Semi-Automatic Generation of Rotoscope Animation
Using SAM and k-means Clustering
Mizuki Sakakibara
1
and Tomokazu Ishikawa
1,2 a
1
Toyo University, 1-7-11 Akabanedai Kita-ku, Tokyo, Japan
2
Prometech CG Research, 3-34-3 Hongo Bunkyo-ku, Tokyo, Japan
Keywords:
Rotoscope, Animation Techniques, Segmentation.
Abstract:
This paper proposes a novel method for automating the rotoscoping process in anime production by combining
SAM (Segment Anything Model) and k-means clustering. Traditional rotoscoping, which involves manually
tracing live-action footage, is time-consuming and labor-intensive. Our method automatically generates line
drawings and coloring regions suitable for anime production workflows through three main steps: line drawing
creation using SAM2, shadow region generation using k-means clustering, and finishing with color design.
Experimental results from 134 participants showed that our method achieved significantly higher ratings in
both “rotoscope-likeness” and “anime-likeness” compared to existing methods, particularly in depicting com-
plex human movements and details. The method also enables hierarchical editing of animation materials and
efficient color application across multiple frames, making it more suitable for commercial anime production
pipelines than existing style transfer approaches. While the current implementation has limitations regarding
segmentation accuracy and line drawing detail, it represents a significant step toward automating and stream-
lining the anime production process.
1 INTRODUCTION
The anime production industry has been expanding its
market scale in recent years and showing significant
vitality. As viewing methods shift from television
to internet streaming, Japanese anime maintains high
popularity all over the world, increasing its impor-
tance as a content industry. Since the online distribu-
tion of animation has become mainstream, the quality
of the work contributes to the number of views and
is directly related to its reputation, so productions are
spending more money on the drawing process so that
we can see well-drawn animation every week. While
approaches to improving animation quality vary de-
pending on the work and direction, rotoscoping exists
as one such technique.
Rotoscope was developed by Max Fleischer in
1915 as an aid to overcome awkward movement in
animation caused by animators’ insufficient skills, in-
volving tracing live-action footage. It continues to
be used even in the past 20 years as animators’ skills
have matured. Cultural factors include the prolifera-
tion of digital devices with cameras like smartphones
a
https://orcid.org/0000-0002-9176-1336
and easy reference material access via the internet.
The universalization of the production technique of
referencing live-action footage has had a major im-
pact on the use of rotoscopes. Rotoscoping adoption
is divided between animation assistance purposes and
expressive technique purposes. The former is used to
animate sophisticated real movements like walking,
dancing, or musical instrument performance, while
the latter is used when need to express the grotesque-
ness that emerges as a side effect of capturing re-
ality without any stylization. It is also sometimes
adopted as pre-visualization to visually express the fi-
nal form in early production stages by determining
layout and acting before animation, helping directors
achieve their intended screen vision.
A problem when using rotoscoping is that man-
ually tracing live-action footage is extremely time-
consuming and labor-intensive. Since what needs
to be drawn is predetermined, creativity is limited
mostly to deciding what lines to take from the im-
age subjects. For animators, this becomes repetitive
manual labor, leading to decreased motivation. There-
fore, this research aims to contribute to reducing an-
imators’ burden by improving animation production
efficiency through automating the rotoscoping work-
Sakakibara, M. and Ishikawa, T.
Semi-Automatic Generation of Rotoscope Animation Using SAM and k-means Clustering.
DOI: 10.5220/0013321000003912
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 1: GRAPP, HUCAPP
and IVAPP, pages 363-370
ISBN: 978-989-758-728-3; ISSN: 2184-4321
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
363

flow. Specifically, we propose a method to automat-
ically generate rotoscope animation by creating line
drawings from live-action footage and separating ma-
terials while identifying line art expressions and col-
oring regions to facilitate incorporation into produc-
tion. In the proposed system, rotoscope animation is
generated through the following steps:
1. Line drawing creation using SAM
2. Creating shadow regions by reducing colors in ba-
sic coloring areas
3. Finishing (coloring) and compositing
The reason for identifying coloring regions is that
generating only line drawings would require coloring
work for each frame. By identifying coloring regions
beforehand, colors can be applied to the entire anima-
tion at once. In anime production, cels are colored one
by one in the finishing process while referring to color
design, which specifies coloring instructions. In this
research, if color design is available, the finishing pro-
cess can be completed immediately. Afterwards, in
the shooting process that composites multiple materi-
als to create the final image or video, cels and back-
grounds are composited and the screen is adjusted.
