Real-Time Editing of Path-Traced Scenes with Prioritized Re-Rendering

Annalena Ulschmid

1 a

, Bernhard Kerbl

1 b

, Katharina Kr

osl

1,2 c

and Michael Wimmer

1 d

TU Wien, Austria

VRVis Forschungs-GmbH, Austria

Keywords:

Path Tracing, Scene Editing, Adaptive Rendering, Empirical Studies.

Abstract:

With recent developments in GPU ray tracing performance and (AI-accelerated) noise reduction techniques,

Monte Carlo Path Tracing at real-time rates becomes a viable solution for interactive 3D scene editing, with

growing support in popular software. However, even for minor edits (e.g., adjusting materials or moving

small objects), current solutions usually discard previous samples and the image formation process is started

from scratch. In this paper, we present two adaptive, priority-based re-rendering techniques with incremen-

tal updates, prioritizing the reconstruction of regions with high importance, before gradually moving to less

important regions. The suggested methods automatically identify and schedule sampling and accumulation of

immediately affected regions. An extensive user study analyzes whether such prioritized renderings are bene-

ﬁcial to interactive scene editing, comparing them with same-time conventional re-rendering. Our evaluation

shows that even with simple priority policies, there is a signiﬁcant preference for such incremental rendering

techniques for interactive editing of small objects over full-screen re-rendering with denoising.

1 INTRODUCTION

Monte Carlo (MC) path tracing (Kajiya, 1986; Pharr

et al., 2016) is a frequently used physically based ren-

dering technique, as it naturally includes many effects

such as global illumination, reﬂections, soft shadows,

and caustics. Ofﬂine path tracing has become a sta-

ple of the movie industry and production rendering in

general. For each pixel of a rendered image, a path

is traced into the scene, and over time, multiple of

these path samples are accumulated. Each additional

sample helps to reduce MC noise, and sampling qual-

ity is often referred to via the number of samples per

pixel (spp). Due to its stochastic nature, path tracing

is unbiased, but a large number of samples is required

for a noise-free image. Thus, even in the best-case

scenario, renderings require several seconds before an

acceptable level of quality is reached. Due to the high

cost of path tracing, rendering use cases that require a

short feedback loop (e.g., previews during interactive

scene editing) rely on real-time approximation meth-

ods (often based on rasterization). These solutions

closely, but not entirely, mimic the results obtained

https://orcid.org/0000-0002-0539-9378

https://orcid.org/0000-0002-5168-8648

https://orcid.org/0000-0002-9939-0517

https://orcid.org/0000-0002-9370-2663

by an eventual path-traced render. However, to pro-

vide more accurate previews, it is desirable to use path

tracing interactively, also during scene editing. Such

a solution would provide a more seamless experience

and productive workﬂow for digital artists, as current

approximating methods cannot fully capture complex

light transport effects.

Schmidt et al. (Schmidt et al., 2014) identify in-

teractive feedback as an open problem in artistic edit-

ing. Much has happened since then: Modern graph-

ics hardware supports ray-tracing pipelines next to

traditional rasterization. However, in real-time set-

tings (30–60 FPS), current GPU hardware imposes

a soft limit of ∼1 spp per frame. Several powerful

noise-reduction techniques have been introduced, al-

lowing not only shorter rendering times for ofﬂine

rendering, but also path tracing in interactive settings.

However, current solutions generally discard all pre-

viously accumulated samples as soon as any part of

the scene changes (camera movement, object modiﬁ-

cations, material edits). Although this is rightly jus-

tiﬁed by physical correctness (as every change can

affect indirect light transport), it results in low im-

age quality during editing and disruptive artifacts, e.g.

ﬂickering or blurriness.

Global image updates are reasonable when the

scene changes overall, but can be unnecessarily dis-

Ulschmid, A., Kerbl, B., Krösl, K. and Wimmer, M.

Real-Time Editing of Path-Traced Scenes with Prioritized Re-Rendering.

DOI: 10.5220/0012324600003660

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2024) - Volume 1: GRAPP, HUCAPP

and IVAPP, pages 46-57

ISBN: 978-989-758-679-8; ISSN: 2184-4321

Incremental Updates

Continuous Refinement

Global Updates

Any

Edit

Camera

Edit

Scene

Edit

Compute Tile Order

based on Importance Metric

Duplicate

Current Tiles in Budget

Reduction Compute Shader

Aggregate

All Tiles

Processed

Replace Updated Tiles

preserving previous values elsewhere

Follow-Up Edit

Tiles left in Queue

Conventional

Incremental (ours)

(a) Reference

(b) Noisy Re-rendering

(d) (Chaitanya et al., 2017)

(e) Incremental Noisy (ours)

(f) Incremental Denoised (ours) (g) Process Overview

Figure 1: (a) – (f): Comparison of different rendering methods with 1 spp sample budget to the ground truth (a) of an in-

teractively edited scene, after slightly raising the cup. In (b) previously accumulated samples were discarded during the

computation of the new visualization and the image is updated globally, resulting in disruptive quality degradation every-

where. (c) and (d) employ denoising to reconstruct high-ﬁdelity images from (b). (e) and (f) use a tile-based approach with

incremental updates, concentrating available samples on a subset of screen-space tiles and processing them in order of impor-

tance, measured by the Chebyshev distance from the edit on screen. The latter two avoid both noise and overblurring in the

entire scene. (g): Diagram of our suggested re-rendering process containing incremental updates.

ruptive for minor modiﬁcations that inﬂuence only a

localized region: The artist must retain their vision

of a desired visual change while having to observe

an instantaneous regression from converged images to

grainy or blurry visuals during the ﬁrst few moments

of image formation. Manually specifying a static re-

gion of interest for a high-quality preview, as some

production renderers allow, cannot react to dynamic

scene changes with large object movements.

In this paper, we propose to automatically iden-

tify areas of high importance and, following an edit,

re-render these ﬁrst at higher quality. An importance

policy is used to identify image areas where changes

have a lower impact and schedules gradual updates

in an incremental fashion. By rendering in the order

of importance, we eventually update the entire image

to achieve physical correctness but allow an artist

to ﬁrst focus on their modiﬁcation. To determine

importance, we use the Chebyshev distance of pixels

from the modiﬁed object’s position on screen. We

implement two ﬂavors of this approach: One noisy

variant, where each MC-integrated image region is

presented immediately, and a denoised variant with

an additional denoising step. Importantly, we aim to

Real-Time Editing of Path-Traced Scenes with Prioritized Re-Rendering

verify the usability of such incremental re-rendering

approaches in an interactive editing setting. Based

on our suggested methods, a detailed user study is

conducted to evaluate beneﬁts of both the noisy and

denoised methods for real-time scene editing. The

incremental path tracing solutions are compared to a

state-of-the-art denoiser that discards all previously

accumulated samples upon edits (conventional

re-rendering). Through this user study, we aim to

answer the following research questions:

• RQ1. Which re-rendering method do artists pre-

fer (w.r.t. perceived workload and focusability)?

