Image2Life: A Model for 3D Mesh Reconstruction from a Single-Image

Lynda Ayachi

and Mohamed Rabia Benarbia

Orange Innovation Tunisia, Sofrecom, Tunis, Tunisia

National School of Computer Science, Manouba, Tunisia

Keywords:

3D Reconstruction, Advanced Modeling Techniques, Large Reconstruction Model LRMs, Feature

Aggregation.

Abstract:

Reconstructing 3D models from a single 2D image is a complex yet fascinating challenge with applications

in areas like computer vision, robotics, and augmented reality. In this work, we propose a novel approach to

tackle this problem, focusing on creating accurate and detailed 3D representations from minimal input. Our

model combines advanced deep learning techniques with geometry-aware methods to extract and translate

meaningful features from 2D images into 3D shapes. By introducing a new framework for feature extraction

and a carefully designed decoding architecture, our method captures intricate details and improves the overall

reconstruction quality. We tested the model extensively on well-known datasets, and the results show signiﬁ-

cant improvements compared to existing methods in terms of accuracy and reliability.

1 INTRODUCTION

The art of creating a digital representation of any

object, real or imagined in three dimensions, is the

fundamental component of 3D modeling. Vertices ,

edges and faces must be placed precisely during this

process. The end product is a ﬂexible 3D model that

can be animated, modiﬁed, and generated to suit a va-

riety of purposes. 3D modeling has reshaped how in-

dustries create, visualize, and interact with digital and

physical spaces. In the entertainment world, it is at the

heart of storytelling, allowing studios such as Pixar

and DreamWorks to craft realistic characters and de-

tailed environments for animations and ﬁlms (Whizzy

Studios, 2023). Similarly, the video game industry

depends on 3D models to design complex worlds and

dynamic characters, enhancing the overall gaming ex-

perience (Cutting Edge R, 2023). In architecture and

construction, 3D modeling provides architects and en-

gineers with tools to visualize designs, conduct virtual

walkthroughs, and efﬁciently plan projects. This ap-

proach helps to communicate ideas clearly and facil-

itates better decision-making during the design pro-

cess (Adobe, 2023). The integration of 3D model-

ing into virtual reality (VR) and augmented reality

(AR) has expanded its applications even further. VR

uses 3D models to create fully immersive digital envi-

ronments, while AR overlays virtual objects onto the

physical world, enhancing areas such as education,

training, and interactive experiences (Cutting Edge

R, 2023). In the medical ﬁeld, 3D modeling is used

to generate highly detailed anatomical representations

that support medical education, surgical planning, and

prosthetic design. These models improve understand-

ing of complex biological structures and contribute

to better patient outcomes (Whizzy Studios, 2023).

Also, the product design and manufacturing indus-

tries leverage 3D modeling to speed up prototyping,

enable greater customization, and streamline produc-

tion. By visualizing and reﬁning designs before man-

ufacturing, 3D modeling helps reduce costs and im-

prove efﬁciency (Adobe, 2023). Our solution uses

generative AI to create detailed 3D models from 2D

images, primarily optimized for electronics and gad-

gets while maintaining effectiveness across other ob-

ject types. This approach enhances reconstruction ﬁ-

delity and realism while providing a more efﬁcient

and accessible alternative to traditional 3D modeling

techniques.

2 RELATED WORK

This section will explore advanced techniques in

3D model generation, focusing on three main areas:

Score Distillation for 3D Generation, 3D Generation

with Sparse View Reconstruction, and Feed-forward

3D Generative Models.

Ayachi, L. and Benarbia, M. R.

Image2Life: A Model for 3D Mesh Reconstruction from a Single-Image.