It is important to emphasize here that the proposed
technology is not a mere style conversion method like
Diffutoon (Duan et al., 2024) or DomoAI (DOMOAI
PTE. LTD, 2024), but can also output intermediate
data such as line drawings in accordance with the ani-
mation production process. The existence of interme-
diate data allows retakes and re-editing, thus replacing
part of the existing animation production pipeline.
2 RELATED WORKS
We propose a novel automated rotoscoping method
that automatically generates line drawings and col-
oring regions suitable for anime production. In this
section, we classify related existing research from the
following three perspectives and analyze their advan-
tages and disadvantages. Then, we clarify the posi-
tioning and novelty of this research.
2.1 Conventional Rotoscoping Methods
Conventional rotoscoping has primarily been per-
formed through manual line tracing. While line
drawing extraction using edge detection like Canny
method (Canny, 1986) has been studied, it faces chal-
lenges in generating closed regions necessary for col-
oring and cannot reproduce anime-specific line art
expressions. Agarwala et al. proposed efficiency
improvements through keyframe interpolation (Agar-
wala et al., 2004), but the manual workload re-
mains substantial. Adobe After Effects’ Roto Brush
tool specializes in silhouette extraction (Dissanayake
et al., 2021; Torrejon et al., 2020) but is not suited for
hierarchical generation of line drawings and coloring
regions needed in anime production.
2.2 Image Anime-Stylization Using
Deep Learning
GAN-based methods like CartoonGAN (Chen et al.,
2018) and AnimeGAN (Chen et al., 2020), and Sta-
ble Diffusion-based methods (Rombach et al., 2022;
Esser et al., 2024) can generate high-quality anime-
style images. Nevertheless, these methods are not
suitable for animation production as they cannot
consider temporal coherence and shape consistency.
Among these methods, the latest stylization tech-
niques are Diffutoon (Duan et al., 2024) and Do-
moAI (DOMOAI PTE. LTD, 2024). These maintain
general temporal consistency and demonstrate high
quality as video generation AI. However, they can-
not separately output line drawings, coloring regions,
and shooting process effects, making integration into
commercial anime production workflows difficult.
2.3 Segmentation Technology and Its
Application to Anime
As emphasized by animator Tatsuyuki Tanaka, anime
expression consists of “symbolic expressions of sim-
ple lines and color separation” (Tanaka, 2021), unlike
realistic paintings. To capture these symbolic expres-
sions, segmentation at the semantic level becomes
crucial. In recent years, deep learning-based seg-
mentation technology has rapidly advanced, enabling
high-precision segmentation. The Segment Anything
Model (SAM) (Kirillov et al., 2023) is a prime ex-
ample, being a versatile model capable of accurately
segmenting various objects at the pixel level.
In Tous’s research (Tous, 2024), they use SAM
to segment various visual features of characters (hair,
skin, clothes, etc.) and combine it with a method
called DeAOT to automatically generate retro-style
rotoscope animations. However, it specializes in
styles composed of limited colors and expressions
like retro games, and since it does not consider gen-
eral anime production or line drawing generation, it
is not suitable for delicate expressions like Japanese
anime created in a line-expression culture.
As shown in Table 1, existing methods do not ad-
equately consider the hierarchical generation of line
drawings and coloring regions necessary for anime
GRAPP 2025 - 20th International Conference on Computer Graphics Theory and Applications
364

Table 1: Comparison with previous research.