Are preferences dependent on speciﬁc editing sce-

narios (small objects, large objects, lights)?

• RQ2. Which kind of updates (incremental or

global) do artists prefer overall?

• RQ3. What is their general attitude towards path

tracing for scene editing?

By answering these questions, we hope to quan-

tify the beneﬁts of our proposed methods and further

gain a better understanding of the artists’ needs. De-

veloping techniques adapted to their workﬂow could

enable them to use the same rendering method during

scene editing as for the ﬁnal product and enhance their

overall productivity. Our main contributions thus are:

• The design of a tile-based, incremental real-time

path tracing pipeline. Scheduled tiles are re-

peated multiple times in the buffer input of a

Path Tracer. This enables replacing conventional,

global re-rendering with same-time, adaptive re-

rendering, requiring only minimal changes to un-

derlying rendering backends.

• Two implementations of automatic, prioritized re-

rendering policies for editing: one displaying the

raw values, which may still be noisy, and one with

an additional denoising step.

• A detailed user study that investigates the pre-

ferred rendering method (conventional or incre-

mental updates) of experienced artists in different

scenarios (scene modiﬁcations with uniform and

non-uniform impact), as well as offering new in-

sights into the artists’ workﬂow and requirements.

2 RELATED WORK

The foundations of MC path tracing are provided by

Kajiya (Kajiya, 1986). A thorough description of im-

plementing an ofﬂine path tracing framework is given

by Pharr et al. (Pharr et al., 2016). The recently in-

troduced Turing architecture (Nvidia, 2018) enables

hardware-accelerated accelerated ray-tracing and fa-

cilitates real-time path tracing at a low sample count.

2.1 Adaptive Sampling

Zwicker et al. (Zwicker et al., 2015) categorize adap-

tive rendering strategies into a-priori and a-posteriori

methods. A-priori approaches incorporate local 3D

geometry for example to compute gradients or ana-

lyze the light transport equations or analytical BRDF

models to decide where to invest in a higher amount

of samples. A-posteriori base this decision on statis-

tics such as the variance of samples thus far accu-

mulated or the mean squared error (MSE). Seminal

work on adaptive sampling to achieve high quality

with reduced sample count was provided by Mitchell

(Mitchell, 1987). Select examples of later methods

focused on path tracing are presented by Overbeck

et al. (Overbeck et al., 2009) (measuring the error

based on 2D wavelet approximations) or Hachisuka et

al. (Hachisuka et al., 2008) (computing an error based

on samples stored in multidimensional path space).

However, these works and later follow-up techniques

target ofﬂine rendering and do not consider interac-

tive aspects. In particular, to the best of our knowl-

edge, none of these adaptive approaches is designed

in the context of scene editing. In contrast, we exploit

the priors of the editing process (information where

changes occurred) to steer the sample distribution.

Strategies for improving importance sampling as

pursued by techniques like ReSTIR (Wyman and Pan-

teleev, 2021) constitute a research direction comple-

mentary to ours, as both approaches could be com-

bined. The aforementioned paper of Wyman and Pan-

teleev (Wyman and Panteleev, 2021) adapts ReSTIR

for production purposes and amongst other things

combines it with the ReLAX denoising approach.

2.2 Denoising Techniques

Much research has been conducted on improving

the computational time of physically-based light-

transport simulations to archive real-time frame rates.

The main approach is to enhance noisy, but fast

low-sample renderings by applying image reconstruc-

tion techniques as a post-processing step. Schied

et al. (Schied et al., 2017) introduced spatiotempo-

ral variance-guided ﬁltering (SVGF), which strives to

output a temporally coherent image sequence by ac-

cumulating multiple samples. They later reﬁne this

technique to reduce lag and ghosting artifacts by in-

corporating sparsely sampled gradients in the deci-

sion of whether to include a previous sample in a pix-

els history buffer (Schied et al., 2018).

GRAPP 2024 - 19th International Conference on Computer Graphics Theory and Applications

Nvidia provides a library consisting of three

real-time denoisers: ReLAX, ReBLUR, and Sigma

(NVIDIA, 2021). Building up on SVGF, ReLAX,

inspired by temporal anti-aliasing techniques, clamps

to the fast history to improve the treatment of disoc-

cluded areas and deal with ghosting artifacts at the ex-

pense of a noisier history buffer. It can be combined

with the often more performant ReBLUR (Zhdan,

2021), which is used in computer games such as Cyp-

berpunk 2077 for enhanced visuals via ray-tracing.

Both can separately denoise diffuse and specular sig-

nals and handle checkerboarded half-resolution input.

Recently, deep learning-based denoisers have

been introduced such as Intel’s Open Image Denoise

(OIDN) (Intel, 2019) used for example by Unreal En-

gine or Nvidia’s OptiX denoiser based on the research

of Chaitanya et al. (Chaitanya et al., 2017). The lat-

ter adopts a recurrent autoencoder additionally using

auxiliary buffers to include for example depth infor-

mation and normals. Kuznetsov et al. propose a joint

adaptive and reconstruction approach for low sam-

pling rates based on convolutional neural networks

(Kuznetsov et al., 2018). Neural methods, however,

while still achieving close to real-time frame rates on

recent hardware, currently involve a slightly higher

computational cost. They are thus mainly used in

ofﬂine or interactive rendering, but not in computer

games. Thomas et al. recently proposed a combined

neural denoising and supersampling approach with

comparable runtimes to SVGF, but higher quality re-

sults (Thomas et al., 2022). Similarly, (Hasselgren

et al., 2020) combine neural denoising with learned

spatiotemporal sampling strategies.

In a dynamic context, all of the methods de-

scribed above discard all previous samples to compute

the current frame whenever something in the scene

changes and only reuse them via temporal reprojec-

tion during the denoising step. Instead of these global

updates, we propose an adaptive rendering approach

with incremental updates tailored to artists using path

tracing during scene editing.

