery tasks, colorization (top), super-resolution (middle) and inpainting (bottom). For each we present a given measurement x,
the ground-truth y, and 10 candidate samples from the (approximated) posterior distribution. These samples fuel the approximation phase in our work.
TABLE I
LOCAL EXPERIMENTS: QUANTITATIVE COMPARISON OF THE MEANS
AND STANDARD DEVIATIONS OF OUR LOCALLY APPLIED PUQ METHOD
ON RGB PATCH RESOLUTION OF 8X8, UTILIZING THE TWO PROPOSED
PROCEDURES. NOTE THAT IN THIS EXPERIMENT d = 8 × 8 × 3 = 192
AND ϵ = 1e −10 FOR THE VOLUME COMPUTATION.
Recons. Risk
Dim. ˆk(x) / K
Uncert. Volume
Colorization
im2im-uq [1]
0
192 / 192
1.6e−1 ± 7.2e−2
Conffusion [2]
0
192 / 192
1.7e−1 ± 1.3e−1
E-PUQ
0
192 / 192
2.3e−3 ± 9.8e−4
DA-PUQ
2.5e−2 ± 5.3e−4
1.6 ± 0.77 / 100
2.4e−11 ± 1.3e−11
RDA-PUQ
1.7e−2 ± 9.7e−4
3.8 ± 1.8 / 11.8 ± 6.4
6.8e−11 ± 3.9e−11
Super-
Resolution
im2im-uq [1]
0
192 / 192
8.8e−2 ± 5.7e−2
Conffusion [2]
0
192 / 192
8.8e−2 ± 6.9e−2
E-PUQ
0
192 / 192
1.3e−2 ± 7.5e−3
DA-PUQ
2.5e−2 ± 3.4e−4
11.1 ± 5.5 / 192
3.4e−10 ± 3.5e−10
RDA-PUQ
2.0e−2 ± 0.0e−0
22.8 ± 6.8 / 70.0 ± 0.0
1.6e−9 ± 1.3e−9
Inpainting
im2im-uq [1]
0
192 / 192
2.8e−1 ± 1.4e−1
Conffusion [2]
0
192 / 192
2.7e−1 ± 1.6e−1
E-PUQ
0
192 / 192
1.9e−2 ± 1.0e−2
DA-PUQ
1.8e−2 ± 1.0e−4
39.4 ± 11.8 / 192
3.6e−8 ± 7.3e−8
RDA-PUQ
1.8e−2 ± 1.6e−3
55.6 ± 10.2 / 72.6 ± 30.3
1.3e−7 ± 8.8e−8
uncertainty volumes across tasks at 8x8 patch resolution. We
also provide visual representations of the uncertainty volume
maps for patches at varying resolutions in Figure 8.
The results shown in Figure 6 and Figure 7 demonstrate
that our method provides smaller uncertainty volumes, and
thus more confined uncertainty regions, when compared to
previous work in all tasks and patch resolutions, and while
satisfying the same statistical guarantees in all cases. More
specifically, Figure 6 compares our exact procedure, E-PUQ,
to baseline methods. Following this figure, one can see that
using the E-PUQ procedure we obtained an improvement of
∼×100 in the uncertainty volumes in colorization and an
improvement of ∼×10 in super-resolution and inpainting,
when applied to the highest resolution of 8x8. Additionally,
as the patch resolution increases, we observe a desired trend
of uncertainty volume reduction, indicating that our method
takes into account spatial correlation to reduce uncertainty.
Note that even a patch size of 1 × 1 brings a benefit in the
evaluated volume, due to the exploited correlation within the
three color channels. E-PUQ reduces trivially to im2im-uq [1]
and Conffusion [2] when applied to scalars (1×1×1 patches).
In Figure 7, we examine our approximation methods, DA-
PUQ and RDA-PUQ, in which we set a relatively small recon-
struction risk of β = 0.05. Observe the significantly smaller
uncertainty volumes obtained; this effect is summarized in
Fig. 6.
Local Experiments: A comparison of E-PUQ (see Section IV-B1)
with previous work – im2im-uq [1] and Conffusion [2]. These methods are
applied locally on patches with α = δ = 0.1. Each column corresponds to a
relevant metric (see Section V-A), and each row corresponds to a specific task.
The uncertainty volume was computed with ϵ = 1e −10. Results indicate
that our approach achieves superior uncertainty volume.
Table I as well. Figure 7 also portrays the dimensionality
of the uncertainty region used with our method using two
overlapping bars. The outer bar in yellow refers to the number
of PCs that need to be constructed, denoted as K in DA-
PUQ and
ˆK in RDA-PUQ. The smaller this number is,
the lower the test time computational complexity. The inner
bar in green refers to the average number of the adaptively
selected PCs, denoted as ˆk(x). A lower value of ˆk(x) indicates
better interpretability, as fewer PCs are used at inference than
those that were constructed. For example, in the colorization
task, it can be seen that the RDA-PUQ procedure is the
most computationally efficient methodology, requiring only
ˆK ≈12 PCs to be constructed at inference, while the DA-PUQ
procedure is the most interpretable results, with uncertainty
regions consisting of only ˆk(x) ∈{1, 2, 3} axes.
In all experiments demonstrated in Figure 6 and Figure 7, it
is noticeable that the standard deviation of the interval-size of
our approach is higher than that of the baseline methods. This
effect happens because a few intervals along the first few PCs

