1D-DiffNPHR: 1D Diffusion Neural Parametric Head Reconstruction
Using a Single Image
Pragati Jaiswal
1,2
, Tewodros Amberbir Habtegebrial
1,2
and Didier Stricker
1,2
1
RPTU, Technische Universit
¨
at Kaiserslautern, Germany
2
DFKI, German Research Center for Artificial Intelligence, Germany
Keywords:
Single-View Reconstruction, 3D Face Reconstruction, 1D-Diffusion.
Abstract:
In the field of 3D reconstruction, recent developments, especially in face reconstruction, have shown consid-
erable promise. Despite these achievements, many of these techniques depend heavily on a large number of
input views and are inefficient limiting their practicality. This paper proposes a solution to these challenges
by focusing on single-view, full 3D head reconstruction. Our approach leverages a 1D diffusion model in
combination with RGB image features and a neural parametric latent representation. Specifically, we train a
system to learn latent codes conditioned on features extracted from a single input image. The model directly
processes the input image at inference to generate latent codes, which are then decoded into a 3D mesh. Our
method achieves high-fidelity reconstructions that outperform state-of-the-art approaches such as 3D Mor-
phable Models, Neural Parametric Head Models, and existing methods for head reconstruction.
1 INTRODUCTION
3D head reconstruction has become a pivotal area
in computer vision due to its applications in vir-
tual/augmented reality, medical imaging, and surgi-
cal planning. Traditional multi-view reconstruction
techniques (Furukawa and Ponce, 2009; Sch
¨
onberger
et al., 2016; Campbell et al., 2008) rely on feature
matching across multiple images, requiring extensive
image overlap and computational resources. While
effective, their high computational cost limits appli-
cability, particularly in single-view scenarios.
To overcome these limitations, single-view ap-
proaches such as 3D Morphable Models (3DMMs)
(Blanz and Vetter, 1999) use PCA to represent fa-
cial geometry in a low-dimensional space, enabling
reconstructions from a single image. Advanced mod-
els like DECA (Feng et al., 2021), MICA (Zielonka
et al., 2022), HI-Face (Chai et al., 2023), and HRN
(Lei et al., 2023) have improved reconstruction qual-
ity but struggle with full-head details, including com-
plex features like hair and nuanced expressions, due
to the inherent limitations of PCA. To address the lim-
itations of 3DMMs, Neural Parametric Head Model
(NPHM) (Giebenhain et al., 2023) offers significant
advantages in 3D head reconstruction by explicitly
learning deformable head shapes and facial expres-
sions but it depends on 3D point cloud data. In con-
Figure 1: We introduce 1D-DiffNPHR, a high-fidelity 3D
head reconstruction method using a single image. The fig-
ure shows the 360
◦
view of the reconstructed 3D geometry
given an image.
trast, the Monocular Neural Parametric Head Model
(MonoNPHM) (Li et al., 2023) generates high-fidelity
3D heads from monocular video, capturing dynamic
expressions but its reliance on optimization in 2D
space decreases the accuracy. The 3D is estimated
using the loss in 2D rendered space, which can result
in inconsistencies with the geometry and appearance,
particularly when compared to 3D ground truth.
Diffusion-based models, such as Morphable Dif-
fusion (Wang et al., 2024) and Diffusion Rig
(Zheng Ding and Zhang, 2023), focus on achieving
efficient novel view synthesis by generating consis-
tent 3D-aware images. While methods like Rodin
Jaiswal, P., Habtegebrial, T. A. and Stricker, D.
1D-DiffNPHR: 1D Diffusion Neural Parametric Head Reconstruction Using a Single Image.
DOI: 10.5220/0013310300003905
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 393-400
ISBN: 978-989-758-730-6; ISSN: 2184-4313
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
393