2.3 Scene Editing and Perception

(Murakami and Hirota, 1992) introduce an incremen-

tal ray-tracing scheme, which tracks changes via a

spatial voxel partition and hash indices. For consis-

tent scene editing, (G

unther and Grosch, 2015) pro-

pose using progressively computed difference images

to identify regions where scene modiﬁcations have

a high impact. Their method is based on stochas-

tic progressive photon mapping, as they focus on

slowly converging effects such as caustics, which are

more negatively affected by a global reset. (Rous-

Figure 2: Tile layout representing traced paths of 1 spp ren-

derings. On the left: Standard approach of sampling the

entire image once (green). On the right: Our approach of re-

peatedly sampling only a few selected tiles (in the depicted

case two in red and blue) during the incremental updates.

The number of repetitions determines sample count within

a tile, which we refer to as the tile quality. Note that the

actual tiles are much smaller than depicted.

selle et al., 2016) reformulate the previous approach

into a control-variate integration scheme and explore

estimator combinations based on covariance in of-

ﬂine editing scenarios with static scenes and gradient-

domain rendering. While they provide a thorough

qualitative analysis of their method, they do not eval-

uate user beneﬁts of such adaptive solutions, e.g., in

the context of interactive editing scenarios. Addition-

ally, the above approaches use uniformly distributed

and not prioritized, adaptive screen-space samples.

In contrast, we outline and thoroughly evaluate re-

rendering methods that automatically identify and pri-

oritize regions of interest. Furthermore, our solutions

are explicitly designed to allow integration into real-

time pipelines for immediate visual feedback.

(Myrodia, 2021) conducted multiple studies to ex-

amine the perception of different noise levels in im-

ages rendered with MC path tracing, discovering that

participants primarily used their most central vision

as well as non-textured and brightest areas of images

to detect noise. Foveated rendering for path tracing

in Virtual Reality (VR) uses eye-tracking to discover

regions of higher interest to the viewer, which then a

higher sample count can be devoted to. The basic con-

cept is similar to our idea and could additionally be

integrated when editing a scene in VR in case artists

focus on a region we identiﬁed as having a lower im-

pact. Similarly to foveated rendering with path trac-

ing, we also exploit the fact that unbiased MC render-

ing allows to straightforwardly average partial results

enabling local image areas with higher quality.

3 PRIORITIZED RE-RENDERING

We propose two variants for the priority-based re-

rendering techniques with incremental updates. To

allow integration in real-time pipelines, our solutions

are designed as modules in Nvidia’s Falcor frame-

Real-Time Editing of Path-Traced Scenes with Prioritized Re-Rendering

work for rapid prototyping of real-time techniques

(Kallweit et al., 2022) (version 5.1.). The code can

be found online (Ulschmid et al., 2023). Falcor

employs the concept of render graphs consisting of

render passes. Our methods adopt the Megakernel

Path Tracer and accumulation pass to change between

three modes of operation: global updates, continuous

reﬁnement, and incremental updates. A global up-

date discards all previously accumulated samples and

renders the new frame with 1 spp uniformly. Dur-

ing continuous reﬁnement, per-pixel samples in each

tile are accumulated by a moving average. Conven-

tional re-rendering operates exclusively using these

two modes, switching to the ﬁrst whenever the scene

is changed or the camera moves. By adding a third

mode, we treat scene modiﬁcations separately by al-

lowing to update the image in an incremental fash-

ion. During the global updates and continuous re-

ﬁnement step, both variants employ Nvidia’s OptiX

(Chaitanya et al., 2017) as a noise-reducing measure.

Our methods have the same runtime complexity as

conventional re-rendering (aside from a small delta

for tile merging and denoising in ours, see below) and

achieve real-time performance (30 frames per second

at 1080p resolution) on state-of-the-art hardware.

3.1 Noisy Incremental Updates

To support incremental updates, we changed the

Megakernel Path Tracer to operate on tiles represent-

ing disjoint parts of the scene in image space instead

of the entire viewport. By repeating a selection of tiles

multiple times in the input buffer of the (unbiased)

path tracer, we can sample this region at a higher qual-

ity without having to adapt the underlying implemen-

tation, including random number generation. Our ap-

proach is designed to operate on a tile budget, thus we

can redirect the original GPU workload (and overall

complexity) of full-screen updates to just a few tiles.

After the path tracing procedure, we use a com-

pute shader to aggregate the content of the dupli-

cated tiles with a tree-based averaging reduction. This

is possible since we target unbiased MC rendering,

which allows for straightforward averaging of partial

results to create higher-quality areas. Our approach

thus yields a few high-quality tiles in roughly the

same time it would take to render a whole scene of

the same size at a low sample rate (not accounting for

variations in scene complexity over the screen). The

buffer layout we use is illustrated by Figure 2. Choos-

ing a higher target tile quality results in fewer tiles be-

ing processed per update and thus a smaller region of

the image that can be re-rendered with a given budget.

The tile size directly inﬂuences the visuals and extent

Figure 3: Left: tiles update simultaneously in resting scene.

Right: a spiral determines the order of tile updates after edit.

Note that the actual tiles are much smaller than depicted.

of each incremental update. As default values, we set

the tile quality to 64 spp and the tile size to 16 pixels.

Falcor’s path tracing pipeline includes a sample

accumulation stage, which aggregates color values

over time as long as scenes remain unchanged and re-

sets them to black otherwise. Consider the case of

a small, localized scene edit. In our incremental ap-

proaches, only the previously processed tiles are re-

placed. The values in the other tiles are kept. This

way, the regions of the image we identiﬁed as most

affected by a modiﬁcation get updated in high quality

as soon as possible, while less inﬂuenced regions are

re-rendered at a later time. Unless the scene is further

modiﬁed, we switch to the default, full-screen accu-

mulation mode after the entire screen has been pro-

cessed once with incremental updates. In the case of

further modiﬁcations, we restart the incremental pro-

cedure at the changed object’s screen position. Dur-

ing the global updates and continuous accumulation

of samples, the tiles are processed from top to bottom

and left to right. If the scene is modiﬁed and an incre-

mental update is triggered, the order of tiles is com-

puted by constructing a spiral starting at the object’s

position projected to image space (see Figure 3).