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
10
Fig. 7. Local Experiments: A comparison of DA-PUQ (see Section IV-B2)
and RDA-PUQ (see Appendix D), when applied locally on 8x8 patches with
α = δ = 0.1, β = 0.05 and q = 0.9. Each column corresponds to a
relevant metric (see Section V-A), and each row corresponds to a specific
task. The uncertainty volume was computed with ϵ = 1e −10. Here, the
dimensionality is presented by two overlapping bars, where the yellow bars
represent the distribution of K in DA-PUQ and ˆ
K in RDA-PUQ, and the
inner bars represent the distribution of ˆk(x) in both cases.
TABLE II
GLOBAL EXPERIMENTS: QUANTITATIVE COMPARISON OF THE MEANS
AND STANDARD DEVIATIONS OF OUR GLOBALLY APPLIED PUQ METHOD
IN THE COLORIZATION TASK, UTILIZING THE PROPOSED DA-PUQ (SEE
SECTION IV-B2) AND RDA-PUQ (SEE APPENDIX D) PROCEDURES.
Recons. Risk
Dim. ˆk(x) / K
Uncert. Volume
im2im-uq [1]
0
49152 / 49152
1.4e−1 ± 3.2e−2
Conffusion [2]
0
49152 / 49152
1.4e−1 ± 5.5e−2
DA-PUQ
5.0e−2 ± 1.1e−3
2.2 ± 0.93 / 100
1.2e−13 ± 5.0e−14
RDA-PUQ
4.3e−2 ± 2.8e−3
5.5 ± 4.5 / 22.3 ± 10.9
3.1e−13 ± 2.4e−13
are wider than those along the remaining PCs. However, the
majority of the interval sizes are significantly smaller, resulting
in a much smaller uncertainty volume. Interestingly, the un-
certainty intervals of the DA-PUQ and RDA-PUQ procedures
in Figure 7 exhibit larger standard deviation compared to the
E-PUQ procedure in Figure 6. We hypothesize that this is
caused when only a few intervals (e.g., 2 intervals) are used
for the calibration process while small miscoverage ratio is set
by the user (α = 0.1). As an example in the case of using 2
intervals with all samples of the calibration set, it is necessary
to enlarge all the intervals to ensure the coverage guarantee,
resulting in wider intervals over the first few PCs.
The heat maps presented in Figure 8 compare the un-
certainty volumes of our patch-based E-PUQ procedure to
baseline methods. Each pixel in the presented heat maps
corresponds to the value of Equation (3) evaluated on its
corresponding patch. The results show that as the patch resolu-
tion increases, pixels with strong correlation structure, such as
pixels of the background area, also exhibit lower uncertainty
volume in their corresponding patches. This indicates that the
proposed method indeed takes into account spatial correlation,
leading to reduced uncertainty volume.
C. Global Experiments on Images
We turn to examine the effectiveness and validity of DA-
PUQ and RDA-PUQ when applied to complete images at a
resolution of 128 × 128. In this case, the E-PUQ procedure
does not apply, as it requires computing and maintaining
d = 128×128×3 PCs. We present results for the colorization
task hereafter, and refer the reader to Appendix G for a similar
analysis related to super-resolution and inpainting.
While all PUQ procedures can be applied locally for any
task, working globally is more realistic in tasks that exhibit
strong pixel correlation. Under this setting, most of the image
variability could be represented via DA-PUQ or RDA-PUQ
while (i) maintaining a small reconstruction risk, and (ii)
using only a few PCs to assess the uncertainty of the entire
images. We should note that the tasks of super-resolution
and inpainting are less-matched to a global mode since they
require a larger number of PCs for an effective uncertainty
representation – more on this is discussed in Appendix G.
Figure 9 visually demonstrates the performance of our approx-
imation methods, also summarized in Table II. These results
demonstrate that our method provides significantly smaller
uncertainty volumes compared to our local results in Figure 7
and previous works, but this comes at the cost of introducing
a small reconstruction risk of up to β = 0.1. Observe how our
approximation methods improve interpretability: the uncer-
tainty regions consist of only 2-5 PCs in the full dimensional
space of the images. The DA-PUQ procedure produces the
tightest uncertainty regions; see the uncertainty volumes in
Table II. In addition, the mean interval-size with our proce-
dures is very small and almost equal to zero, indicating that
the constructed uncertainty regions are tight and narrow due
to strong correlation structure of pixels. However, similar to
the previous results, the standard deviation of interval-size is
spread across a wide range. This is because few of the first
PCs have wide intervals. The RDA-PUQ procedure is the most
computationally efficient as it required to construct only ∼30
PCs during inference to ensure statistical validity.
Figure 10 presents selected uncertainty regions that were
provided by our proposed RDA-PUQ procedure when applied
globally. As can be seen, the projected ground-truth images
using only ˆk(x) PCs results in images that are very close
to the originals. This indicates that the uncertainty region can
describe the spread and variability among solutions with small
reconstruction errors. The first two axes of our uncertainty
regions exhibit semantic content, which is consistent with a
method that accounts for spatial pixel correlation. The fact
that these PCs capture foreground/background or full-object
content highlights a unique strength of our approach. We
provide the importance weights of the first two PCs, indicating
impressive proportions of variability among projected samples
onto these components (see Section IV-A). For example, in
the third row, we observe that 77% of the variability in ˆPy|x
is captured by ˆv1(x), which mostly controls a linear color
range of the pixels associated with the hat in the image. In
Figure 11 we visually compare samples that were generated
from the corresponding estimated uncertainty regions, by
sampling uniformly a high dimensional point (i.e., an image)

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
11
x
y
im2im-uq
Conffusion
Our-1x1x3 Our-2x2x3 Our-4x4x3 Our-8x8x3
Fig. 8.
Local Experiments: Uncertainty volume maps for patches applied in image colorization (top), super-resolution (middle), and inpainting (bottom)
with E-PUQ, im2im-uq [1] and Conffusion [2]. Each pixel in the maps corresponds to the uncertainty volume, defined in Equation (3), of its corresponding
patch. These results expose the effectiveness of our method that incorporates spatial correlations, resulting in a reduction of the uncertainty volume.
Fig. 9. Global Experiments: A comparison of DA-PUQ (see Section IV-B2)
and RDA-PUQ (see Appendix D), when applied globally on the colorization
task with α = β = δ = 0.1 and q = 0.95. The uncertainty volume was
computed with ϵ = 1e −10.
x
y
Recons.
ˆv1(x)
ˆv2(x)
Fig. 10. Global Experiments: Visual presentation of uncertainty regions pro-
vided by RDA-PUQ (Appendix D) when applied globally for the colorization
task. The reconstructed image is given by ˆµ(x) + Pˆk(x)
i=1 ˆvi(x)T ycˆvi(x),
where yc := y −ˆµ(x). The values of ˆk(x), ˆw1(x) and ˆw2(x) are shown in
the top left corners of the corresponding columns.
from the corresponding hyper-rectangle. For further details
regarding this study, please refer to Appendix F. As can be
seen, the samples extracted from our uncertainty region are of
high perceptual quality, whereas im2im-uq [1] and Conffusion
[2] produce highly improbable images. This testifies to the fact
that our method provides much tighter uncertainty regions,
|—————— Our ——————-|
im2im-uq
Conffusion
Fig. 11. Global Experiments: Images sampled uniformly from the estimated
global uncertainty regions, referring to the colorization task. Using RDA-PUQ
results with high-perceptual images, while im2im-uq [1] and Conffusion [2]
produce unlikely images. These results indicate that our uncertainty regions
are significantly more confined than those of previous works.
whereas previous work results in exaggerated regions that
contain unlikely images. In addition to the above, we present
in Appendix J a visualization of the lower and upper corners
of the uncertainty regions produced by our method, comparing
them to those produced by previous work [1], [2].
VI. CONCLUDING REMARKS
This paper presents “Principal Uncertainty Quantification”
(PUQ), a novel and effective approach for quantifying un-
certainty in any image-to-image task. PUQ takes into account
the spatial dependencies between pixels in order to achieve
significantly tighter uncertainty regions. The experimental
results demonstrate that PUQ outperforms existing methods
in image colorization, super-resolution and inpainting, by
improving the uncertainty volume. Additionally, by allowing
for a small reconstruction error when recovering ground-
truth images, PUQ produces tight uncertainty regions with