Method Input Output Technique
Symbolic
Expression
Motion
Expression
Hierarchical
Editing
Agarwala et al. Keyframes Animation Interpolation High High Not Possible
CartoonGAN Image Anime-style image GAN High Low Not Possible
AnimeGAN Image Anime-style image GAN High Low Not Possible
Stable Diffusion Text + image Image Diffusion Model High Low Not Possible
Dissanayake et al. Live-action video Silhouette Deep Learning Low Low Not Possible
Tous et al. Live-action video Retro anime SAM, DeAOT Medium High Possible
Proposed Method Live-action video
Line drawing
+ Coloring regions
SAM2, k-means High High Possible
Figure 1: Processing flow of the proposed method.
production, nor their integration with color design. In
this research, we propose a novel method that com-
bines SAM2 (Ravi et al., 2024) and k-means cluster-
ing to address these challenges. Leveraging SAM2’s
high segmentation accuracy and interactive versatil-
ity, we accurately extract user-specified objects and
generate line drawings and coloring regions. Further-
more, by using k-means clustering to reduce colors
in coloring regions, we automatically separate basic
color regions and shadow regions, facilitating color
design. This significantly streamlines the rotoscop-
ing process in anime production and supports high-
quality animation production.
Our research aims to hierarchically generate line
drawings and coloring regions that can be integrated
into Japanese anime production workflows, achieving
high-precision segmentation and efficient color de-
sign using SAM2 and k-means clustering.
3 PROPOSED METHOD
We propose the method to convert live-action video
into drawable materials with symbolic expressions of
simple lines and color separation, the processing flow
of the proposed method is shown in Figure 1. The
proposed method consists of a line drawing process
using SAM2, shading using the k-means method, and
finishing. Details of each process are described below.
3.1 Line Drawing Creation Using SAM2
We use Segment Anything Model 2 (SAM2) (Ravi
et al., 2024) to generate line drawings and color-
ing regions. SAM2 is a segmentation model that
can handle both images and videos, enabling high-
precision and fast processing. It particularly excels at
identifying spatial and temporal ranges of objects in
videos, capable of high-precision segmentation even
for fast-moving objects or partially occluded objects
necessary for rotoscoping. Additionally, as it is de-
signed for interactive interfaces, users can easily spec-
ify target objects with simple operations and obtain
desired segmentation results. This interactive se-
mantic segmentation capability combined with high-
precision segmentation ability is the main reason we
adopted SAM2 for this research. Specifically, we
first extract frame images from the input video and
use SAM2 to segment target objects (target silhou-
ettes, hair, clothes, accessories, etc.). In this pro-
cess, users instruct SAM2 on segmentation targets us-
ing prompts. In this case, we executed SAM2’s im-
age segmentation using user clicks on frame images
as input and used the obtained segmentation areas as
prompts for video segmentation. For wide shots, we
segment main parts such as hair, skin, clothes, and
shoes, while for close-ups, we additionally segment
details like eyes, mouth, and accessories. SAM2 gen-
erates segmentation areas for each object based on the
specified prompts. The contours obtained by apply-
Semi-Automatic Generation of Rotoscope Animation Using SAM and k-means Clustering
365

Table 2: Properties of videos used in the experiment.
Video ID Contents Resolution (W × H) fps Time
1 A woman dancing in a long shot 1920×1080 25 11s
2 A preening penguin in a long shot 1080×1920 25 10s
ing the Canny method to these segmentation areas are
adopted as line drawings and used as the basis for gen-
erating coloring areas in subsequent processing.
3.2 Shading by Color Reduction in
Basic Coloring Areas
In anime, shade is typically expressed as regions col-
ored differently from the basic colors. To reproduce
this characteristic, we perform color reduction of col-
oring regions using k-means clustering to extract ba-
sic color regions and shadow regions.
First, using the SAM2 segmentation areas ob-
tained in Sec. 3.1, we extract only the target object
regions from the original live-action video. We apply
k-means clustering to this extracted image, dividing
it into two clusters. This extracts high-brightness ar-
eas as basic color regions and low-brightness areas as
shadow regions.
3.3 Final Coloring and Compositing
Using the line drawings and coloring regions (basic
color regions, shadow regions) generated in Sec. 3.1
and Sec. 3.2, we generate the final animation materi-
als. First, based on the basic color regions and shadow
regions obtained in Sec. 3.2, users develop a color de-
sign and assign appropriate colors to each region by
specifying colors in a palette.