Our applied importance metric thus deﬁnes the

priority P of a tile t with center (x

, y

) in image space

using the Chebyshev distance from the selected object

o with center (x

, y

) as follows:

P(t) = max(|x

− x

|, |y

− y

|) (1)

3.2 Denoised Incremental Updates

In addition to the above method, we designed a ”de-

noised” version. In addition to denoising during con-

tinuous reﬁnement, it also applies the OptiX denoiser

during incremental updates, resizing its domain to the

space of already re-rendered tiles. However, as re-

sizing and running a denoiser designed for full-image

denoising can be costly, we introduce a conﬁgurable

threshold parameter. A value of zero means the de-

noiser is resized and run each frame. Values above

zero indicate the number of pixels the updated region

GRAPP 2024 - 19th International Conference on Computer Graphics Theory and Applications

has to grow on either side before the denoiser is re-

sized and executed. In order to not mix noisy and

denoised tiles, we separately store two images: One

containing all noisy updated tiles and another contain-

ing only the region that has so far been updated and

denoised. The ﬁrst one is the same as for the noisy in-

cremental method. If the threshold value is reached,

the denoiser is resized to and executed upon the en-

tire region that has been updated so far. For temporal

stability, the smaller, previous denoised region is then

placed on top of the frame again.

Higher threshold values can lead to slightly more

noticeable edges between the old and newly updated

regions and less reactive, block-like visuals. Overall,

however, it increases the performance of the method

such that it only deviates from the noisy incremen-

tal approach by a minor delta that does not affect

perceived responsiveness during editing. We used a

threshold value of 200 as a default.

3.3 Interaction and User Control

To enable the user to modify the scene, we added an

additional interaction pass to Falcor. It allows select-

ing one or multiple objects and exposes their trans-

form and material properties via the Dear ImGUI user

interface (UI). These entries can be manipulated via

clicking and dragging or directly entering a value.

Changes in the parameters are internally applied to

the objects as an animation. The UI stage we added

to the Falcor path tracing pipeline also communicates

information to the rendering stages, such as selected

object center and whether an interaction occurred.

Artifacts can occur when a user interrupts the in-

cremental update procedure by continuously chang-

ing the object position with very high tile quality set-

tings. If the resulting tile size is smaller than the ob-

ject, moving it can result in ghosting artifacts, as parts

of the object in its old position persist. In this case, the

tile quality can be reduced or the tile size increased

to process more tiles in one incremental update. Au-

tomating this parameter tuning is left as future work.

4 EVALUATION

Especially for smaller edits, reusing tiles from pre-

vious, converged images for re-rendering can sig-

niﬁcantly increase quality as measured by metrics,

e.g., PSNR and PSNR-HVS-M, (Ponomarenko et al.,

2007; Ponomarenko et al., 2011). Figure 4 illustrates

this for the edit in Figure 1, also displayed by the

video in the suplementary material. However, our

main focus in this work lies on verifying whether

0 25 50 75 100

Frame

PSNR

0 25 50 75 100

Frame

PSNR−HVS−M

Noisy

Incremental

Denoised

Incremental

Optix

Denoiser

SVGF

Comparison during Scene Editing

Figure 4: Comparison of different global and incremental

methods during the edit in Figure 1 all starting at 1024 spp.

The dotted line represents 1 spp baseline without denoising.

As our method replaces converged tiles before the edit with

new ones, image accuracy degrades gracefully while mov-

ing the cup.

such policies also result in an improved usability and

interactive editing experience. Thus, we conducted

an extensive user study to compare conventional and

priority-based re-rendering.

4.1 User Study Overview

Referring to our three research questions stated in sec-

tion 1, we investigate which rendering method artists

prefer and if their preference depends on a speciﬁc

editing scenario (RQ1). The following three scenar-

ios are considered:

1. Small edits (modify localized geometry)

2. Large edits (modify large scene objects)

3. Modify a scene’s light sources

To evaluate the artists’ preferences, we instructed

them to execute multiple tasks with global and in-

cremental re-rendering methods under the scenarios

listed above (see Table 1). Afterward, we let the par-

ticipants rate the experienced combination of method

and scenario to extract a participant-wise score. We

start from the initial premise that edits on small ob-

jects primarily affect a local region of the scene and

it might thus be obstructive for a ﬂuid workﬂow to

re-render the entire image. On the other hand, light

source edits mostly have a global inﬂuence and are

more likely to require a full image update to convey a

complete impression of their effect. Thus, one goal of

this study is to evaluate the following two hypotheses:

Artists prefer (w.r.t. perceived workload and focus-

ability, measured as a participant-wise score)

1. an incremental rendering method to state-of-the-

art solutions (i.e., denoised full-screen updates)

when editing small objects.

2. a global update approach when manipulating light

sources.

Real-Time Editing of Path-Traced Scenes with Prioritized Re-Rendering

We test these hypotheses via ANOVA and pair-

wise t-tests. By performing an a-priori power analysis

using G*Power (Faul et al., 2007), we calculated the

number of participants needed for a power of 1 − β =

0.8 with a medium effect of η

= 0.06 as 15 for an

ANOVA (repeated measures, within factors) and as

34 for a two-tailed, paired t-test with a medium effect

of d = 0.5.

The preference of users when editing larger ob-

jects as well as the remaining aspects are researched

in an exploratory fashion. We therefore analyse cen-

tral tendencies and effect sizes to identify trends in the

pairwise differences between the scores. To identify

which kind of updates are preferred overall (RQ2),

we compute a favored method based on the afore-

mentioned score via majority vote. Furthermore, we

directly asked the participants to select their favorite

method, as well as to rate them based on whether they

would buy them, to construct a metric similar to a Net

Promoter Score (NPS). To answer what is the general

attitude towards path tracing for scene editing (RQ3),

we asked questions about the frequency of working

with path tracing and reasons for not using it during

scene editing. Moreover, we conducted structured in-

terviews with the artists at the end of the study, asking

why they preferred a method, how they estimate the

future relevance of physically-based rendering in their

ﬁeld of work, and what they would consider helpful

during scene editing with real-time path tracing.

4.2 Technical Setup and Participation

The survey was implemented as a self-hosted Dru-

pal Webform and the full questionnaire, as well as

the anonymized collected data and R code used for

the evaluation, can be found online (Ulschmid et al.,

2023). We use three different scenes from the Bitterli

Rendering Resources (Bitterli, 2016): The Contem-

porary Bathroom, the Country Kitchen, and The Grey

& White Room. We refer to them simply as Bathroom,

Kitchen, and Living Room.

Since hardware-accelerated path tracing requires

an Nvidia RTX GPU or comparable, we offered the

artists three ways to participate: They could attend in

person to perform the study on one of our PCs with an

Nvidia RTX 3070 Ti, or online by either executing the

program on their own hardware or streaming from our

PC by remote controlling it via Parsec (Parsec Cloud

Inc., ). Under all three circumstances, we recorded

screen and audio for the duration of the study, to make

sure the experiment conditions were comparable for

all participants, as well as for further evaluation. In

the online cases, we supervised the experiment via in-

ternet telephony while participants working on their

Table 1: Number of tasks per edit type and scene. Transfor-

mational edits include moving and rotating objects. Mate-

rial edits involve changing albedo and roughness values of

objects or emissive parameters of lights/environment maps.