(Zhao et al., 2022) and Rodin HD (Zhao et al., 2023)
employ triplane-based approaches to produce high-
quality 3D-aware synthesis while aiming to optimize
efficiency compared to traditional volumetric render-
ing. However, the potential of using a more efficient
approach is still open. This limitation motivates our
approach, which introduces a diffusion-based model
using a 1D diffusion architecture to streamline the re-
construction process.
We introduce 1D-DiffNPHR, a novel method for
high-fidelity 3D head reconstruction from a single im-
age. Our contributions are as follows:
• 1D Diffusion for Full-Head Reconstruction:
Our 1D diffusion model achieves state-of-the-art
(SOTA) accuracy and inference time in recon-
structing detailed full 3D head geometry, includ-
ing facial expressions and its transfer from a sin-
gle RGB image.
• Enhanced Feature Integration: By directly in-
tegrating RGB image features into the diffusion
process and calculating losses in a compact 1D la-
tent space, we eliminate the need for complex 2D
or 3D loss computations, optimizing the training.
2 RELATED WORK
2.1 3D Morphable Models
3DMMs have been pivotal for facial reconstruction,
using linear subspaces derived from 3D scans to de-
fine shape and texture (Cao et al., 2013; Booth et al.,
2017; Booth et al., 2016). Subsequent advance-
ments leveraged larger datasets and enhanced statis-
tical models for controllable expressions and finer de-
tails (Ploumpis et al., 2019; Tran et al., 2019; Tran
and Liu, 2018; Li et al., 2017). Despite progress,
3DMMs struggle with representing complex topolo-
gies like full heads and diverse hairstyles, requiring
more flexible approaches. Neural Head (Grassal et al.,
2022) addresses this by combining 3DMM priors with
deep neural networks for high-frequency detail. How-
ever, PCA’s inherent smoothing limits 3DMMs’ abil-
ity to capture realistic and diverse facial structures.
2.2 Neural Parametric Head Model
NPHM and Deferred Diffusion (Kirschstein et al.,
2023) demonstrate significant potential in 3D head
reconstruction. NPHM excels in high-quality 3D re-
construction by learning deformable head shapes and
expressions, but its reliance on 3D point cloud data
can hinder practicality. Deferred Diffusion integrates
multi-view video data with diffusion techniques for
view consistency but requires substantial effort in data
capture and preprocessing. Both methods face chal-
lenges in real-world scenarios due to high data re-
quirements and preprocessing demands.
2.3 Diffusion-Based Models
Diffusion models have shown promise in generating
3D. Techniques like DreamFusion (Poole et al., 2022)
and Diffrf (M
¨
uller et al., 2023) generate 3D objects as
radiance fields, but often lack fine detail and realistic
texture. Control3Diff (Gu et al., 2023) uses 2D diffu-
sion to sample triplanes for multiview rendering but
struggles with resolution, occlusion, and view incon-
sistencies.
In contrast, 1D diffusion models excel in spatio-
temporal applications like time-series forecasting and
trajectory generation, as demonstrated by RecFu-
sion (B
´
en
´
edict et al., 2023) and DiffTraj (Zhu et al.,
2023b). While effective in 1D contexts, these models
have not been explored for 3D reconstruction. Our
work extends the concept of 1D diffusion by applying
it to full 3D head reconstruction, leveraging the sim-
plicity and efficiency of 1D diffusion to overcome the
complexity of higher-dimensional diffusion models in
3D facial and head modelling.
3 METHOD
We introduce a 3D head reconstruction pipeline as
shown in Fig. 2, which takes an RGB image to pro-
duces a high-fidelity 3D head mesh that can be used
directly in standard graphics pipelines. Leveraging
pre-trained 2D facial models for visual details and la-
tent representations for geometry, the approach gener-
ates a morphable 3D head suitable for animations and
deformations.
3.1 Face Image Embeddings
Given an input image I with dimensions 512 × 512 ×
3, we utilize FaRL (Face Representation Learning)
(Zheng et al., 2021) we generate a conditional em-
bedding z
cond
∈ R
d
, where d is the dimension of the
embedding vector (in our case, d = 512). This em-
bedding encodes critical facial attributes, including
identity and expression-specific features, essential for
guiding the diffusion process of 3D head reconstruc-
tion.
z
cond
= M(I) (1)
where M(I) is the function for obtaining the condi-
tional embedding from the input image.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
394