DOI: 10.5220/0013349500003890

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

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 1319-1326

ISBN: 978-989-758-737-5; ISSN: 2184-433X

1319

• Score Distillation for 3D Generation:

Score distillation leverages large-scale image diffu-

sion models to iteratively reﬁne and produce high-

quality 3D models without requiring extensive 3D

datasets. DreamFusion (Poole et al., 2022) pio-

neered this approach with Score Distillation Sam-

pling (SDS), which uses feedback from image diffu-

sion models to align 3D models with target images

or descriptions. This iterative reﬁnement process en-

ables highly detailed results without relying on pre-

existing 3D data. ProliﬁcDreamer (Wang et al., 2023)

builds on this foundation by introducing Variational

Score Distillation (VSD), addressing over-saturation

issues and enhancing output diversity. However, these

techniques face signiﬁcant computational challenges

- generating a single model on an NVIDIA RTX 4090

GPU can take several minutes with DreamFusion,

while complex objects may require hours of process-

ing time, limiting practical applications where speed

is crucial.

• 3D Generation with Sparse View Reconstruc-

tion:

Recent advancements in 3D reconstruction have in-

troduced methods that generate multi-view consistent

images from single-view inputs, which are then trans-

formed into detailed 3D models. SyncDreamer (Liu

et al., 2023c) employs a synchronized multi-view dif-

fusion model to produce images with consistent ge-

ometry and color across various perspectives, facil-

itating accurate 3D reconstruction without extensive

training data. Building upon this, Wonder3D (Long

et al., 2023) utilizes a cross-domain diffusion ap-

proach to generate multi-view normal maps and cor-

responding color images, ensuring coherence across

different views. This method effectively handles com-

plex scenes and textures, resulting in high-ﬁdelity tex-

tured meshes from single-view images. Zero123++

(Shi et al., 2023) further enhances multi-view image

generation by leveraging reﬁned diffusion techniques

to manage dynamic lighting and shadows, produc-

ing realistic and cohesive images from diverse view-

points. For the reconstruction phase, frameworks like

NeuS and Neuralangelo have demonstrated signiﬁ-

cant capabilities. NeuS (Liu et al., 2023c) special-

izes in reconstruction of the neural surface using a

signed distance function to implicitly represent ge-

ometry, which is particularly beneﬁcial in handling

sparse views and self-occlusions. Neuralangelo (Liu

et al., 2023a) advances 3D surface reconstruction by

employing multi-resolution 3D hash grids and neu-

ral surface rendering, achieving high ﬁdelity through

coarse-to-ﬁne optimization strategies.

Despite these advancements, challenges persist,

notably the computational intensity required for pro-

Figure 1: Wonder3D Multi-View Results.

Figure 2: Examples of 3D Models Reconstructed by NeuS.

cessing. For instance, NeuS requires approximately

two hours to process a single object on an NVIDIA

RTX 4090 GPU, while Neuralangelo takes around 40

minutes. Sparse view inputs can lead to incomplete

reconstructions, and ensuring geometric consistency

across different views remains critical.

• Feed-forward 3D Generative Models:

Feed-forward 3D generative models offer a faster and

more efﬁcient alternative for generating 3D objects,

bypassing the iterative optimization processes typi-

cally used in traditional methods. By employing a di-

rect feed-forward approach, these models drastically

reduce the time required for 3D object generation,

making them particularly valuable for applications

where rapid results are essential. Recent advance-

ments in this domain have demonstrated signiﬁcant

progress, aided by extensive datasets like Objaverse

(Deitke et al., 2023), which provides diverse and ro-

bust training data.

One notable example is One-2-3-45 (Liu et al.,

2023b), which transforms single images into de-

tailed textured 3D meshes using Zero123 diffusion

model for multi-view prediction and SparseNeuS for

neural surface reconstruction. It generates results

in 45 seconds without iterative optimization. The

Large Reconstruction Model (LRM) (Hong et al.,

2023) advances feed-forward 3D generation through

transformer-based architecture with reﬁned attention

mechanisms, generating high-quality 3D objects in

10 seconds on an NVIDIA RTX 4090 GPU. While

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1320

the ofﬁcial model is private . Building on LRM,

Triposr (Tochilkin et al., 2023) enhances its perfor-

mance through improved data preprocessing, model

optimization, and training strategies while maintain-

ing the core LRM architecture. These reﬁnements al-

low it to achieve superior reconstruction quality in ap-

proximately the same inference time.