Scene Edit

Small

Objects

Large

Objects

Lights

Kitchen

Transform 1 1 -

Material 1 1 2

Living

Room

Transform 1 1 -

Material 1 - 1

Sum 4 3 3

own hardware instead of streaming from ours were

sharing their screens.

The artists were asked to voice their thoughts dur-

ing the study, and we conducted a structured inter-

view directly after the experiment. For evaluation

purposes, the voice recordings were transcribed us-

ing OpenAIs speech-to-text system Whisper (Radford

et al., 2022).

Throughout the entire questionnaire as well as in

the Falcor framework we obfuscated the names of the

rendering methods using letters. We also hid all UI el-

ements except the needed transformational, material,

and rendering parameters.

4.3 User Study Design

For the representative of conventional re-rendering

with denoising, we use the AI-accelerated Nvidia Op-

tiX denoiser (Chaitanya et al., 2017), as it produced

higher-quality results than SVGF in our quantitative

analysis, while still being interactive.

Before beginning the experiment, participants

were asked to sign a consent form. The study itself

consists of three parts: an introductory tutorial, the

task-based evaluation of the three methods (denoiser,

noisy incremental, denoised incremental) in three dif-

ferent editing scenarios (small objects, large objects,

lights), as well as some general and demographic

questions at the end. We intentionally searched for

artists from diverse backgrounds and different experi-

ence levels to achieve better generalizability.

Tutorial. The tutorial explained these necessary el-

ements of the UI to the participants (see section 3) and

allowed them to practice by executing four exemplary

tasks in the Bathroom scene using the basic 1 spp ren-

dering without any additional features enabled. The

artists were also instructed to use predeﬁned cam-

era viewports but were allowed to move the camera

if they really needed it to not interrupt their artistic

workﬂow. However, we ensured that they would al-

ways stay at the same distance from the object. The

GRAPP 2024 - 19th International Conference on Computer Graphics Theory and Applications

Table 2: Format of the Likert items for the adapted scale

used for the extended NASA-TLX questions and their nu-

merical mapping applied during the evaluation.

Worse

than

1spp

Same

1spp

Slightly

better

than 1spp

Better,

but still

ﬂawed

Mostly

like con-

verged

Almost

equivalent

Same as

converged

-1

1 2

tutorial also provided a short text explaining the cur-

rent limitations of real-time path tracing and showed

two prerecorded videos of a cup moving in the Bath-

room scene, one with a 1 spp noisy re-rendering and

one with 4096 spp plus denoising with OptiX. The lat-

ter was supposed to serve as a reference for an imagi-

nary approach that instantly converges.

Task-based Evaluation. During the task-based

evaluation, participants performed multiple editing

tasks in both the Kitchen and Living Room scene (see

Table 1) for each of the three editing scenarios and the

three rendering methods, yielding nine combinations

in total. After each condition, the artists were asked

to rate the experienced method. For the incremental

approaches, we additionally asked which rendering

parameters (tile quality/size, threshold) they changed

and which settings they preferred (see subsection 3.1).

The order of the tasks and scenarios was ﬁxed,

whereas we randomized the order of the rendering

methods so participants would either ﬁrst encounter

the conventional or the incremental techniques (in the

order noisy then denoised).

To rate the experienced combination of scenario

and method, we used a questionnaire based on the

standardized NASA Task Load Index (NASA-TLX)

(Hart and Staveland, 1988), which is comprised of six

questions measuring the perceived workload of tasks.

We added two questions to assess focus and distrac-

tion and dropped the question on physical demand as

we deemed it less relevant for our purposes, resulting

in seven questions in total.

Scale Design. After conducting a small pilot study

with four participants, we also decided to change the

scale of the answers from the original range with

21 gradations to Likert items with seven levels as

can be seen in Table 2 following the construction

guidelines in (South et al., 2022). The pilot partic-

ipants were confused about the comparison point to

use for the ﬁrst encountered method. In response to

this, we introduced the tutorial, where they experi-

ence the conventional 1 spp noisy re-rendering ap-

proach themselves and watch a video of how one of

the editing tasks would look like if we could produce

high-quality renderings in real-time. We requested

the artists to rate the experienced rendering methods

in comparison to the basic 1 spp approach and the

imaginary converged solution. In using these descrip-

tions for our adapted seven-point scale, the partici-

pants were provided with the reference points they re-

ported as missing during the pilot study.

4.4 Results

We use an alpha level of α = 0.05 for all statistical

tests and, in case of multiple testing, report adjusted

p-values for better readability. As effect sizes we re-

port, depending on applicability, Cohen’s d (< 0.5

small, [0.5, 0.8] medium, > 0.8 large), Cram

er’s V

(0.1 small, 0.3 medium, 0.5 large) and the general-

ized eta squared η

(Cohen, 1988). As we did not ﬁnd

any similar studies on our topic, we use the standard

interpretation of η

for η

(< 0.06 small, [0.06, 0.14]

medium, > 0.14 large).

For the task-based evaluation, following the orig-

inal formulation of Likert-typed scales (South et al.,

2022), we aggregated the extended NASA-TLX ques-

tions by averaging the individual Likert items. The

aggregation is computed for each combination of sce-

nario and method on participant level, resulting in

nine scores for each participant. This procedure is

also referred to as Raw TLX (Hart, 2006). As the

same participant ranked all of the combinations dur-

ing the study and the individual ratings are thus de-

pendent on the subject, we use paired or repeated

measure methods.

Study Population. Out of the 21 artists (m=15, f=4,

d=2) who completed the survey, ten were profes-

sionals from the industry, and eleven were students

from universities. Their experience measured in hours

spent using 3D rendering software ranged from ’50-

99’ to ’10,000+’ hours (Median Mdn: ’1,000-4,999’,

interquartile range IQR: [’100-499’, ’5,000-9,999’])

and participants’ ages ranged from 21 to 51 (Mean

M: 30.48, standard deviation SD: 7.79).

Eleven executed the tasks with the denoiser ﬁrst

and ten always the incremental methods ﬁrst. Eight

artists participated via Parsec or on their own hard-

ware, while ﬁve performed the tasks in person at our

ofﬁce. The results of further questions about the study

population are reported in Table 3. Most artists com-

pleted the study in roughly one hour.