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
12
a few axes and thus improves computational complexity and
interpretability at inference. As a result, PUQ achieves state-
of-the-art performance in uncertainty quantification for image-
to-image problems.
Referring to future research, more sophisticated choices that
rely on recent advancements in stochastic image regression
models could be explored, so as to improve the complexity
of our proposed approximation phase. Further investigation
into alternative geometries for uncertainty regions could be
interesting in order to reduce the gap between the provided
region of uncertainty and the high-density areas of the true
posterior distribution. This includes an option to divide the
spatial domain into meaningful segments, while minimizing
the uncertainty volume, or consider a mixture of Gaussians
modeling of the samples of the estimated posterior distribu-
tion. Additionally, exploring alternative diffusion models and
various conditional stochastic samplers presents an interesting
path for future investigation. This could involve comparing dif-
ferent conditional samplers, potentially offering an alternative
approach to the utilization of FID scores.
REFERENCES
[1] Anastasios N Angelopoulos, Amit Pal Kohli, Stephen Bates, Michael
Jordan, Jitendra Malik, Thayer Alshaabi, Srigokul Upadhyayula, and
Yaniv Romano.
Image-to-image regression with distribution-free un-
certainty quantification and applications in imaging. In International
Conference on Machine Learning, pages 717–730. PMLR, 2022.
[2] Eliahu Horwitz and Yedid Hoshen. Conffusion: Confidence intervals for
diffusion models. arXiv preprint arXiv:2211.09795, 2022.
[3] Roger Koenker and Gilbert Bassett Jr. Regression quantiles. Economet-
rica: journal of the Econometric Society, pages 33–50, 1978.
[4] Swami Sankaranarayanan, Anastasios N Angelopoulos, Stephen Bates,
Yaniv Romano, and Phillip Isola.
Semantic uncertainty intervals for
disentangled latent spaces. arXiv preprint arXiv:2207.10074, 3, 2022.
[5] Mehdi Mirza and Simon Osindero. Conditional generative adversarial
nets. arXiv preprint arXiv:1411.1784, 2014.
[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on
image synthesis. Advances in Neural Information Processing Systems,
34:8780–8794, 2021.
[7] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya
Ganguli. Deep unsupervised learning using nonequilibrium thermody-
namics. In International Conference on Machine Learning, pages 2256–
2265. PMLR, 2015.
[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion prob-
abilistic models. Advances in Neural Information Processing Systems,
33:6840–6851, 2020.
[9] Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Moham-
mad Norouzi, and Tim Salimans. Cascaded diffusion models for high
fidelity image generation. J. Mach. Learn. Res., 23(47):1–33, 2022.
[10] Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J
Fleet, and Mohammad Norouzi.
Image super-resolution via iterative
refinement.
IEEE Transactions on Pattern Analysis and Machine
Intelligence, 2022.
[11] Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan
Ho, Tim Salimans, David Fleet, and Mohammad Norouzi.
Palette:
Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference
Proceedings, pages 1–10, 2022.
[12] Vladimir Vovk, Alexander Gammerman, and Glenn Shafer. Algorithmic
learning in a random world, volume 29. Springer, 2005.
[13] Jing Lei, Max G’Sell, Alessandro Rinaldo, Ryan J Tibshirani, and
Larry Wasserman. Distribution-free predictive inference for regression.
Journal of the American Statistical Association, 113(523):1094–1111,
2018.
[14] Anastasios N Angelopoulos and Stephen Bates. A gentle introduction
to conformal prediction and distribution-free uncertainty quantification.
arXiv preprint arXiv:2107.07511, 2021.
[15] Anastasios N Angelopoulos, Stephen Bates, Emmanuel J Cand`es,
Michael I Jordan, and Lihua Lei. Learn then test: Calibrating predictive
algorithms to achieve risk control.
arXiv preprint arXiv:2110.01052,
2021.
[16] St´ephane Lathuili`ere, Pablo Mesejo, Xavier Alameda-Pineda, and Radu
Horaud. A comprehensive analysis of deep regression. IEEE transac-
tions on pattern analysis and machine intelligence, 42(9):2065–2081,
2019.
[17] Venkataraman Santhanam, Vlad I Morariu, and Larry S Davis. Gen-
eralized deep image to image regression. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pages 5609–
5619, 2017.
[18] Xinchen Yan, Jimei Yang, Kihyuk Sohn, and Honglak Lee.
At-
tribute2image: Conditional image generation from visual attributes. In
Computer Vision–ECCV 2016: 14th European Conference, Amsterdam,
The Netherlands, October 11–14, 2016, Proceedings, Part IV 14, pages
776–791. Springer, 2016.
[19] Karol Gregor, Ivo Danihelka, Alex Graves, Danilo Rezende, and Daan
Wierstra.
Draw: A recurrent neural network for image generation.
In International conference on machine learning, pages 1462–1471.
PMLR, 2015.
[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David
Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Gen-
erative adversarial networks. Communications of the ACM, 63(11):139–
144, 2020.
[21] Yang Song and Stefano Ermon.
Generative modeling by estimating
gradients of the data distribution.
Advances in neural information
processing systems, 32, 2019.
[22] Zahra Kadkhodaie and Eero P Simoncelli.
Solving linear inverse
problems using the prior implicit in a denoiser.
arXiv preprint
arXiv:2007.13640, 2020.
[23] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek
Kumar, Stefano Ermon, and Ben Poole.
Score-based generative
modeling through stochastic differential equations.
arXiv preprint
arXiv:2011.13456, 2020.
[24] Bahjat Kawar, Gregory Vaksman, and Michael Elad. Stochastic image
denoising by sampling from the posterior distribution. In Proceedings
of the IEEE/CVF International Conference on Computer Vision, pages
1866–1875, 2021.
[25] Bahjat Kawar, Gregory Vaksman, and Michael Elad.
Snips: Solving
noisy inverse problems stochastically. Advances in Neural Information
Processing Systems, 34:21757–21769, 2021.
[26] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. De-
noising diffusion restoration models. arXiv preprint arXiv:2201.11793,
2022.
[27] Zahra Kadkhodaie, Florentin Guth, St´ephane Mallat, and Eero P Si-
moncelli. Learning multi-scale local conditional probability models of
images. arXiv preprint arXiv:2303.02984, 2023.
[28] Yaniv Romano, Evan Patterson, and Emmanuel Candes. Conformalized
quantile regression. Advances in neural information processing systems,
32, 2019.
[29] Victor Chernozhukov, Kaspar W¨uthrich, and Yinchu Zhu. Distributional
conformal prediction. Proceedings of the National Academy of Sciences,
118(48):e2107794118, 2021.
[30] Matteo Sesia and Yaniv Romano. Conformal prediction using condi-
tional histograms. Advances in Neural Information Processing Systems,
34:6304–6315, 2021.
[31] Chirag Gupta, Arun K Kuchibhotla, and Aaditya Ramdas.
Nested
conformal prediction and quantile out-of-bag ensemble methods. Pattern
Recognition, 127:108496, 2022.
[32] Danijel Kivaranovic, Kory D Johnson, and Hannes Leeb.
Adaptive,
distribution-free prediction intervals for deep networks. In International
Conference on Artificial Intelligence and Statistics, pages 4346–4356.
PMLR, 2020.
[33] Stephen Bates, Anastasios Angelopoulos, Lihua Lei, Jitendra Malik,
and Michael Jordan. Distribution-free, risk-controlling prediction sets.
Journal of the ACM (JACM), 68(6):1–34, 2021.
[34] Anastasios N Angelopoulos, Stephen Bates, Adam Fisch, Lihua Lei, and
Tal Schuster. Conformal risk control. arXiv preprint arXiv:2208.02814,
2022.
[35] Jacopo Teneggi, Matthew Tivnan, Web Stayman, and Jeremias Sulam.
How to trust your diffusion model: A convex optimization approach
to conformal risk control.
In International Conference on Machine
Learning, pages 33940–33960. PMLR, 2023.
[36] Yves Meyer.
Orthonormal wavelets.
In Wavelets: Time-Frequency
Methods and Phase Space Proceedings of the International Conference,
Marseille, France, December 14–18, 1987, pages 21–37. Springer, 1990.
[37] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive
growing of gans for improved quality, stability, and variation.
arXiv
preprint arXiv:1710.10196, 2017.