3 Image2Life MODEL

ARCHITECTURE

In this section, we present the detailed architecture

of our model, illustrated in ﬁgure 3, highlighting its

components and their functionalities.

3.1 Image Encoder

The DINO (Self-Distillation with No Labels) model

(Touvron et al., 2023) is utilized as the image encoder

within the architecture due to its ability to learn ro-

bust visual representations without requiring labeled

data. Speciﬁcally, the integration of DINOv2 (Tou-

vron et al., 2023), an enhanced version of DINO, ad-

dresses the demands of dense visual predictions and

ﬁne-grained feature extraction. Object views are pro-

cessed through the encoder to extract feature vec-

tors that encapsulate critical visual details, forming

the foundation for subsequent stages in the 3D re-

construction pipeline. The DINOv2 model incorpo-

rates an optimized self-distillation process that signif-

icantly enhances its proﬁciency in capturing intricate

details and high-resolution features. Input images are

segmented into patches and passed through a multi-

layer transformer comprising 12 layers, each with a

dimensionality of 768 (Touvron et al., 2023). This

process generates detailed feature vectors that repre-

sent the visual properties of each patch while preserv-

ing a coherent understanding of the entire image. The

maintenance of spatial coherence is ensured through

the use of register tokens, which uphold spatial rela-

tionships within the image. This mechanism is criti-

cal for preserving structural integrity, allowing the re-

constructed 3D model to accurately reﬂect the input

image. The self-distillation approach employed by

DINOv2 provides interpretable attention over image

structures and textures, enabling detailed structural

information to guide the reconstruction of geometry

and color in 3D space.

3.2 Feature Aggregator

The Multi-Branch Attention Aggregator is a critical

module that we designed to integrate features from

multiple views into a uniﬁed representation, leverag-

ing the complementary information provided by each

view. This architecture is particularly effective in sce-

narios where different perspectives offer unique in-

sights, enhancing the quality of the aggregated fea-

tures for downstream tasks. Inspired by the patch-

based processing of the DINO encoder (Touvron

et al., 2023), the Feature Aggregator employs a multi-

branch attention mechanism to efﬁciently combine in-

formation from various views. Each view is processed

through a dedicated attention layer comprising se-

quential linear layers interspersed with ReLU activa-

tion functions, enabling each branch to independently

highlight important features. Attention weights are

dynamically calculated using a softmax layer, which

ensures that the most relevant features across all views

are emphasized. The aggregated features are subse-

quently reﬁned through a ﬁnal fully connected output

layer, which adjusts the dimensionality to suit speciﬁc

application requirements. This architecture offers en-

hanced feature representation by adaptively focusing

on the most informative aspects of each view, result-

ing in a richer, more robust output. Additionally, the

modular design of the aggregator allows for ﬂexibil-

ity in the number of views and the dimensionality of

the features, making it highly adaptable across various

settings. This approach signiﬁcantly improves the ef-

ﬁciency and effectiveness of multi-view feature inte-

gration, aligning well with modern 3D reconstruction

pipelines.

3.3 Image-to-Triplane Decoder

The Image-to-Triplane Decoder plays a critical

role in converting high-dimensional image features

into structured 3D representations. By utilizing a

transformer-based architecture, this decoder projects

image features onto learnable spatial-positional em-

beddings, effectively mapping them to a triplane

structure. This approach addresses the inherent chal-

lenges of single-image 3D reconstruction by pro-

viding a comprehensive representation of geomet-

ric and appearance information. The triplane rep-

resentation consists of three planes aligned with the

axes: T

, T

Y Z

, T

, each with a spatial resolution

of 64 × 64 and d

feature channels. Features for

any 3D point within the NeRF bounding box [−1, 1]

are derived through bilinear interpolation, followed

by processing with a multi-layer perceptron (MLP)

to generate NeRF color and density values (Milden-

hall et al., 2020). Learnable spatial-positional em-

beddings are integrated into the decoder, with dimen-

sions 3 × 32 × 32 × d

, to bridge the gap between

image and 3D space. These embeddings query im-

Image2Life: A Model for 3D Mesh Reconstruction from a Single-Image

1321

Figure 3: Image2Life Architecture.