Figure 2: 1D-DiffNPHR: The pipeline for 3D head reconstruction using a 1D diffusion model. On the left is the illustration of
the process for deriving ground truth latent codes (z
geo
, z
app
, and z
exp
). They are randomly initialized and iteratively optimized
using a frozen decoder. On the right, we have the training phase employing a 1D U-Net with attention to denoise latent
representations, guided by time-step and conditional embeddings from a pre-trained embedding model. During inference
phase, the diffusion model generates latents from the input image, which are decoded to reconstruct a detailed 3D head.
3.2 Conditional 1D Latent Diffusion
Our model employs a 1D U-Net with attention layers
to generate latent z
true
∈ R
4456×1
(see Section 4.1.1)
guided by a conditional embedding z
cond
∈ R
512×1
.
At each timestep t, the U-Net refines the latent using:
Latent representation z
t
: Refined progressively
to reduce noise.
Conditional embedding z
cond
: Derived from the
input image I to ensure retention of unique facial fea-
tures of each individual.
Time-step embedding γ(t): Encodes timestep for
noise management.
The U-Net processes z
t
via attention and residual
blocks to estimate noise
ˆ
ε
t
, iteratively refining z
t
to
produce the final latent z
final
. Training minimizes the
Mean Squared Error (MSE) between z
model
and z
true
,
with the conditional embedding z
cond
guiding model
to generate accurate 3D head representation. The loss
function for this training process is defined as:
L
total
=
T
∑
t=0
λ
t
L(t) (2)
where λ
t
is a timestep-specific weighting factor, and
L(t) represents the MSE loss at timestep t.
3.3 3D Reconstruction via MonoNPHM
Decoder
Once we obtain the denoised latent from diffu-
sion, z
final
, it is split back into its individual sub-
components z
geo
, z
exp
, and z
app
, and scaled for the
decoder (see Section 4.1.1). These act as input to
the MonoNPHM decoder that converts these latents
to 3D mesh. The decoder combines these latents to
generate:
Geometry: SDF(x) to represent the 3D geometry
of the head, including both static structural features
and dynamic expressions, and
Appearance: RGB(x) to add realistic colour to
the reconstructed face.
This provides a high-quality 3D head, preserving
identity, expressions, and appearance from the input
image.
4 TRAINING AND INFERENCE
4.1 Dataset
For our training, we used the NPHM dataset, compris-
ing over 9,200 high-quality head scans from 488 iden-
tities. Captured with a custom 3D scanning setup, the
dataset features diverse facial shapes and expressions,
enabling the model to learn detailed and generalized
representations of human faces.
4.1.1 Latent Representation of 3D Meshes
To encode essential attributes of facial geometry, ex-
pression, and appearance, each 3D mesh is con-
verted into a compact latent representation using the
MonoNPHM decoder. This process yields three types
of latent codes:
Geometric Latent (z
geo
∈ R
2176
): Encodes the
structural features of the 3D face,
Appearance Latent (z
app
∈ R
2176
): Captures the
surface colour, and
1D-DiffNPHR: 1D Diffusion Neural Parametric Head Reconstruction Using a Single Image
395

Expression Latent (z
exp
∈ R
100
): Captures
changes in facial morphology.
The latent codes are obtained by initialising ran-
dom values for each 3D head and then iteratively op-
timising these initial values. The optimization process
involves passing the random latent vectors through
the frozen MonoNPHM decoder and minimizing loss.
For calculating the loss we precompute SDF(x) and
RGB(x) values for near surface points x and optimize
for latent codes:
argmin
z
geo
,z
exp
,z
app
∑
x∈P
λ
SDF


D
geo
(D
exp
(x)) − SDF(x)


+ λ
RGB


D
app
(D
geo
(D
exp
(x))) − RGB(x)


(3)
After optimization, the latent codes are rescaled
from an arbitrary range to [−1, 1] for diffusion train-
ing stability. For example, the geometric latent is
rescaled as:
z
geo rescaled
= 2 ×