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
13
Fig. 12. A visual representation demonstrating the intuition behind utilizing principal components (PCs) as the basis, ˆB(x), in Equation (1) for the colorization
task. The left part illustrates that the PCs incorporate spatial correlation, with ˆv1(x) primarily controlling the hat color, ˆv2(x) governing the background
color, and ˆv3(x) influencing the clothing color. On the right side, an illustration of the uncertainty region is presented, composed of these axes, where the
origin is ˆµ(x), and each image is defined by ˆµ(x) + ˆvi(x)T yc + a, where yc := y −ˆµ(x), and a ∈R is a controllable parameter that moves along the axis.
[38] Tero Karras, Samuli Laine, and Timo Aila.
A style-based generator
architecture for generative adversarial networks. In Proceedings of the
IEEE/CVF conference on computer vision and pattern recognition, pages
4401–4410, 2019.
APPENDIX
A. Visualizing the Principal Component Vectors
Figure 12 depicts the role of the Principal Component (PCs)
vectors in the context of the image colorization task. This
figure provides an intuition behind employing these vectors for
the uncertainty quantification. We show the estimation of the
first three PCs using our globally applied PUQ and visualize
the uncertainty region formed by these axes. Our approach
facilitates efficient exploration within the uncertainty region,
thanks to the linear axes that incorporate spatial correlation,
as illustrated by the visualization of ˆv1(x), ˆv2(x), and ˆv3(x).
B. Coverage Loss Justification
This section aims to justify our choice for the loss-function
for tuning λ in Equation (5), and the weights used in it, ˆwi(x).
Recall, this expression is given as:
L(x, y; λ) :=
d
X
i=1
ˆwi(x) · 1
n
ˆvi(x)T y ̸∈Tλ(x; ˆB(x))i
o
.
Our starting point is the given d-dimensional hyper-
rectangle obtained from the approximation phase, oriented
along the d PC directions. This shape serves as our ini-
tially estimated uncertainty region. Given the calibration data,
Scal := {(xi, yi)}n
i=1, our goal is to inflate (or deflate, if this
body proves to be exaggerated) this shape uniformly across
all axes so that it contains the majority of the ground truth
examples.
Focusing on a single pair from this dataset, (x, y), the
degraded image x is used to ignite the whole approximation
phase, while the ground truth y serves for assessing the
obtained hyper-rectangle, by considering the projected coor-
dinates {ˆvi(x)T yc}d
i=1, where yc := y −ˆµ(x). The following
function measures a potential deviation in the i-th axis,
hi(x, y) := max

ˆvi(x)T yc −ˆu(x)i, 0
	
(11)
+ max

−ˆvi(x)T yc + ˆl(x)i, 0
	
.
Written differently, this expression is also given by
hi(x, y) :=





ˆvi(x)T yc −ˆu(x)i
if ˆvi(x)T yc > ˆu(x)i > 0
ˆl(x)i −ˆvi(x)T yc
if ˆvi(x)T yc < −ˆl(x)i < 0
0
otherwise.
If positive, this implies that in this axis the example spills
outside the range of the rectangle, and the value itself is the
distance from it’s border.
The following expression quantifies the weighted amount
of energy that should be invested in projecting back the
{ˆvi(x)T y}d
i=1 coordinates to the closest border point:
Energy(x, y) =
d
X
i=1
ˆσi(x)2hi(x, y)2.
(12)
Note that in our weighting we prioritize high-variance axes,
in which deviation from the boundaries is of greater impact.
Naturally, we should tune λ, which scales ˆu(x)i and ˆl(x)i,
so as to reduce this energy below a pre-chosen threshold,
thus guaranteeing that the majority of ground truth images fall
within the hyper-rectangle. While this expression is workable,
it suffers from two shortcomings: (1) It is somewhat involved
to compute; and (2) The threshold to use with it is hard to
interpret and thus to choose. Therefore, similar to previous
approaches [1], [2], we opted in this work for a binary version
of Equation (11) of the form
bi(x, y) :=
(
1
if hi(x, y) > 0
0
otherwise.
(13)
In addition, we divide the energy expression, defined in
Equation (12), by the sum of squares of all the singular values,
and this way obtain exactly L(x, y; λ) as in Equation (5).

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
14
Observe that, by definition, we get that 0 ≤L(x, y; λ) ≤1,
where the bottom bound corresponds to a point fully within
the rectangle, and the upper bound for the case where the
point is fully outside in all axes. Therefore, thresholding
the expectation of this value with α ≪1 is intuitive and
meaningful.
C. Reconstruction Loss Justification
This section aims to discuss our choice for the loss function
for tuning λ1 in Equation (8), given by
L1(x, y; λ1) := ˆQq







ˆk(x;λ1)
X
j=1
ˆvj(x)T ycˆvj(x) −yc

i



d
i=1


.
Recall the process: We begin with K ≤d PCs obtained from
the approximation phase, and then choose ˆk(x; λ1) ≤K of
them as instance-specific number of PCs for the evaluation of
the uncertainty.
Given a calibration pair (x, y), x is used to derive
ˆk(x; λ1), defining a low-dimensional subspace
ˆV (x)
:=
[ˆv1(x), . . . , ˆvˆk(x;λ1)(x)] ∈Rˆk(x;λ1)×d. This, along with the
conditional-mean, ˆµ(x), represent Py|x as an affine subspace.
The ground-truth image y is then projected onto this slab via:
Projection(y) := ˆµ(x) + ˆV (x) ˆV (x)T yc
(14)
= ˆµ(x) +
ˆk(x;λ1)
X
j=1
ˆvj(x)T ycˆvj(x) ,
where yc := y −ˆµ(x).
The parameter λ1 should be tuned so as to guarantee that
this projection entails a bounded error, dist(y, Projection(y))
in expectation. A natural distance measure to use here is
the L2-norm of the difference, which aligns well with our
choice to use SVD in the approximation phase. However,
L2 accumulates the error over the whole support, thus los-
ing local interpretability. An alternative is using L∞which
quantifies the worst possible pixelwise error induced by the
low-dimensional projection,
dist(y, Projection(y)) :=