Figure 4: Architecture of The Feature Aggregator.

age features via cross-attention mechanisms, enabling

the projection of image features to the triplane struc-

ture. Given the disparity in token counts between

the initial embeddings 3 × 32 × 32 and the ﬁnal tri-

plane representation 3 × 64 × 64, the transformer out-

put is upsampled to complete the triplane represen-

tation. The decoder employs a transformer archi-

tecture beginning with a token sequence of dimen-

sions (3 × 32 × 32) × 1024, where the token count

3 × 32 × 32 corresponds to the number of spatial to-

kens and 1024 represents the hidden dimension of the

transformer. This conﬁguration includes 1025 image

feature tokens for generating keys and values in the

cross-attention layers. The transformer consists of 16

layers, each with 16 attention heads of dimension 64.

In accordance with methods proposed by Touvron et

al. (Touvron et al., 2023), bias terms are omitted, and

a pre-normalization structure is employed, deﬁned as:

x + f ({LayerNorm}(x)). Each transformer layer con-

sists of a cross-attention sub-layer to integrate image

and triplane features, a self-attention sub-layer to cap-

ture internal triplane relationships, and an MLP sub-

layer for non-linear feature transformations. The op-

erations are deﬁned as (Hong et al., 2023):

cross

= CrossAttn



ModLN

cam

(LayerNorm( f

));{h

}

i=1



+ f

(1)

self

= SelfAttn



ModLN

cam

(LayerNorm( f

cross

))



+ f

cross

(2)

out

= MLP

tfm

(LayerNorm( f

self

)) + f

self

(3)

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1322

This combination of components ensures a robust

and detailed 3D representation derived from high-

dimensional image features.

Then, camera features are constructed by ﬂatten-

ing the 4 × 4 extrinsic matrix E and concatenating it

with the focal length foc and principal point pp, cre-

ating a 20-dimensional vector c. This vector is nor-

malized and then transformed into a high-dimensional

embedding

c via an MLP. These camera features are

used to modulate the triplane tokens through adap-

tive layer normalization. The true camera parameters

c are incorporated into the Image-to-Triplane decoder

only during the training phase. In the inference phase,

these features are substituted with an encoding of the

standard, ﬁxed camera parameters.

γ, β = MLP

mod

(

c) (4)

ModLN

( f

) = LayerNorm( f

) · (1 +γ) + β (5)

After processing through all transformer layers,

the ﬁnal triplane features f

out

are upsampled using a

learnable de-convolution layer to form the ﬁnal tri-

plane representation T . This layer transforms the

transformer output from (3 × 32 × 32) × 1024 to 3 ×

(64 ×64) × 80. The de-convolution layer has a kernel

size of 2, a stride of 2, and no padding.

The Triplane-NeRF (Neural Radiance Fields)

component is designed to estimate RGB values and

density σ from point features derived from the tri-

plane representation T . This approach is based on

the triplane framework described by (Chan et al.,

2022), utilizing a multi-layer perceptron (MLP) to

process these point features effectively. The architec-

ture of the MLP consists of three main parts. First,

the input comprises point features extracted from the

triplane representation T . These features are then

passed through 10 hidden layers, each employing lin-

ear transformations interleaved with ReLU activation

functions and having a dimensionality of 64. Finally,

the output layer generates a 4-dimensional vector,

where the ﬁrst three dimensions represent the RGB

color values, and the fourth corresponds to the density

σ. This structured design enables the MLP to process

the input features and produce precise estimations of

the color and density values.

In terms of the detailed process, querying point

features is a critical step. For each point within the

3D bounding box, the point is mapped onto the tri-

plane planes (T

, T

Y Z

, T

). Bilinear interpolation

is then applied to extract feature values from these

planes, ensuring accurate and efﬁcient computation

for 3D representation tasks.