Next, when creating animation, we use the line
drawings obtained in Sec. 3.1 and the sequential im-
ages of colored basic color regions and shadow re-
gions as materials. By compositing these materials
with background images and applying shooting pro-
cesses (effects, color correction, etc.) as needed, we
complete the final animation.
4 EXPERIMENTS AND RESULTS
We describe the experimental conditions of the pro-
posed method and the results and their evaluation.
The evaluation of the proposed method was con-
ducted using both quantitative and qualitative assess-
ments. For quantitative evaluation, we measured the
processing time of each processing step. For qualita-
tive evaluation, we conducted a subjective evaluation
Table 3: Experimental environment.
OS Windows 11 Home 64bit
CPU Intel Core i7-13700KF @3.0GHz
RAM 32GB
GPU
NVIDIA GeForce
RTX 4070Ti 12GB
Software Adobe After Effects
experiment with 134 participants, assessing two crite-
ria: ”rotoscope-likeness” and ”anime-likeness.”
4.1 Experimental Conditions
To verify the effectiveness of our proposed method,
we conducted experiments using two types of videos
containing diverse movements and subjects. Table 2
describes the content and conditions of the videos
used. The experimental environment is as shown in
Table 3. We used default values for all parameters of
SAM2 and the Canny method. For k-means cluster-
ing, we specified the number of clusters as 2.
4.2 Result Images
Line drawings, coloring area, and still images of the
final animation generated using the proposed method
are shown in Figures 2.
Limitations: Accurate segmentation cannot be per-
formed when the segmentation target is blurred or un-
Figure 2: The result of processing on long shot video of
a dancing woman. From top to bottom: live-action video,
generated line drawing, after shading, and after coloring.
GRAPP 2025 - 20th International Conference on Computer Graphics Theory and Applications
366

clear. The method needs to be applied to videos where
the rotoscoping targets are clearly visible.
4.3 Processing Time
Using the experimental environment shown in Ta-
ble 3, we measured the processing time for two steps:
segmentation using SAM2, and line drawing extrac-
tion using the Canny method combined with shadow
region generation using k-means clustering. The
user prompt input time, while not included in these
measurements, averaged approximately 1 minute per
video for initial segmentation setup. This operation
time is proportional to the number of segmentation
regions. This initial investment significantly reduces
manual rotoscoping time, which typically requires
15-20 minutes per frame for traditional methods. The
average FPS for the segmentation portion of the pro-
cess using SAM2 was 3.01 fps, and the average FPS
per segmentation area was 1.16 fps. The number of
segmented regions for Video 1 was 6, Video 2 was
7. On the other hand, the processing using the Canny
method and k-means clustering for line drawing and
shadow generation achieved an average FPS of 3.69
fps, with an average FPS per segmentation area of
0.63 fps. The results suggested that the segmenta-
tion processing using SAM2 is significantly affected
by the increase in the number of segmentation areas.
4.4 Quality Evaluation
To evaluate the quality of rotoscope animation pro-
duced by our proposed method, we conducted a sub-
jective evaluation experiment with 134 participants,
both male and female university students in their 20s.
Participants were shown videos processed using three
different methods of converting live-action footage to
anime - our proposed method (Ours), the cartoon ef-
fect in Aftereffects (Cartoon), and k-means clustering
+ Canny method (k-means) - for each of 2 different
video sections (Video 1 and 2). Since the k-means
clustering + Canny method is a classic method of
transforming cartoon-like images, it was used as the
comparison target in this experiment. In this method,
let k be 6. They were asked to evaluate two crite-
ria, ”rotoscope-likeness” and ”anime-likeness,” on a
5-point scale (1: Strongly disagree to 5: Strongly
agree).
To help them judge the rotoscope-like nature of
the project, we explained what rotoscope was before-
hand and had them watch line drawings and anima-
tions created using rotoscope. The order in which
the methods were presented was randomized in each
video section. Figures 3 and 4 show the input video
and the results of image processing by each method.