ˆk(x;λ1)
X
j=1
ˆvj(x)T ycˆvj(x) −yc







∞
.
While this measure is applicable in many tasks, there are
cases (e.g., inpainting) in which controlling a small maximum
error requires the use of a large number of PCs, ˆk(x; λ1). To
address this, we propose a modification by considering the
maximum error over a user-defined ratio of pixels, q ∈(0, 1),
a value close to 1. This is equivalent to determining the q-th
empirical quantile, ˆQq, of the error values among the pixels,
providing a more flexible and adaptive approach, which also
aligns well with the rationale of uncertainty quantification, in
which the statistical guarantees are given with probabilistic
restrictions.
D. Reduced Dimension-Adaptive PUQ
The DA-PUQ procedure (see Section IV-B2) reduces the
number of PCs to be constructed to K ≤d while using
ˆk(x; ˆλ1) ≤K PCs, leading to increased efficiency in both time
and space during inference. However, determining manually
the smallest K value that can guarantee both Equation (2)
and Equation (4) with high probability can be challenging.
To address this, we propose an expansion of the DA-PUQ
procedure; the Reduced Dimension-Adaptive PUQ (RDA-
PUQ) procedure that also controls the maximum number
of PCs required for the uncertainty assessment. While this
approach is computationally intensive during calibration, it is
advantageous for inference as it reduces the number of samples
required to construct the PCs using Algorithm 2.
Specifically, for each input instance x and its corresponding
ground-truth value y in the calibration data, we use the estima-
tors obtained in the approximation phase, to estimate ˆKλ3 PCs
of possible solutions, denoted by ˆB(x), their corresponding
importance weights, denoted by ˆw(x), the conditional mean
denoted by ˆµ(x), and the lower and upper bounds denoted by
˜l(x) and ˜u(x), respectively. Note that these estimates are now
depend on λ3, we omit the additional notation for simplicity.
Then, for each choice of λ3, we use these ˆKλ3-dimensional
estimates exactly as in the DA-PUQ procedure to achieve both
the coverage and reconstruction guarantees of Equation (2) and
Equation (4) with high probability.
Similar to previous approaches, we aim to minimize the
uncertainty volume, defined in Equation (3), for the scaled
ˆKλ3-dimensional intervals where any additional axis (d−ˆKλ3
axes) is fixed to zero. We denote the uncertainty volume
in this setting as Vλ1,λ2,λ3. The minimization of Vλ1,λ2,λ3
is achieved by minimizing λ1, λ2 and λ3, while ensuring
that the guarantees of Equation (2) and Equation (4) are
satisfied with high probability. This can be provided using a
conformal prediction scheme, for example, through the LTT
[15] calibration scheme, which ensures that the following
holds:
P

E[L1(x, y; ˆλ1, ˆλ3)] ≤β
E[L2(x, y; ˆλ1, ˆλ2, ˆλ3)] ≤α

≥1 −δ ,
(15)
where ˆλ1, ˆλ2 and ˆλ3 are the minimizers for the uncertainty
volume among valid calibration parameter results, ˆΛ, obtained
through the LTT procedure. Note that the loss functions,
L1 and L2, in the above are exactly those of the DA-PUQ
procedure, defined in Equation (8) and Equation (9), while
replacing K with ˆKλ3.
Intuitively, Equation (15) guarantees that a fraction q of the
ground-truth pixel values is recovered with an error no greater
than β using no more than ˆKˆλ3 principal components, and a
fraction of more than 1−α of the projected ground-truth values
onto the first ˆk(x; ˆλ1) principal components (out of ˆKˆλ3) are
contained in the uncertainty intervals, with a probability of at
least 1−δ. The RDA-PUQ procedure is formally described in
Algorithm 5.
E. Experimental Details
This section provides details of the experimental method-
ology employed in this study, including the datasets used,
architectures implemented, and the procedural details and
hyperparameters of our method.
1) Datasets and Preprocessing: Our machine learning sys-
tem was trained using the Flickr-Faces-HQ (FFHQ) dataset
[38], which includes 70,000 face images at a resolution of
128x128. We conducted calibration and testing on the CelebA-
HQ (CelebA) dataset [37], which also consists of face images

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
15
Algorithm 5 Reduced Dimension-Adaptive PUQ Proc.
Input: Calibration set Scal
:= {xi, yi}n
i=1. Scanned calibration
parameter values Λ1 ←[1 . . . λ1max], Λ2 ←[1 . . . λ2max] and
Λ3 ←[1 . . . λ3max]. Maximal PCs number K ≤d. Approx-
imation phase estimators ˆB, ˆw, ˆµ, ˜u, ˜l. Recovered pixels ratio
q ∈(0, 1). Reconstruction error β ∈(0, 1). Misscoverage ratio
α ∈(0, 1). Calibration error level δ ∈(0, 1).
1: for (x, y) ∈Scal do
2:
for λ3 ∈Λ3 do
▷Reduce dimensionality
3:
ˆKλ3 ←⌊K · λ3⌋
4:
ˆB(x), ˆw(x), ˆµ(x), ˜u(x), ˜l(x) ←Apply Algorithm 2
to x, with the choice of ˆKλ3 samples
5:
for λ1 ∈Λ1 do
▷Compute adaptive dimensionality, Equation (7)
6:
ˆk(x; λ1, λ3) ←
mink

k : P ˆ
Kλ3
i=1 ˆwi(x) ≥λ1

▷Compute reconstruction loss, Equation (8)
7:
yc ←y −ˆµ(x)
8:
L1(x, y; λ1, λ3) ←
ˆQq