z
geo
− z
geo min
z
geo max
− z
geo min

− 1 (4)
where: z
geo min
, z
geo max
are dataset specific extrema.
The appearance and expression latents are rescaled
similarly.
The rescaled latents are concatenated into a uni-
fied vector:
z
true
= [z
geo rescaled
, z
app rescaled
, z
exp rescaled
, 0
4
] (5)
where 0
4
is a zero-padding vector to ensure a fixed
dimensionality of 4456.
This final latent vector serves as the ground truth
for training the diffusion model, encapsulating de-
tailed structural and expressive features for accurate
3D head representation.
4.1.2 Rendering Input Images from 3D Meshes
To generate inputs for the latent diffusion process,
high-quality 3D meshes are rendered into 2D images
using Trimesh (Dawson-Haggerty et al., ) and Pyren-
der (Matl et al., 2019). A scene is set up with the 3D
mesh at the origin, a perspective camera for proper
framing, and a point light source for illumination. The
resulting images are used for embedding generation
and serve as diffusion model conditions.
4.2 Inference Process
During inference, the input image I is processed by
FaRL to generate a conditional embedding z
cond
∈
R
512×1
, which guides the diffusion model. The model
iteratively refines the latent representation to gener-
ate z
final
, which is decoded by the MonoNPHM de-
coder to reconstruct the 3D mesh geometry of the in-
put face. The results (Fig. 3) show accurate geome-
try, contours, and features, with minimal deviation in
Figure 3: 3D reconstruction comparison against ground
truth. Our model accurately reconstructs 3D geometry from
a single image, closely matching the ground truth 3D mesh.
Figure 4: Results of expression transfer. Given an identity
image and expression image, transfer of expression on iden-
tity.
complex regions like the nose, mouth, and eyes. This
demonstrates the model’s capability to handle diverse
geometries and fine details, supporting realistic 3D fa-
cial modelling.
For the transfer of expression from one image to
a different identity, FaRL is used to extract embed-
dings from both an identity image and an expression
image. After the diffusion process, geometry (z
geo
)
and appearance (z
app
) latents are taken from the iden-
tity image, while expression (z
exp
) is derived from
the expression image. These combined latents form
the final representation, decoded by the MonoNPHM
decoder to produce a 3D head with the first image’s
identity and the second image’s expression. This en-
ables precise manipulation of facial identity and ex-
pressions, accurately transferring expressions while
preserving identity, as shown in Fig. 4.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
396

4.3 Implementation Details
We utilise Pytorch (Paszke et al., 2019) for the imple-
mentation. For the model architecture, we employed a
U-Net (Ronneberger et al., 2015) as the primary diffu-
sion backbone. A single input image of 512 × 512 res-
olution was used, optimizing data usage while main-
taining detailed output fidelity. We use pre-trained
FaRL that was trained on LAIONFace20M (Zheng
et al., 2022) for 64 epochs. We train our model for
4500 epochs, using a batch size of 24 and a learning
rate of 8e−5 for 48 hours on 2 H100 GPUs.
5 BENCHMARKING AGAINST
SOTA
In this section, we present and analyze the re-
sults of our model and compare its performance
with other SOTA methods in 3D head reconstruction
such as Morphable Diffusion (Diffusion-based), HRN
(3DMM-based), and MonoNPHM (NPHM-based).
We evaluate reconstruction fidelity, focusing on high-
precision geometry and expression accuracy, as well
as assessing the efficiency of our model in terms of
inference time and accuracy-to-time ratio.
5.1 Evaluation Details
5.1.1 Test Data
We evaluated on the NPHM and Facescape (Zhu et al.,
2023a) datasets. For NPHM, a separate test set of
six diverse identities was used, including challenging
cases like long hair and facial hair. Each subject had
a random and neutral expression for testing general-
ization. On FaceScape, real images of varied subjects
were used without retraining, assessing cross-dataset
generalization to different settings. This dual evalua-
tion highlights the model’s adaptability to new identi-
ties and real-world variations.
5.1.2 Metrics
To quantify reconstruction accuracy, we employ the
unidirectional L1-Chamfer distance (in centimetres)
as our primary evaluation metric.
In addition to accuracy, we evaluate model
efficiency by calculating inference time and the
accuracy-to-time ratio.
5.1.3 Evaluation Protocol
For a fair comparison, we align each generated mesh
with the ground truth using a two-step alignment pro-
cess: initial manual registration followed by refine-
ment with Iterative Closest Point (ICP) (Besl and
McKay, 1992).
Inference time was measured on an NVIDIA RTX
4090 GPU under identical conditions. The evalua-
tion was conducted using the default configurations
of each method, with an input image of fixed reso-
lution. Multiple runs were performed to calculate the
average inference time, reducing the impact of system
variability.
5.2 Quantitative Results
As shown in Table 1, our model achieves the low-
est L1-Chamfer distances, with 37.5% and 38.4% im-
provements for neutral and expressive faces respec-
tively compared to its closest competitor. Evaluations
on the FaceScape dataset confirm this trend, with a
41.3% improvement, demonstrating strong general-
ization Table 2.
Table 1: Quantitative comparison on NPHM test dataset for
a neutral and a random expression against SOTA. Reported
values are Chamfer distance with average surface error in
centimetres.
Accuracy: Neutral Face (↓)
Subject
Ids
Model
MonoNPHM HRN Morphable
Diffusion
Ours
1 1.28 9.55 1.64 0.80
2 0.79 7.91 0.84 0.38
3 0.71 8.60 1.35 0.57
4 1.69 9.78 2.09 1.05
5 1.02 8.95 1.53 0.67
6 1.23 9.46 1.88 0.75
Average 1.12 9.04 1.56 0.70
Accuracy: Expressive Face (↓)
1 (Lips Down) 1.05 8.12 1.87 0.68
2 (Squeeze) 0.89 8.16 1.03 0.39
3 (Grin) 0.76 8.17 1.00 0.76
4 (Mouth Stretch) 1.97 10.30 2.35 1.44
5 (Angry) 1.12 8.21 1.20 0.49
6 (Smile) 1.26 8.63 1.62 0.58
Average 1.17 8.60 1.51 0.72
Table 2: Quantitative comparison on FaceScape dataset
against SOTA. Reported values are Chamfer distance with
average surface error in centimetres.
Accuracy (↓)
Subject
Ids
Model
MonoNPHM HRN Morphable
Diffusion
Ours
1 2.55 8.27 1.60 1.22
2 1.26 6.21 1.99 0.88
Average 1.90 7.24 1.79 1.05
1D-DiffNPHR: 1D Diffusion Neural Parametric Head Reconstruction Using a Single Image
397