The mean evaluation values for each method in
each video are shown in Figures 5 and 6. For each
video and evaluation criterion, we used Friedman’s
test to determine whether there were significant dif-
ferences among the methods. Subsequently, we con-
ducted Wilcoxon signed-rank tests with Bonferroni
correction as post-hoc tests to perform pairwise com-
parisons between methods. Figures 5 and 6 are
marked with ∗ for pairs with p-values below the 5%
significance level, ∗∗ for pairs with p-values below
the 1% significance level, and ∗ ∗ ∗ for pairs below
the 0.1% significance level, respectively. Addition-
ally, we calculated Cliff’s delta as an effect size mea-
sure and evaluated its magnitude in four levels: negli-
gible, small, medium, and large.
These results showed that our proposed method
received significantly higher evaluations in both
”rotoscope-likeness” and ”anime-likeness” compared
to the other two methods in most cases. The supe-
riority of our proposed method was particularly no-
table in Video 1 (woman dancing) , where the effect
size is large. The reason why Video 2’s video eval-
uation was comparable to the cartoon filter might be
because the penguin’s inherently limited color palette
already made it work as a symbolic expression of sim-
ple lines and color separation. These results suggest
that our proposed method can effectively express both
rotoscope-likeness and anime-likeness when generat-
ing animation materials from live-action footage, par-
ticularly in depicting human movements and details.
5 CONCLUSIONS AND FUTURE
WORK
In this research, we proposed a novel method that
automates the rotoscoping process in anime pro-
duction by combining SAM2 and k-means cluster-
ing. The experimental results suggested that our pro-
posed method could achieve superior results in both
rotoscope-likeness and anime-likeness compared to
conventional methods. This effect was particularly
notable in depicting complex human movements and
details.
Future challenges include improving segmenta-
tion accuracy, enhancing line drawing expression, and
accommodating various anime styles. Improving line
drawing expression involves enhancing line drawing
details. In this study, we gave up on adding appropri-
ate lines for finger shapes and clothing wrinkles. This
was because it was difficult to add symbolic lines as
intended by humans, and because k-means clustering-
based shadowing increased the information content of
Semi-Automatic Generation of Rotoscope Animation Using SAM and k-means Clustering
367

Figure 3: Comparison of the input video (Video 1) and the results after each image processing. From left: input video, ours,
comic, k-means.
Figure 4: Comparison of the input video (Video 2) and the
results after each image processing. From left: input video,
ours, comic, k-means.
cells, making it better not to add lines when consider-
ing information control. While k-means clustering-
based shadowing is highly rated for its accurate
edge-line shadowing and three-dimensional shadow-
ing necessary for anime drawing, it appears stiff as
it fails to appropriately reduce the image information
content. To make it more anime-like, we could con-
sider moderately enhancing line drawing details while
maintaining appropriate and accurate shadowing.
Additionally, having options to change only faces
or clothing in footage with other elements would
make it more manageable in production. It would
also be beneficial to be able to change the lighting of
subjects. We expect image generation AI to provide
these two options. We also envision adapting to other
styles, such as creating two shadow regions instead of
one and adding highlights to the drawings.
We believe that there should be an option to re-
move the number of frames in the video appropri-
ately. The advantage of rotoscoping is that it en-
ables anime creation with frame-by-frame shooting
that fills all frames within a second. However, if
it’s possible to select frames capturing lively sub-
jects from the footage, reducing frames might im-
prove quality. Since there are a number of existing
studies (Koroku and Fujishiro, 2022; Miura et al.,
2014) on frame dropping, We believe that introduc-
ing them into the proposed system will help create
more animated images. By addressing these chal-
lenges from this research, our proposed method is ex-
pected to significantly contribute to the efficiency of
animation production.
GRAPP 2025 - 20th International Conference on Computer Graphics Theory and Applications
368

(a) Video 1
(b) Video 2
Figure 5: Results of subjective evaluation experiment
(rotoscope-likeness).
ACKNOWLEDGEMENTS
This work was supported by Toyo University Top Pri-
ority Research Program.