The features obtained from the three planes are

aggregated into a single feature vector that repre-

sents the 3D point. This combined feature vector is

then processed using the MLP

NeRF

, which consists

of a series of 10 linear layers with ReLU activation

functions. The output from the MLP

NeRF

is a 4-

dimensional vector, where the ﬁrst three dimensions

correspond to the RGB color values, and the fourth

dimension represents the density σ. This structured

pipeline ensures the efﬁcient transformation of the

triplane features into meaningful 3D representations

suitable for rendering.

The design of this framework offers several ad-

vantages. First, the triplane representation provides

a compact yet expressive encoding of 3D structures,

enabling efﬁcient querying of point features. Second,

the use of MLP

NeRF

facilitates accurate predictions of

color and density, resulting in high-quality 3D recon-

structions and renderings. Finally, the modular nature

of the triplane and MLP components ensures scalabil-

ity to complex scenes and objects, making it a versa-

tile solution for diverse 3D tasks.

4 EXPERIMENTATIONS

4.1 Training

The training process for the proposed model is con-

ducted in two stages.

Stage 1: The ﬁrst stage focuses on training the fea-

ture aggregator, a critical component of the model’s

architecture. This stage involves ﬁne-tuning key train-

ing parameters to enhance the model’s efﬁciency and

ensure robust feature extraction and integration. The

training process incorporates techniques and param-

eters derived from the LRM framework, comple-

mented by improvements inspired by recent advance-

ments in Triposr.

To optimize the balance between computational

efﬁciency and reconstruction quality, instead of ren-

dering full-resolution images at a 512×512 resolu-

tion, the model processes smaller 128×128 patches,

prioritizing patches more likely to cover foreground

regions. This approach ensures that the model fo-

cuses on critical areas, enabling detailed surface re-

constructions while maintaining computational efﬁ-

ciency. The rendering process was conﬁgured with

128 samples per ray, a radius of 0.87, an exponential

density activation function, and a density bias of -1.0.

The model was trained over 27 epochs .

Stage 2: In the second stage, we ﬁne-tune the last

ﬁve layers of the Image-to-Triplane Decoder and the

Triplane-NeRF while freezing the rest of the model to

Image2Life: A Model for 3D Mesh Reconstruction from a Single-Image

1323

Figure 5: Training and Validation Loss During Stage 1.

adapt it to speciﬁc tasks. This stage uses the same hy-

perparameters and renderer settings as the ﬁrst stage,

ensuring consistency and stability during training.

4.2 Evaluation

4.2.1 Qualitative Results

We conducted qualitative comparisons to evaluate our

model’s performance against established benchmarks.

Figure 6 compares OpenLRM and our model with

only the feature aggregator trained, using images

sourced from OpenLRM and the MyTek website.

The results demonstrate that our model, leveraging

Zero123++ (Shi et al., 2023), generates more accu-

rate and superior views.

Figures 7 compare our ﬁne-tuned model with

OpenLRM on electronic items from retailers like

Amazon and Orange Store. Our model consis-

tently delivers more realistic geometry and appear-

ance, signiﬁcantly improving results on electronic

objects while generalizing well across various cate-

gories. In Figures 8 comparisons with baselines such

as CRM (Wang et al., 2024) and One-2-3-45 highlight

our model’s advantages.

TripoSR struggles with imaginative capabilities and

often produces degraded textures, while Image2Life

use of 80 channels in the Triplane-NeRF architecture

ensures detailed geometry and textures.

CRM, while recent, fails to generate smooth surfaces,

whereas our model, aided by mask loss during train-

ing, excels in smooth and realistic reconstructions.