Table 3: Comparison against SOTA based on inference time
and efficiency (accuracy/time ratio).
Method MonoNPHM HRN Morphable
Diffusion
Ours
Time(mm:ss) 06:55 00:48 02:51 1:11
Accuracy/Time 0.0021 0.0023 0.0038 0.0197
Table 3 highlights our model’s superior balance of
speed and accuracy. Unlike multi-step methods like
Morphable Diffusion and MonoNPHM, our approach
operates directly without added overhead. While
HRN is faster, it sacrifices accuracy. Our method’s
efficiency and performance make it ideal for practical
applications.
5.3 Qualitative Results
Fig. 5 provides a qualitative comparison of re-
constructed meshes and their L1-Chamfer distances
across methods. MonoNPHM generates facial
meshes that resemble human faces but often fail to
preserve identity and detailed expressions, resulting
in higher Chamfer distances for expressive images.
HRN better captures expressions but exhibits rough
reconstructions and lacks accuracy in modelling facial
hair. Morphable Diffusion introduces high-frequency
noise, reducing the fidelity of its 3D meshes.
Our model, however, consistently preserves high-
fidelity geometry and subtle expressions. Simi-
Figure 5: Qualitative results on NPHM test dataset against MonoNPHM, HRN, and Morphable Diffusion. Odds rows are
neutral and even rows are with expression. The colour code is in centimetres.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
398

Figure 6: Qualitative results on FaceScape dataset against MonoNPHM, HRN, and Morphable Diffusion. The colour code is
in centimetres.
lar trends appear in tests on the FaceScape dataset
(Fig. 6), highlighting its robustness in capturing di-
verse facial structures and expressions under varying
conditions.
6 LIMITATIONS
Our model, while strong in 3D head reconstruction
and expression transfer, has notable limitations. The
use of a 1D latent vector z
app
∈ R
2176
constrains de-
tail and realism of texture. Relying on vertex colours
without photorealistic textures or material properties
like specular reflections. A modular approach could
decouple appearance encoding from geometry and ex-
pression, enabling photorealistic texture generation
with dedicated models.
Finally, hair information is insufficiently captured
in the latent representation or conditional embed-
dings, leading to hairstyle reconstruction inaccura-
cies. Specialized mechanisms or hair-specific condi-
tioning inputs could address this gap.
7 CONCLUSION
In this work, we have introduced a 1D diffusion-
based method for high-fidelity 3D head reconstruc-
tion, capable of generating detailed 3D meshes from
a single input image. By leveraging 1D diffusion
framework combined with robust conditioning, our
approach achieves accurate reconstruction of facial
geometry and expressions. The proposed method en-
sures high-quality structural fidelity and expression
accuracy, offering a versatile and scalable solution for
realistic 3D head reconstruction. Through compre-
hensive quantitative and qualitative evaluations, we
demonstrate that our method consistently outperforms
state-of-the-art approaches, excelling in capturing ge-
ometric details and expressive facial variations. This
positions our framework as a powerful tool for ap-
plications that demand precise and detailed 3D head
modelling. Our work marks a significant advance-
ment in 3D head reconstruction, showcasing the po-
tential of small diffusion-based models for generating
high-quality 3D meshes. By bridging the gap between
accuracy and adaptability, this approach lays the foun-
dation for future innovations.
ACKNOWLEDGEMENT
This work was co-funded by the European Union
under Horizon Europe, grant number 101092889,
project SHARESPACE. We sincerely thank Dr. Ren
´
e
Schuster for their invaluable assistance in writing and
proofreading this paper.
REFERENCES
Besl, P. and McKay, N. D. (1992). A method for registration
of 3-d shapes. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 14(2):239–256.
Blanz, V. and Vetter, T. (1999). A morphable model for
the synthesis of 3d faces. In Proceedings of the 26th
annual conference on Computer graphics and inter-
active techniques, pages 187–194.
Booth, J., Antonakos, E., Ploumpis, S., Trigeorgis, G.,
Panagakis, Y., and Zafeiriou, S. (2017). 3d face
morphable models” in-the-wild”. In Proceedings of
the IEEE conference on computer vision and pattern
recognition, pages 48–57.
Booth, J., Roussos, A., Zafeiriou, S., Ponniah, A., and
Dunaway, D. (2016). A 3d morphable model learnt
from 10,000 faces. In Proceedings of the IEEE con-
ference on computer vision and pattern recognition,
pages 5543–5552.
B
´
en
´
edict, G., Jeunen, O., Papa, S., Bhargav, S., Odijk, D.,
and de Rijke, M. (2023). Recfusion: A binomial dif-
fusion process for 1d data for recommendation.
Campbell, N. D., Vogiatzis, G., Hern
´
andez, C., and Cipolla,
R. (2008). Using multiple hypotheses to improve
depth-maps for multi-view stereo. In Computer
Vision–ECCV 2008: 10th European Conference on
1D-DiffNPHR: 1D Diffusion Neural Parametric Head Reconstruction Using a Single Image
399