Assumption Testing. The applied statistical tech-

niques require the data to fulﬁll the following assump-

tions: no signiﬁcant outliers, normality, and spheric-

ity. The sphericity assumption requires that the vari-

ances of the pairwise differences are equal. To test

Real-Time Editing of Path-Traced Scenes with Prioritized Re-Rendering

Editing Scenario

Small Objects Large Objects Lights

Favor Optix Denoiser

Favor Incremental

−4

−2

Small Objects Large Objects Lights

Editing Scenario

Difference in the aggregated Score

Favorite Incremental − Optix Denoiser

Favor Optix Denoiser

Favor Noisy Incremental

−4

−2

Small Objects Large Objects Lights

Editing Scenario

Difference in the aggregated Score

Noisy Incremental − Optix Denoiser

Favor Optix Denoiser

Favor Denoised Incremental

−4

−2

Small Objects Large Objects Lights

Editing Scenario

Difference in the aggregated Score

Denoised Incremental − Optix Denoiser

Favor Denoised Incremental

Favor Noisy Incremetal

−4

−2

Small Objects Large Objects Lights

Editing Scenario

Difference in the aggregated Score

Noisy Incremetal − Denoised Incremental

p=0.372 p=0.364 p=0.101 p=0.057 p=0.723 p=0.096 p=0.44 p=0.177 p=0.089 p=2.4e−04 p=0.081 p=0.051

Pairwise Differences between the Rendering Methods by Editing Scenario

Figure 5: Top row: Box plots of the pairwise differences in the averaged score of the task-based evaluation. If the difference

between two methods is positive, the method named ﬁrst was favored. If it is negative, the method named second was favored.

The respective means are depicted with a diamond shape and outliers as stars, whereas extreme outliers are colored red.

Bottom row: quantile-quantile (QQ) Plots of the pairwise differences by editing scenario and results of Shapiro-Wilk tests. If

the p-value is larger than the assumed signiﬁcance level of α = 0.05, we may assume a normal distribution.

Table 3: Artistic background of the study population. For

all questions, except the path tracer usage, selecting multi-

ple answers was possible.

Results of the general questions

Field of Work Used Denoisers Path Tracer Usage

Games 14 OptiX 12 Never 4

Digital Art 11 Intel 5 Occasionally 9

Animation 9 None 2 Regularly 7

Architecture Vis. 6 Vdenoise 2 Every time 1

Other 9 Other 5

Used Software Reasons for not using a Path Tracer

Blender 18 Waiting time for clear image too long 16

Unity 3D 15 Real-time approx. good enough 10

Unreal Engine 10 Used software has no Path Tracer 7

Maya 9 Only using for ﬁnal image 6

3ds Max 7 Initial noise/blurryness is distracting 5

ZBrush 6 Necessary hardware not available 2

Other 14

this we use Mauchly’s Test, which is signiﬁcant (p =

0.041, ε = 0.743), meaning the assumption is violated.

There are no outliers on the aggregated values.

However, when looking at the pairwise differences

between the methods by participant and scenario de-

picted in Figure 5, we found two signiﬁcant outliers:

one on the difference between noisy incremental and

denoised incremental for small objects and one on the

difference between the denoiser and denoised incre-

mental for lights. As removing the respective values

led to similar overall results, we left the outliers in the

data for the main analysis, but reported both results

where appropriate.

To test if the data is normally distributed we use

a Shapiro-Wilk test on the pairwise differences. The

resulting p-values, as well as QQ plots, can be seen

in Figure 5. The only pairwise difference violating

the normality assumption is between noisy incremen-

tal and denoised incremental for small objects. By

looking at the QQ plot we identify the same outlier as

above as being responsible. Removing it would lead

to a value of p = 0.527, however with the same rea-

soning as before we decided to leave it in the data.

ANOVA and t-Test Results. To analyze the in-

teraction between the rendering method and edit-

ing scenario on the score aggregated by participants,

we performed a two-way repeated measures ANOVA

with the scenario as the focal variable and the used

method as the moderator variable. As the spheric-

ity assumption is violated, we apply the conservative

Greenhouse–Geisser correction. We found a statisti-

cally signiﬁcant interaction between the rendering ap-

proach and editing scenario on the aggregated value

(F(2.97, 59.42) = 5.29, p = 0.003, η

= 0.03).

Therefore, we analyzed the effect of the used

method for each scenario, reporting p-values ad-

justed using the Bonferroni multiple testing correc-

tion method. Using one-way ANOVAs, the simple

main effect of the rendering approach was found to

be only signiﬁcant for small objects (p = 0.03, η

0.06). Pairwise comparisons using paired t-tests show

that for small objects, the mean aggregated value was

signiﬁcantly different between using the denoiser and

the denoised incremental method (p = 0.026, d =

0.63) and between the noisy and denoised incremental

GRAPP 2024 - 19th International Conference on Computer Graphics Theory and Applications

Small Objects Large Objects Lights

Editing Scenario

Aggregated Score

Method

Optix

Denoiser

Noisy

Incremental

Denoised

Incremental

Anova, F(2.97,59.42) = 5.29, p = 0.003, η

= 0.03

Averaged Task Question Score by Rendering Method and Editing Scenario

pwc: T test; p.adjust: Bonferroni

(a)

Incremental

Update

Global

Update

Preference

Number of Participants

Method

Optix

Denoiser

Noisy

Incremental

Denoised

Incremental

Selected in the General Section

Overall prefered Method

(b)

Figure 6: (a) Box plot of the aggregated score of the task-based evaluation part of the survey (♦: mean, *: α < 0.05). (b) Bar

plot showing the favorite rendering approach selected in the general section at the end of the questionnaire.

method (p = 0.05, d = 0.57; with outlier removed p =

0.025, d = 0.66). There is no signiﬁcant difference for

light edits between using the denoiser and the noisy (p

= 1, d = 0.15) or denoised (p = 0.69, d = 0.27; with

outlier removed p = 0.079, d = 0.54) incremental ren-

dering technique (see Figure 6a).

Analysis of Differences. When looking at the pair-

wise differences between the rendering methods at

per-participant level for each scenario depicted in Fig-

ure 5, we observed a trend in central tendency be-

tween the OptiX denoiser and the denoised incremen-

tal method that the participants of our study seemed to

favor the denoised incremental method for small ob-

jects (M: 0.948, SD: 1.50) and the denoiser for light

edits (M: -0.464, SD: 1.72). By looking at the dif-

ferences between the noisy and denoised incremental

approach, however, it can be noticed that while for

small objects the denoised method was preferred (M:

-0.564, SD: 0.988), for light edits the noisy method

was favored slightly more (M: 0.25, SD: 0.958).