4.2.2 Quantitative Results

To evaluate the quantitative performance of our

model, we conducted an extensive analysis using 50

unseen objects from our curated dataset and an ad-

ditional 50 objects from the High-Quality Alignment

Subset of Objaverse-XL. For each object, 15 refer-

ence views were processed, and ﬁve of these views

Figure 6: Comparison between OpenLRM and our model

with only the feature aggregator trained.

were individually used as input for our model to re-

construct the corresponding 3D object. The ren-

dered images were then evaluated against all 15 ref-

erence views to assess the quality and accuracy of

the reconstruction. This evaluation methodology, in-

spired by the approach utilized in the LRM frame-

work, provides a robust assessment of our model’s

reconstruction precision and robustness. Notably,

our model demonstrates signiﬁcant computational

efﬁciency, achieving high-quality object generation

within just 10 seconds. The rendered images were

captured from the following angles:

As shown in Table 2, Image2Life surpasses base-

line models in 2D novel view synthesis metrics,

achieving higher SSIM and LPIPS scores, which

highlight its superior perceptual quality. It also leads

in FID and CLIP-Similarity, demonstrating its ability

to produce visually appealing and contextually accu-

rate images. While its PSNR is slightly lower than

the best-performing baseline, this is attributed to the

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1324

Figure 7: Comparison of ﬁne-tuned model and OpenLRM

on electronic objects.

Table 1: Rendered View Angles (Az: Azimuth, El: Eleva-

tion).

View Az El View Az El

Back 180° 0° Right 270° 0°

Back Left 135° 0° Top 0° 90°

Back Right 225° 0° Top Back 180° 45°

Front 0° 0° Top Front 0° 45°

Front Left 45° 0° Top Left 90° 45°

Front Right 315° 0° Top Right 270° 45°

Left 90° 0°

multi-view diffusion model (Zero123++) generating

less pixel-perfect but perceptually richer views. We

argue that perceptual quality outweighs pixel-level ac-

curacy for novel view synthesis, given the inherent

variability of such views.

Figure 8: Comparison of Our Model with One-2-3-45 and

CRM.

For 3D geometric metrics, Image2Life markedly

outperforms baselines in both Chamfer Distance (CD)

and F-Score (FS), demonstrating superior shape ﬁ-

delity. The addition of mask loss during training ef-

fectively minimizes ’ﬂoater’ artifacts, resulting in a

signiﬁcant improvement in CD.

5 CONCLUSION

In conclusion, our model consistently demonstrates

exceptional performance in generating realistic ge-

ometry and visually appealing appearances, particu-

larly for electronic objects sourced from various on-

line retailers such as Amazon and Orange Store. Its

robust generalization capabilities extend beyond elec-

tronic items, achieving superior results across a di-

verse range of object categories. Qualitative com-

parisons have shown that our ﬁne-tuned model out-

performs baseline approaches by producing smoother

and more accurate reconstructions.

Comprehensive quantitative evaluations further

validate the effectiveness of our approach. Our model

excels in key perceptual quality metrics such as SSIM,

LPIPS, FID, and CLIP-Similarity, while also achiev-

ing state-of-the-art performance in geometric ﬁdelity

metrics like Chamfer Distance (CD) and F-Score

(FS).

Despite these achievements, there is room for im-

provement. Addressing challenges such as textural in-

Image2Life: A Model for 3D Mesh Reconstruction from a Single-Image

1325

Table 2: Quantitative comparison of our method with baseline models across multiple metrics. ↑ indicates that higher values

are better, while ↓ indicates that lower values are better.

Method PSNR ↑ SSIM ↑ LPIPS ↓ CD ↓ FS ↑ FID ↓ CLIP-Similarity ↑

TripoSR 23.681 0.872 0.204 0.246 0.879 25.459 0.812

DreamGaussian 19.204 0.789 0.277 0.382 0.635 57.257 0.815

CRM 22.790 0.891 0.137 0.201 0.802 23.846 0.880

One-2-3-45 18.558 0.726 0.296 0.421 0.633 98.261 0.720

Ours 23.297 0.901 0.124 0.179 0.892 22.547 0.930

Table 3: Comparison of our model with only the feature

aggregator trained against OpenLRM using CD and LPIPS

metrics.