Computer Vision, Marseille, France, October 12-
18, 2008, Proceedings, Part I 10, pages 766–779.
Springer.
Cao, C., Weng, Y., Zhou, S., Tong, Y., and Zhou, K. (2013).
Facewarehouse: A 3d facial expression database for
visual computing. IEEE Transactions on Visualization
and Computer Graphics, 20(3):413–425.
Chai, Z., Zhang, T., He, T., Tan, X., Baltru
ˇ
saitis, T., Wu, H.,
Li, R., Zhao, S., Yuan, C., and Bian, J. (2023). Hiface:
High-fidelity 3d face reconstruction by learning static
and dynamic details.
Dawson-Haggerty et al. trimesh.
Feng, Y., Wu, F., Shao, X., Wang, Y., and Zhou, X. (2021).
Deca: Deep face model with 3d morphable models.
In Proceedings of the IEEE/CVF Conference on Com-
puter Vision and Pattern Recognition (CVPR), pages
7668–7677.
Furukawa, Y. and Ponce, J. (2009). Accurate, dense, and ro-
bust multiview stereopsis. IEEE transactions on pat-
tern analysis and machine intelligence, 32(8):1362–
1376.
Giebenhain, S., Kirschstein, T., Georgopoulos, M., R
¨
unz,
M., Agapito, L., and Nießner, M. (2023). Learning
neural parametric head models. In Proc. IEEE Conf.
on Computer Vision and Pattern Recognition (CVPR).
Grassal, P.-W., Prinzler, M., Leistner, T., Rother, C.,
Nießner, M., and Thies, J. (2022). Neural head avatars
from monocular rgb videos. In Proceedings of the
IEEE/CVF Conference on Computer Vision and Pat-
tern Recognition, pages 18653–18664.
Gu, J., Gao, Q., Zhai, S., Chen, B., Liu, L., and Susskind, J.
(2023). Control3diff: Learning controllable 3d diffu-
sion models from single-view images.
Kirschstein, T., Giebenhain, S., and Nießner, M. (2023).
Diffusionavatars: Deferred diffusion for high-fidelity
3d head avatars. arXiv preprint arXiv:2311.18635.
Lei, B., Ren, J., Feng, M., Cui, M., and Xie, X. (2023). A
hierarchical representation network for accurate and
detailed face reconstruction from in-the-wild images.
Li, T., Bolkart, T., Black, M. J., Li, H., and Romero, J.
(2017). Learning a model of facial shape and expres-
sion from 4d scans. ACM Trans. Graph., 36(6):194–1.
Li, Y., Zhang, Y., Chen, W., Zheng, Y., Li, H., and Yu, J.
(2023). Mono-nphm: Monocular neural parametric
head model for high-fidelity 3d head reconstruction.
arXiv preprint arXiv:2312.06740.
Matl, M. et al. (2019). Pyrender. GitHub Repository.
M
¨
uller, N., Siddiqui, Y., Porzi, L., Bul
`
o, S. R.,
Kontschieder, P., and Nießner, M. (2023). Diffrf:
Rendering-guided 3d radiance field diffusion.
Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J.,
Chanan, G., Killeen, T., Lin, Z., Gimelshein, N.,
Antiga, L., Desmaison, A., Kopf, A., Yang, E., De-
Vito, Z., Raison, M., Tejani, A., Chilamkurthy, S.,
Steiner, B., Fang, L., Bai, J., and Chintala, S. (2019).
Pytorch: An imperative style, high-performance deep
learning library. In Wallach, H., Larochelle, H.,
Beygelzimer, A., d'Alch
´
e-Buc, F., Fox, E., and Gar-
nett, R., editors, Advances in Neural Information Pro-