To analyze in a more general fashion whether a

global or incremental update method is preferred, we

extracted the favorite incremental method per partici-

pant and scenario by choosing the one with the high-

est aggregated score. We then compared this favorite

to the denoiser. The results for edits on small objects

approximately remain the same, whereas for light ed-

its the differences become closer to zero (M: 0.014,

SD: 1.53). When editing large objects, the differences

between the rendering approaches are roughly zero-

centered for all combinations.

Overall Preference. We furthermore computed

which rendering method artists preferred overall ac-

cording to the task-based evaluation. For this, we ﬁrst

extracted the favored method for each of the three

editing scenarios by taking the one with the highest

aggregated score. Then the overall favorite is deter-

mined by a majority vote. The results can be seen

in Figure 7a. Most participants preferred incremen-

tal over global updates, which is conﬁrmed by di-

rectly asking the artists at the end of the study as

well, as can be seen in Figure 6b. There is, how-

ever, no statistically signiﬁcant relation between the

two favorites (Fisher’s exact test (FET), p = 0.46, V

= 0.37). We identiﬁed a signiﬁcant association be-

tween the favorite for the light editing scenario and

the overall preferred method stated at the end of the

survey (FET, p = 0.043, V = 0.47). The majority vote

favorite is also signiﬁcantly associated with the pref-

erence for edits on large objects (FET, p = 0.004, V

= 0.65) and lights (FET, p = 0.004, V = 0.68). These

values are illustrated in Figure 7c.

Net Promoter Score. To compute a metric similar

to a NPS based on the question if artists would buy

the respective rendering approach as a feature in a 3D

software, we subtracted the percentage of participants

rating higher than seven on a scale from zero to ten

from those rating lower than six. We decided to use a

slightly lower threshold than for a normal NPS, as the

overall ratings were also rather low. The computed

scores are depicted in Figure 7b.

Parameter Settings. Regarding the rendering pa-

rameters, participants favored higher tile qualities for

editing smaller objects (noisy Mdn: 64, denoised

Mdn: 128) and lower quality for larger objects and

light edits (Mdn: 16). The median setting for tile size

was 16, except for the denoised incremental method

in the small objects scenario, where it was 24. For

Real-Time Editing of Path-Traced Scenes with Prioritized Re-Rendering

Incremental

Update

Global

Update

Undecided

Preference

Number of Participants

Method

Optix

Denoiser

Noisy

Incremental

Denoised

Incremental

Undecided

Computed via Majority Vote

Overall prefered Method

(a)

nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1nps = −38.1 nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9nps = −61.9 nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81nps = −23.81

Optix

Denoiser

Noisy

Incremental

Denoised

Incremental

Method

NPS Score

Method

Optix

Denoiser

Noisy

Incremental

Denoised

Incremental

Raw Data for Net Promoter Score

(b)

0.08

0.32

0.53

0.33

0.21

0.65

0.38

0.68

0.47 0.37

* *

small obj.

large obj.

lights

majority vote

large obj.

lights

majority vote

stated

Association (Cramer's V) between Favorites

(c)

Figure 7: (a) Overall preferred rendering approach of the artists extracted by a majority vote between the three methods

with the highest average score for each editing scenario. (b) Box plot showing the distribution of ratings when asking artists

whether they would buy the respective rendering method as a feature in a 3D software. Based on these ratings we computed

a metric similar to a NPS, which is stated on top of the box plots. (c) Association computed via Cramer’s V between the

favorite method for a certain task derived from the averaged scores (Figure 6a) and the overall favorite either calculated via a

majority vote or stated by the artists themselves in the survey.

the threshold parameter, larger values were preferred

(Mdn: ’101-200’).

5 DISCUSSION & CONCLUSION

Regarding our ﬁrst research question RQ1, we found

a statistically signiﬁcant increase in a score mea-

suring participant-wise preference for using the de-

noised incremental approach when editing small ob-

jects by performing an ANOVA and pairwise t-tests.

For larger objects, no clear favorite method could be

determined. In the case of editing the lighting of a

scene, there is a minor, however not statistically sig-

niﬁcant trend in central tendencies to favor the tech-

nique with global updates. The study is slightly un-

derpowered w.r.t. the a-priory power analysis for pair-

wise t-tests. However, we observed larger effect sizes

than assumed.

Overall and in answer to our second research

question RQ2, rendering with incremental updates

seems to be favored according to a preference both

computed via majority vote between the three inves-

tigated scenarios and extracted by directly asking the

participants at the end, as well as to a NPS-like metric.

On the other hand, artists mentioned during a struc-

tured interview that none of the methods perfectly

caters to their needs so far and they would thus still

mainly edit scenes using less ideal real-time approx-

imations, resolving RQ3. However, the future rele-

vance of accurate, physically based methods for ﬁnal

render results is still high for the movie industry and

even inﬁltrates high-performance businesses such as

computer games.

In summary, we presented two adaptive, priority-

based re-rendering approaches designed for interac-

tively editing a 3D scene with MC path tracing. Us-

ing a tile-based approach, we can efﬁciently update

image regions with higher quality in an incremental

fashion, applying an importance metric based on the

Chebyshev distance. Our quantitative and qualitative

evaluation shows that for localized edits, in particular,

our incremental approach reduces noise levels and is

preferred by artists. While we tried to choose a di-

verse study population with different backgrounds, as

well as various editing scenarios in multiple scenes,

our results may not generalize. Future work could

extend the building blocks of the incremental meth-

ods by combining them with higher quality denoisers,

such as Intel’s OIDN (Intel, 2019). Furthermore, the

priority policy could be reﬁned, based on fast approx-

imations of global illumination, which would allow

for better incremental updates for indirect (lighting)

effects. Gradient-based techniques and difference im-

ages could be integrated to enable automatic param-

eter selection (Rousselle et al., 2016). Overall, our

evaluation provides a strong motivation to pursue fur-

ther methods for automatic, priority-based, and inter-

active re-rendering techniques for creative processes.

ACKNOWLEDGEMENTS

This research was funded by FWF project F77 (SFB

Advanced Computational Design SP4) and by WWTF

project ICT22-055 (Instant Visualization and Inter-

action for Large Point Clouds). We also thank our

students Pascal Hann and Nina Semmelrath for their

contributions.

GRAPP 2024 - 19th International Conference on Computer Graphics Theory and Applications

REFERENCES

Bitterli, B. (2016). Rendering resources. https://

benedikt-bitterli.me/resources/.