Method CD ↓ LPIPS ↓

OpenLRM 0.271 0.209

Our Model (Only Feature Aggregator Trained) 0.194 0.153

distinctness in occluded regions and reducing the de-

pendency on the quality of multi-view images gener-

ated by Zero123++ are crucial next steps. Enhancing

the initial stages of multi-view image generation or

developing alternative strategies could further boost

the performance and reliability of the 3D reconstruc-

tion process. These advancements will pave the way

for even more robust and precise 3D reconstruction

capabilities in future work.

REFERENCES

Adobe (2023). What is 3d modelling & what is it used for?

Chan, E. R. et al. (2022). Efﬁcient triplane-nerf for 3d ob-

ject reconstruction. In Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern Recogni-

tion (CVPR).

Cutting Edge R (2023). 10 exciting applications of 3d mod-

eling in various industries.

Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel,

O., VanderBilt, E., Schmidt, L., Ehsani, K., Kemb-

havi, A., and Farhadi, A. (2023). Objaverse: A uni-

verse of annotated 3d objects. pages 13142–13153.

Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D.,

Liu, F., Sunkavalli, K., Bui, T., and Tan, H. (2023).

Lrm: Large reconstruction model for single image to

3d. arXiv preprint arXiv:2311.04400.

Liu, J., Zhang, Z., Wang, X., Li, S., Zhang, Z. Y.,

Yang, M.-Y., Kautz, J., Hilliges, O., and Tulsiani, S.

(2023a). Neuralangelo: High-ﬁdelity neural surface

reconstruction. arXiv preprint arXiv:2306.03092.

Liu, M., Xu, C., Jin, H., Chen, L., Varma, M. T., Xu, Z., and

Su, H. (2023b). One-2-3-45: Any single image to 3d

mesh in 45 seconds without per-shape optimization.

arXiv preprint arXiv:2306.16928.

Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T.,

and Wang, W. (2023c). Syncdreamer: Generating

multiview-consistent images from a single-view im-

age. arXiv preprint arXiv:2309.03453.

Long, X., Liu, Y., Lin, C., Zeng, Z., Liu, L., Komura,

T., and Wang, W. (2023). Wonder3d: Single image

to 3d using cross-domain diffusion. arXiv preprint

arXiv:2310.15008.

Mildenhall, B., Srinivasan, P. P., Tancik, M., Barron, J. T.,

Ramamoorthi, R., and Ng, R. (2020). Nerf: Repre-

senting scenes as neural radiance ﬁelds for view syn-

thesis. In Proceedings of the European Conference on

Computer Vision, pages 405–421.

Poole, B., Jain, A., Barron, J. T., and Mildenhall, B. (2022).

Dreamfusion: Text-to-3d using 2d diffusion. arXiv

preprint arXiv:2209.14988.

Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X.,

Chen, L., Zeng, C., and Su, H. (2023). Zero123++: A

single image to consistent multi-view diffusion base

model. arXiv preprint arXiv:2310.15110.

Tochilkin, A. et al. (2023). Triposr: High-efﬁciency 3d re-

construction from minimal data inputs. arXiv preprint

arXiv:2308.12045.

Touvron, H., Bojanowski, P., Caron, M., Misra, I., Mairal,

J., and Joulin, A. (2023). Dinov2: Learning ro-

bust visual features without labels. arXiv preprint

arXiv:2304.07193.

Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., and

Zhu, J. (2023). Proliﬁcdreamer: High-ﬁdelity and di-

verse text-to-3d generation with variational score dis-

tillation. In Advances in Neural Information Process-

ing Systems (NeurIPS).

Wang, Z., Wang, Y., Chen, Y., Xiang, C., Chen, S., Yu, D.,

Li, C., Su, H., and Zhu, J. (2024). Crm: Single image

to 3d textured mesh with convolutional reconstruction

model. arXiv preprint arXiv:2403.05034.

Whizzy Studios (2023). Applications of 3d modeling.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1326