Pˆk(x;λ1,λ3)
j=1
ˆvj(x)T ycˆvj(x) −yc

i
d
i=1

9:
for λ2 ∈Λ2 do
▷Scale uncertainty intervals
10:
ˆu(x) ←λ2˜u(x) and ˆl(x) ←λ2˜l(x)
11:
Tλ2(x; ˆB(x)) ←
Equation (1) using ˆµ(x), ˆu(x), ˆl(x)
▷Compute weighted coverage loss, Equation (5)
12:
L2(x, y; λ1, λ2, λ3) ←Pˆk(x;λ1,λ3)
i=1
ˆwi(x)·
1
n
ˆvi(x)T y ̸∈Tλ2(x; ˆB(x))i
o
13:
end for
14:
end for
15:
end for
16: end for
17: ˆΛ
←
Extract
valid
λs
from
LTT
[15]
applied
on
{(L1(x, y; λ1, λ3), L2(x, y; λ1, λ2, λ3))
:
(x, y) ∈Scal, λ1 ∈Λ1, λ2 ∈Λ2, λ3 ∈Λ3}
at
risk
levels
(β, α).
▷Compute the minimizers for the uncer. volume, Equation (3)
18: ˆλ1, ˆλ2, ˆλ3 ←arg minλ1,λ2,λ3∈ˆΛ
n
1
n
Pn
i=1 Vλ1,λ2(xi; ˆB(xi))
o
Output: Given a new instance x ∈X, obtain valid uncertainty
intervals for it, Tˆλ2(x; ˆB(x)) over ˆk(x; ˆλ1) ≤ˆKˆλ3 PCs.
and was resized to match the resolution of our training data. To
this end, we randomly selected 2,000 instances from CelebA,
of which 1,000 were used for calibration and 1,000 for testing.
For the colorization experiments, a grey-scale transformation
was applied to the input images. For the super-resolution
experiments, patches at a resolution of 32x32 were averaged
to reduce the input image resolution by a factor of 4 in
each dimension. For the inpainting experiments, we randomly
cropped pixels from the input images during the training phase,
either in squares or irregular shapes; while for the calibration
and testing data, we cropped patches at a resolution of 64x64
at the center of the image.
2) Architecture and Training: In all our experiments, we
applied the approximation phase using recent advancements in
conditional image generation through diffusion-based models,
while our proposed general scheme in Algorithm 1 can accom-
modate any stochastic regression solvers for inverse problems,
such as conditional GANs [5]. In all tasks, we utilized the
framework for conditional diffusion-based models proposed
in the SR3 work [10], using a U-Net architecture. For each of
the three tasks, we trained a diffusion model separately and
followed the training regimen outlined in the code of [10].
To ensure a valid comparison with the baseline methods, we
implemented them using the same architecture and applied the
same training regimen. All experiments, including the baseline
methods, were trained for 10,000 epochs with a batch size of
1,024 input images.
3) PUQ Procedures and Hyperparameters: Our experi-
mental approach follows the general scheme presented in
Algorithm 1 and consists of 2 sets of experiments: local
experiments on patches and global experiments on entire
images. For the local experiments, we conducted 4 experi-
ments of the E-PUQ procedure (detailed in Section IV-B1) on
RGB patch resolutions of 1x1, 2x2, 4x4, and 8x8. We used
K = 3, K = 12, K = 48, and K = 192 PCs for each
resolution, respectively. We set α = δ = 0.1 to be the user-
specified parameters of the guarantee, defined in Equation (6).
In addition, we conducted another 2 experiments of the DA-
PUQ (detailed in Section IV-B2), and RDA-PUQ (detailed in
Appendix D) procedures on RGB patch resolution of 8x8. We
set q = 0.9, β = 0.05 and α = δ = 0.1, to be the user-
specified parameters of the guarantees of both Equation (10)
and Equation (15). In total, we conducted 18 local experiments
across three tasks. For the global experiments, we used entire
images at a resolution of 128x128, in which we applied the
DA-PUQ and the RDA-PUQ procedures. As global working
is suitable for tasks that exhibit strong pixel correlation, we
applied these experiment only on the task of image coloriza-
tion. We set q = 0.95, β = α = δ = 0.1, to be the user-
specified parameters of the guarantees of both Equation (10)
and Equation (15). Both locally and globally, for the DA-PUQ
and RDA-PUQ experiments, we used K = 100 PCs in the
colorization task and K = 200 PCs in super-resolution and
inpainting. We note that in the RDA-PUQ experiments, we
used ˆK PCs during inference, as discussed in Appendix D. In
all experiments we used ϵ = 1e −10 for the computation of
the uncertainty volume, defined in Equation (3).
F. Comparative Samples from Uncertainty Regions
We provide here more details referring to the experiment
involving a visualization of samples drawn from the uncer-
tainty regions of baseline methods [1], [2] and our proposed
approach. We note that the baseline methods lack such an
experiment.
This experiment was conducted across entire images, show-
ing that our uncertainty region is much tighter, containing
highly probable image candidates, compared to the pixelwise
baseline methods. These methods tend to generate exaggerated
uncertainty regions that encompass a range of noisy images,
diverging from the posterior distribution of images given
a measurement. Our success in producing more confined
regions, encompassing the ground truth within them, is a direct
consequence of the incorporation of spatial correlations.
To justify this claim, we trained the identical architecture
for each baseline method and applied the same training regime

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
16
Fig. 13. Global Experiments: A comparison of DA-PUQ (see Section IV-B2)
and RDA-PUQ (see Appendix D), when applied globally on super-resolution
and inpainting tasks with α = β = δ = 0.1, where in super-resolution we
set q = 0.95 and in inpainting we set q = 0.8. The uncertainty volume was
computed with ϵ = 1e −10.
x
y
Recons.
ˆv1(x)
ˆv2(x)
Fig. 14.
Global Experiments: Visual presentation of uncertainty regions
provided by DA-PUQ when applied globally for the super-resolution task.
The reconstructed image is given by ˆµ(x) + Pˆk(x)
i=1 ˆvi(x)T ycˆvi(x), where
yc := y −ˆµ(x). The values of ˆk(x), ˆw1(x) and ˆw2(x) are shown in the top
left corners of the corresponding columns.
that was utilized in our approach, leveraging the official code
of both methods. Each baseline method generates uncertainty
intervals via pixel-based uncertainty maps, which is equivalent
to our general definition of uncertainty intervals defined by
Equation (1), while employing standard basis vectors. There-
fore, we uniformly sampled values within the uncertainty
intervals of each approach, including our own, and showcased
the resulting images.
G. Additional Global Experiments
In Section V we presented global studies of our DA-PUQ and
RDA-PUQ, focusing on their deployment in the colorization
task. Here, we extend this analysis by presenting additional
global studies for super-resolution and inpainting, ensuring a
more comprehensive assessment of our methods.
It is worth mentioning that the tasks of super-resolution
and inpainting differ in nature from colorization. In super-
resolution and inpainting, the decay in the associated singular
values of each posterior distribution occurs relatively slowly,
|—————— Our ——————-|
im2im-uq
Conffusion
Fig. 15. Global Experiments: Images sampled uniformly from the estimated
global uncertainty regions, referring to the super-resolution task. Using DA-
PUQ results with high-perceptual images, while im2im-uq [1] and Conffusion
[2] produce unlikely images. These results indicate that our uncertainty regions
are significantly more confined than those of previous works.
x
y
Recons.
ˆv1(x)
ˆv2(x)
Fig. 16.
Global Experiments: Visual presentation of uncertainty regions
provided by DA-PUQ when applied globally for the inpainting task. The
reconstructed image is given by ˆµ(x)+Pˆk(x)
i=1 ˆvi(x)T ycˆvi(x), where yc :=
y −ˆµ(x). The values of ˆk(x), ˆw1(x) and ˆw2(x) are shown in the top left
corners of the corresponding columns.
indicating a more localized impact. This contrasts with the
colorization task, where the decay in singular values is more
rapid and pronounced, implying stronger pixel correlations.
Consequently, constructing global representations of uncer-
tainty regions in the colorization task is effective, with strong
guarantees involving small reconstruction errors over a large
number of pixels using far fewer axes.
Nevertheless, we have applied our DA-PUQ and RDA-PUQ
globally to the tasks of super-resolution and inpainting, and
the quantitative results are depicted in Figure 13. In both