REFERENCES
Agarwala, A., Hertzmann, A., Salesin, D. H., and Seitz,
S. M. (2004). Keyframe-based tracking for ro-
toscoping and animation. ACM Trans. Graph.,
23(3):584–591.
Canny, J. (1986). A computational approach to edge de-
tection. IEEE Transactions on Pattern Analysis and
Machine Intelligence, PAMI-8(6):679–698.
Chen, J., Liu, G., and Chen, X. (2020). Animegan: A novel
lightweight gan for photo animation. In Artificial In-
telligence Algorithms and Applications: 11th Interna-
tional Symposium, ISICA 2019, Guangzhou, China,
November 16–17, 2019, Revised Selected Papers 11,
pages 242–256. Springer.
Chen, Y., Lai, Y.-K., and Liu, Y.-J. (2018). Cartoongan:
(a) Video 1
(b) Video 2
Figure 6: Results of subjective evaluation experiment
(anime-likeness).
Generative adversarial networks for photo cartooniza-
tion. In Proceedings of the IEEE conference on com-
puter vision and pattern recognition, pages 9465–
9474.
Dissanayake, S., Ayoob, M., and Vekneswaran, P. (2021).
Autoroto: Automated rotoscoping with refined deep
masks. In 2021 International Conference on Smart
Generation Computing, Communication and Net-
working (SMART GENCON), pages 1–6.
DOMOAI PTE. LTD (2024). DomoAI. https://domoai.
app/. (Accessed on 11/20/2024).
Duan, Z., Wang, C., Chen, C., Qian, W., and Huang,
J. (2024). Diffutoon: High-resolution editable toon
shading via diffusion models. In Larson, K., editor,
Proceedings of the Thirty-Third International Joint
Conference on Artificial Intelligence, IJCAI-24, pages
7645–7653. International Joint Conferences on Artifi-
cial Intelligence Organization.
Esser, P., Kulal, S., Blattmann, A., Entezari, R., M
¨
uller, J.,
Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F.,
et al. (2024). Scaling rectified flow transformers for
high-resolution image synthesis. In Forty-first Inter-
national Conference on Machine Learning.
Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C.,
Gustafson, L., Xiao, T., Whitehead, S., Berg, A. C.,
Semi-Automatic Generation of Rotoscope Animation Using SAM and k-means Clustering
369

Lo, W.-Y., Doll
´
ar, P., and Girshick, R. (2023). Seg-
ment anything. arXiv:2304.02643.
Koroku, Y. and Fujishiro, I. (2022). Anime-like motion
transfer with optimal viewpoints. In SIGGRAPH Asia
2022 Posters, SA ’22, New York, NY, USA. Associa-
tion for Computing Machinery.
Miura, T., Kaiga, T., Katsura, H., Tajima, K., Shibata, T.,
and Tamamoto, H. (2014). Adaptive keypose extrac-
tion from motion capture data. Journal of Information
Processing, 22(1):67–75.
Ravi, N., Gabeur, V., Hu, Y.-T., Hu, R., Ryali, C., Ma,
T., Khedr, H., R
¨
adle, R., Rolland, C., Gustafson, L.,
Mintun, E., Pan, J., Alwala, K. V., Carion, N., Wu,
C.-Y., Girshick, R., Doll
´
ar, P., and Feichtenhofer, C.
(2024). Sam 2: Segment anything in images and
videos. arXiv preprint arXiv:2408.00714.
Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and
Ommer, B. (2022). High-resolution image synthesis
with latent diffusion models. In Proceedings of the
IEEE/CVF conference on computer vision and pattern
recognition, pages 10684–10695.
Tanaka, T. (2021). Ani Man LLust – Tatsuyuki Tanaka Art
Techniques. SMIRAL Co.,Ltd. p. 4, line 15.
Torrejon, O. E., Peretti, N., and Figueroa, R. (2020). Roto-
scope automation with deep learning. SMPTE Motion
Imaging Journal, 129(2):16–26.
Tous, R. (2024). Lester: Rotoscope animation through
video object segmentation and tracking. Algorithms,
17(8):330.
GRAPP 2025 - 20th International Conference on Computer Graphics Theory and Applications
370