Chaitanya, C. R. A., Kaplanyan, A. S., Schied, C., Salvi,

M., Lefohn, A., Nowrouzezahrai, D., and Aila, T.

(2017). Interactive reconstruction of monte carlo im-

age sequences using a recurrent denoising autoen-

coder. ACM Trans. Graph., 36(4).

Cohen, J. (1988). Statistical Power Analysis for the Behav-

ioral Sciences. Lawrence Erlbaum Associates, 2nd

edition.

Faul, F., Erdfelder, E., Lang, A.-G., and Buchner, A. (2007).

G*power 3: A ﬂexible statistical power analysis pro-

gram for the social, behavior, and biomedical sci-

ences. Behavior Research Methods Instruments &

Computers, 39:175–191.

unther, T. and Grosch, T. (2015). Consistent Scene Editing

by Progressive Difference Images. Computer Graph-

ics Forum.

Hachisuka, T., Jarosz, W., Weistroffer, R. P., Dale, K.,

Humphreys, G., Zwicker, M., and Jensen, H. W.

(2008). Multidimensional adaptive sampling and re-

construction for ray tracing. ACM Trans. Graph.,

27(3):1–10.

Hart, S. (2006). Nasa-task load index (nasa-tlx); 20 years

later. volume 50.

Hart, S. G. and Staveland, L. E. (1988). Development of

nasa-tlx (task load index): Results of empirical and

theoretical research. In Human Mental Workload, vol-

ume 52 of Advances in Psychology, pages 139–183.

Hasselgren, J., Munkberg, J., Salvi, M., Patney, A., and

Lefohn, A. (2020). Neural temporal adaptive sam-

pling and denoising. Computer Graphics Forum,

39:147–155.

Intel (2019). Open image denoise. https://www.

openimagedenoise.org/.

Kajiya, J. T. (1986). The rendering equation. SIGGRAPH

Comput. Graph., 20(4):143–150.

Kallweit, S., Clarberg, P., Kolb, C., Davidovi

c, T., Yao,

K.-H., Foley, T., He, Y., Wu, L., Chen, L., Akenine-

oller, T., Wyman, C., Crassin, C., and Benty, N.

(2022). The Falcor rendering framework. https:

//github.com/NVIDIAGameWorks/Falcor.

Kuznetsov, A., Kalantari, N. K., and Ramamoorthi, R.

(2018). Deep adaptive sampling for low sample count

rendering. Computer Graphics Forum, 37.

Mitchell, D. P. (1987). Generating antialiased images at

low sampling densities. SIGGRAPH Comput. Graph.,

21(4):65–72.

Murakami, K. and Hirota, K. (1992). Incremental Ray Trac-

ing, pages 17–32. Springer Berlin Heidelberg.

Myrodia, V. (2021). Psychophysical studies on Monte Carlo

rendering-noise visual perception. PhD thesis.

Nvidia (2018). NVIDIA TURING GPU ARCHITEC-

TURE. Technical report.

NVIDIA (2021). Nvidia real-time denoisers. https://

developer.nvidia.com/rtx/ray-tracing/rt-denoisers.

Overbeck, R. S., Donner, C., and Ramamoorthi, R. (2009).

Adaptive wavelet rendering. ACM Trans. Graph.,

28(5):1–12.

Parsec Cloud Inc. Parsec. https://parsec.app/.

Pharr, M., Jakob, W., and Humphreys, G. (2016). Physi-

cally Based Rendering: From Theory to Implementa-

tion. Morgan Kaufmann Pub. Inc., 3rd edition.

Ponomarenko, N., Ieremeiev, O., Lukin, V., Egiazarian, K.,

and Carli, M. (2011). Modiﬁed image visual quality

metrics for contrast change and mean shift account-

ing. In The Experience of Designing and Application

of CAD Systems in Microelectronics, pages 305–311.

Ponomarenko, N., Silvestri, F., Egiazarian, K., Carli, M.,

Astola, J., and Lukin, V. (2007). On between-

coefﬁcient contrast masking of dct basis functions.

Proc. of the 3rd Int. Workshop on Video Processing

and Quality Metrics for Consumer Electronics.

Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey,

C., and Sutskever, I. (2022). Robust speech recogni-

tion via large-scale weak supervision.

Rousselle, F., Jarosz, W., and Nov

ak, J. (2016). Image-

space control variates for rendering. ACM Trans.

Graph., 35(6).

Schied, C., Kaplanyan, A., Wyman, C., Patney, A., Chai-

tanya, C. R. A., Burgess, J., Liu, S., Dachsbacher,

C., Lefohn, A., and Salvi, M. (2017). Spatiotempo-

ral variance-guided ﬁltering: Real-time reconstruction

for path-traced global illumination. In Proc. of High

Performance Graphics. ACM.

Schied, C., Peters, C., and Dachsbacher, C. (2018). Gra-

dient estimation for real-time adaptive temporal ﬁlter-

ing. Proc. ACM Comput. Graph. Interact. Tech., 1(2).

Schmidt, T.-W., Pellacini, F., Nowrouzezahrai, D., Jarosz,

W., and Dachsbacher, C. (2014). State of the art in

artistic editing of appearance, lighting, and material.

Computer Graphics Forum, 35.

South, L., Saffo, D., Vitek, O., Dunne, C., and Borkin,

M. A. (2022). Effective Use of Likert Scales in Vi-

sualization Evaluations: A Systematic Review. CGF.

Thomas, M. M., Liktor, G., Peters, C., Kim, S.,

Vaidyanathan, K., and Forbes, A. G. (2022). Tempo-

rally stable real-time joint neural denoising and super-

sampling. Proc. ACM Comput. Graph. Interact. Tech.,

5(3).

Ulschmid, A., Kerbl, B., Kr

osl, K., and Wim-

mer, M. (2023). https://github.com/cg-tuwien/

Prioritized-ReRendering.

Wyman, C. and Panteleev, A. (2021). Rearchitecting spa-

tiotemporal resampling for production. In High-

Performance Graphics - Symposium Papers. The Eu-

rographics Association.

Zhdan, D. (2021). Reblur: A hierarchical recurrent de-

noiser. In Ray Tracing Gems II, pages 823–844.

Zwicker, M., Jarosz, W., Lehtinen, J., Moon, B., Ra-

mamoorthi, R., Rousselle, F., Sen, P., Soler, C., and

Yoon, S.-E. (2015). Recent advances in adaptive sam-

pling and reconstruction for monte carlo rendering.

Comput. Graph. Forum, 34(2):667–681.

Real-Time Editing of Path-Traced Scenes with Prioritized Re-Rendering