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
17
|—————— Our ——————-|
im2im-uq
Conffusion
Fig. 17. Global Experiments: Images sampled uniformly from the estimated
global uncertainty regions, referring to the inpainting task. Using DA-PUQ
results with high-perceptual images, while im2im-uq [1] and Conffusion [2]
produce unlikely images. These results indicate that our uncertainty regions
are significantly more confined than those of previous works.
studies, we utilized 1500 samples for the calibration data
and 500 samples for the test data. This is different from
the global colorization study, where we used 1000 samples
for both calibration and test data. This adjustment aims to
narrow the gap between the true risks of unseen data and
the concentration bounds employed in the calibration scheme,
ultimately allowing us to provide more robust guarantees,
including small coverage and reconstruction risks with high
probability.
Additionally, we set α = β = δ = 0.1 in both studies.
However, in the super-resolution study, we maintained q =
0.95, which is consistent with the setting used in the global
colorization. In contrast, for inpainting, we chose q = 0.8,
indicating a softer reconstruction guarantee applicable to 80%
of the pixels within the missing window.
The results depicted in Figure 13 reveal that our method
consistently yields significantly smaller uncertainty volumes
compared to our local results presented in Section V and pre-
vious research. However, this reduction in uncertainty volume
comes at the cost of introducing a reconstruction risk, reaching
a maximum of β = 0.1, which applies to 95% of the pixels
in super-resolution and 80% of the pixels in inpainting.
Observe the improvement in interpretability that our DA-
PUQ method brings to the table. Notably, the uncertainty
regions generated by DA-PUQ consist of only ∼10 PCs within
the full-dimensional space of the images. In contrast, the
uncertainty regions produced by our RDA-PUQ experiments
comprise ∼100 PCs, indicating a slower decay in the singu-
lar values of the posterior distribution associated with each
uncertainty region.
Figures 14 and
16 showcase selected uncertainty regions
provided by our proposed DA-PUQ when applied globally
to the super-resolution and inpainting tasks, respectively. No-
tably, the projected ground-truth images using only ˆk(x) PCs
resemble the originals. This observation indicates that the un-
certainty region effectively captures the spread and variability
among solutions while maintaining satisfying reconstruction
errors.
In the inpainting task presented in Figure 16, the first two
axes of our uncertainty regions exhibit semantic content, an
indicator to our method’s ability to consider spatial pixel
correlation. The PCs effectively capture features such as
sunglasses, eyebrows, and forehead, highlighting the unique
strength of our approach in terms of interpretability. However,
in the super-resolution task depicted in Figure 14, localized
PCs emerge, implying that only a few pixel values are affected
in each axis of uncertainty.
In Figures 15 and 17, we visually compare samples gen-
erated from the corresponding estimated uncertainty regions.
These samples are obtained by uniformly sampling from the
respective hyper-rectangle, then transforming to the image
domain. For further details regarding this study, please refer
to Appendix F. This visual comparison4 shows that sam-
ples extracted from our uncertainty regions exhibit higher
perceptual quality compared to those generated by im2im-
uq [1] and Conffusion [2]. This observation implies that our
method provides tighter uncertainty regions, whereas previous
work results in exaggerated uncertainty regions that contain
improbable images.
H. Ablation Study
We turn to introduce an ablation study on the user-specified
parameters: α, β, q, and δ. These parameters are used in the
context of the statistical guarantees provided by our proposed
method, and our objective is to offer a comprehensive under-
standing of how to select these parameters and their resulting
impact on performance. To elaborate, α ∈(0, 1) is employed
to ensure coverage, as indicated in Equation (2), while both
β ∈(0, 1) and q ∈(0, 1) play a role in establishing the recon-
struction guarantee, as defined in Equation (4). Additionally,
the parameter δ ∈(0, 1) is used for controlling the error rate
associated with both guarantees over the calibration data.
An effective calibration process relies on these user-
specified parameters, α, β, and δ approaching values close
to zero, while q should ideally approach 1. The choice of
these parameters is guided by the amount of available calibra-
tion data. In cases where a substantial calibration dataset is
accessible, it becomes feasible to establish robust statistical
assessments. This is manifested by the ability to employ
smaller values for α, β, and δ, while favoring a higher value
for q. For instance, achieving a 90% coverage rate (α = 0.1),
with a reconstruction error threshold of 5% (β = 0.05) across
95% of the image pixels (q = 0.95) serves as an illustrative
example of such robust assessments.
It is worth noting that our primary aim in this work is
to enhance the interpretability of the uncertainty assessment
within the context of the inverse problems. This is achieved
through the methods we propose, DA-PUQ and RDA-PUQ.
Consequently, we strive to provide the user with a more
4These results are better seen by zooming in, and especially so for the
super-resolution task.

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
18
Fig. 18. An ablation study of DA-PUQ in a locally applied colorization task on 8x8 RGB patch resolution. We examine the user-defined parameters α, β, q,
and δ, showcasing their impact on the mean uncertainty volume, mean dimensionality, coverage risk, and reconstruction risk. The default values are α = 0.05,
β = 0.05, q = 0.95, and δ = 0.05 using K = 200 PCs. Our results are depicted in green, with threshold values for guarantees highlighted in dashed black.
concise set of uncertainty axes, referred to as the selected
axes denoted as ˆB(x) = {ˆv1(x), ˆv2(x), . . . , ˆvˆk(x)(x)}. Our
approach for selecting the reconstruction guarantee is geared
towards a balance between precision and interpretability. On
one hand, we aim to establish a robust and stringent recon-
struction guarantee to accurately capture the uncertainty of
the posterior distribution across the d dimensions. On the
other hand, we aim to incorporate a softer reconstruction
guarantee that results in providing fewer axes of uncertainty
thus enhancing interpretability.
Figure 18 illustrates the quantitative results of the abla-
tion study conducted on DA-PUQ, where we investigate the
influence of the user-defined parameters, α, β, q, and δ,
on DA-PUQ’s performance. It is important to note that the
default settings in each study are the following: α = 0.05,
β = 0.05, q = 0.95, and δ = 0.05, representing a spectrum of
strengthening and softening parameter choices.
Analyzing the results, we observe that α primarily controls
the coverage aspect. As α increases, the uncertainty intervals
become narrower, leading to more tightly constrained uncer-
tainty regions. This trend is evident in the reduction of the
uncertainty volume metric. However, it is noteworthy that α
has no impact on the reconstruction error, as the dimension and
reconstruction risk remain relatively consistent across different
choices of α.
The parameter β influences the reconstruction error, with
even slight alterations affecting the number of selected axes,
denoted as ˆk(x). On the other hand, the parameter q has a rel-
atively minor effect on performance. Adjusting q does impact
the uncertainty volume, with smaller dimensions resulting in
a reduction in uncertainty volume. Higher values for q lead
to the selection of more PCs for the uncertainty assessment,
involving more pixels in the reconstruction guarantee.
Referring to the parameter δ, we observe minor changes
in the coverage risk, while the reconstruction risk undergoes
more significant changes. This suggests that errors in the
uncertainty assessments tend to be more focused on the
reconstruction guarantee rather than the coverage guarantee.
In terms of the precision and interpretability trade-off, the
ideal scenario would involve selecting the smallest possible
value for β, as demonstrated in Figure 18 with β = 0.02.
However, such a stringent guarantee would require the use
Fig. 19. An analysis illustrating the precision-complexity trade-off of global
DA-PUQ in the colorization task. The complexity aspect is presented by
varying values of K, while precision is represented by the mean number
of PCs provided for the user, denoted as ˆk(x). The parameters setting is:
α = 0.1, β = 0.1, q = 0.95 and δ = 0.1. Smaller ˆk(x) values correspond
to more accurate PCs, while lower values of K indicate improved method
complexity.
of approximately 188.3 PCs, which can harm interpretability.
In this case, a softer guarantee, such as β = 0.04, results in
the use of only around 5.13 PCs, striking a more balanced
trade-off between precision and interpretability.
I. Precision and Complexity Trade-off
We now discuss the trade-off between precision and com-
plexity in our work. Precision here stands for the ability to
accurately capture uncertainty within the posterior distribution
across the d dimensions, as reflected by the reconstruction
risk. Conversely, complexity involves two key aspects: the
complexity associated with our diffusion model for generating
posterior samples and the computational demands of PCA.
Both of these aspects are influenced by the chosen value of
K ≤d, which serves both as the number of drawn samples
and the overall number of initial PC’s to work with. As for the
complexity: (i) Assuming that a single diffusion iteration can
be achieved in constant time, the complexity of generating K
samples is given by O(IK), where I ∈N denotes the number
of iterations in the diffusion algorithm; and (ii) For the PCA,
the complexity is provided by O(d2K + d3), where K refers
to the number of PCs.
Therefore, the value of K ≤d plays a pivotal role in
governing the precision-complexity trade-off across all our
proposed methods: E-PUQ, DA-PUQ, and RDA-PUQ, all of
which involve sampling and PCA. The greater the number of
PCs employed, the more precise our uncertainty assessment, at
the expense of computational complexity, as discussed above.