cessing Systems, volume 32. Curran Associates, Inc.
Ploumpis, S., Wang, H., Pears, N., Smith, W. A., and
Zafeiriou, S. (2019). Combining 3d morphable mod-
els: A large scale face-and-head model. In Proceed-
ings of the IEEE/CVF Conference on Computer Vision
and Pattern Recognition, pages 10934–10943.
Poole, B., Jain, A., Barron, J. T., and Mildenhall, B. (2022).
Dreamfusion: Text-to-3d using 2d diffusion.
Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:
Convolutional networks for biomedical image seg-
mentation.
Sch
¨
onberger, J. L., Zheng, E., Frahm, J.-M., and Pollefeys,
M. (2016). Pixelwise view selection for unstructured
multi-view stereo. In Computer Vision–ECCV 2016:
14th European Conference, Amsterdam, The Nether-
lands, October 11-14, 2016, Proceedings, Part III 14,
pages 501–518. Springer.
Tran, L., Liu, F., and Liu, X. (2019). Towards high-fidelity
nonlinear 3d face morphable model. In Proceedings
of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition, pages 1126–1135.
Tran, L. and Liu, X. (2018). Nonlinear 3d face morphable
model. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pages 7346–
7355.
Wang, Y., Hu, Y., Guo, Y., Zhang, Z., and Yu, J. (2024).
Morphable diffusion: 3d face reconstruction from a
single image. arXiv preprint arXiv:2401.04728.
Zhao, Y., Liu, Z., Yang, M., and Chen, L. (2022). Rodin:
A realistic 3d face reconstruction method using diffu-
sion models. International Journal of Computer Vi-
sion, 130(3):707–724.
Zhao, Y., Liu, Z., Yang, M., and Chen, L. (2023). Rodinhd:
High-resolution 3d face reconstruction using diffusion
models. arXiv preprint arXiv:2305.05555.
Zheng, Y., Yang, H., Zhang, T., Bao, J., Chen, D., Huang,
Y., Yuan, L., Chen, D., Zeng, M., and Wen, F. (2021).
General facial representation learning in a visual-
linguistic manner. arXiv preprint arXiv:2112.03109.
Zheng, Y., Yang, H., Zhang, T., Bao, J., Chen, D., Huang,
Y., Yuan, L., Chen, D., Zeng, M., and Wen, F. (2022).
General facial representation learning in a visual-
linguistic manner. In Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recogni-
tion, pages 18697–18709.
Zheng Ding, Cecilia Zhang, Z. X. L. J. Z. T. and Zhang,
X. (2023). Diffusionrig: Learning personalized priors
for facial appearance editing. In Proceedings of the
IEEE/CVF Conference on Computer Vision and Pat-
tern Recognition.
Zhu, H., Yang, H., Guo, L., Zhang, Y., Wang, Y., Huang,
M., Wu, M., Shen, Q., Yang, R., and Cao, X.
(2023a). Facescape: 3d facial dataset and benchmark
for single-view 3d face reconstruction. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence
(TPAMI).
Zhu, Y., Ye, Y., Zhang, S., Zhao, X., and Yu, J. J. Q.
(2023b). Difftraj: Generating gps trajectory with dif-
fusion probabilistic model.
Zielonka, W., Bolkart, T., and Thies, J. (2022). Towards
metrical reconstruction of human faces.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
400