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
19
x
y
ˆµ(x)
lo
up
lo
up
lo
up
|——— Our ———| |—– im2im-uq ——| |—– Conffusion —–|
Fig. 20.
Visual analysis (left) of the lower and upper corners generated by our global DA-PUQ across three tasks: image colorization, super-resolution,
and inpainting. On the right side, a 2D example illustrates an uncertainty region constructed by our approach in contrast to one produced by the pixelwise
approach, demonstrating the distinction between the lower and upper corners in each approach.
In the case of E-PUQ, we achieve a complete uncertainty
assessment at the computation cost of K = d. This leads to
the effective reduction of the reconstruction risk to zero for
all image pixels. However, it’s essential to recognize that in
practical scenarios, such as the global applications illustrated
in our empirical study in Section V and in Appendix G, con-
ducting sampling and PCA with K = d on high-dimensional
data, such as d = 3 × 128 × 128, becomes unfeasible.
Hence, we introduced DA-PUQ to enhance the method’s
computational efficiency by allowing K ≪d, thereby mitigat-
ing complexity. To further enhance interpretability, we intro-
duced ˆk(x) in Equation (7), which aims to reduce the number
of PCs to be used (out of the already constructed K PCs),
while ensuring that the reconstruction guarantee is maintained
with as few PCs as possible. This balance is demonstrated
in Figure 19, where various values of K showcase that the
reconstruction risk remains unaffected, yet more uncertainty
axes, ˆk(x), are needed to uphold this equilibrium.
Given the challenge of determining an appropriate value
for K that ensures robust statistical guarantees, we introduced
RDA-PUQ. This variant tunes K to the lower value that fulfills
the necessary statistical guarantees.
In Figure 19, we visually depict the precision-complexity
trade-off through experiments involving different values of K
in the context of DA-PUQ’s global application in colorization.
Here, we illustrate the complexity of our method through the
selection of varying K values, where higher values imply
higher complexity, as they require the generation of more sam-
ples and the construction of more PCs. Meanwhile, precision is
assessed by examining the resulting ˆk(x) values, where higher
ˆk(x) values correspond to situations where the uncertainty
assessment is less accurate, signifying a higher reconstruction
risk when employing all the K PCs. Consequently, more axes
are needed to maintain a balanced risk.
J. Lower and Upper Corners
We conclude this paper by providing a visual comparative
analysis of lower and upper corners within uncertainty regions
applied globally across the three tasks: image colorization,
super-resolution, and inpainting.
Formally, the lower and upper corners within the image
domain of an uncertainty region, are constructed using the
intervals outlined in Equation (1). These are defined as the
following expressions: the lower corner is defined as ˆµ(x) −
ˆV (x)ˆl(x), and the upper corner is defined as ˆµ(x)+ ˆV (x)ˆu(x).
Here, ˆV (x) is a matrix comprising of the K selected PCs from
ˆB(x) as it’s columns, and ˆl(x), ˆu(x) are column vectors of
length K.
For example, when choosing to work within the pixel
domain by selecting the standard basis, ˆB(x) = e1, e2 . . . ed,
where ei ∈Rd represents the one-hot vector with a value of
1 in the ith entry, the lower and upper corners align with the
lower and upper bounds presented in prior work [1], [2] that
operates in the pixel domain.
It is essential to note that in our work, we use the term
corners to emphasize that the lower and upper corners in the
image domain do not establish intervals. This is in contrast
to the pixelwise approach, which constructs intervals around
each pixel, making the terminology “lower and upper bound
images” more conceptually suitable.
In Figure 20 (right), we illustrate the difference between the
lower and upper corners of our uncertainty region (depicted as
green dots) and the lower and upper bounds of the pixelwise
approach (depicted as red dots). This comparison is presented
through a 2D example, demonstrating the process of construct-
ing an uncertainty region for a posterior distribution using our
method, in contrast to the pixelwise approach.
In Figure 20 (left), we provide a visual comparison between
the lower and upper corners generated by DA-PUQ and the
lower and upper bounds produced by [1], [2]. It is evident that
our lower and upper corners exhibit a higher perceptual quality
compared to the lower and upper bounds from earlier pixel
domain approaches. This suggests that the lower and upper
corners represent more probable samples than those generated
by the pixelwise approach. Therefore, the uncertainty regions
constructed by our approach are more confined compared to
those constructed using the pixelwise approach.
Interestingly, by traversing between the two corners of DA-
PUQ by their convex combination, we essentially walk in the
main “boulevard” of the uncertainty region. Figure 21 shows
the resulting images in this path for the three applications con-

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
20
t = 0.0
t = 0.1
t = 0.2
t = 0.3
t = 0.4
t = 0.5
t = 0.6
t = 0.7
t = 0.8
t = 0.9
t = 1.0
Fig. 21.
Visualization of the main “boulevard” within the uncertainty regions of DA-PUQ applied globally across three tasks: image colorization, super-
resolution, and inpainting. The traversal along this path is obtained by a convex combination of the lower and upper corners, given by: (1−t)·lo(x)+t·up(x),
where t ∈[0, 1].
sidered: image colorization, super-resolution and inpainting.
