Hunyuan3D 2.

0: Scaling Diffusion Models for High

Resolution Textured 3D Assets Generation

“ Living out everyone’s imagination on creating and manipulating 3D assets.”

— Hunyuan3D Team ∗

Abstract
We present Hunyuan3D 2.0, an advanced large-scale 3D synthesis system for gen-
erating high-resolution textured 3D assets. This system includes two foundation
components: a large-scale shape generation model – Hunyuan3D-DiT, and a large-
scale texture synthesis model – Hunyuan3D-Paint. The shape generative model,
built on a scalable flow-based diffusion transformer, aims to create geometry that
properly aligns with a given condition image, laying a solid foundation for down-
stream applications. The texture synthesis model, benefiting from strong geometric
and diffusion priors, produces high-resolution and vibrant texture maps for either
generated or hand-crafted meshes. Furthermore, we build Hunyuan3D-Studio – a
versatile, user-friendly production platform that simplifies the re-creation process
of 3D assets. It allows both professional and amateur users to manipulate or even
animate their meshes efficiently. We systematically evaluate our models, show-
ing that Hunyuan3D 2.0 outperforms previous state-of-the-art models, including
the open-source models and closed-source models in geometry details, condition
alignment, texture quality, and etc. Hunyuan3D 2.0 is publicly released in order
to fill the gaps in the open-source 3D community for large-scale foundation gen-
erative models. The code and pre-trained weights of our models are available at:
https://github.com/Tencent/Hunyuan3D-2.

∗ Hunyuan3D team contributors are listed in the end of report.

 Hunyuan3D-Paint
Generated Texture Synthesis

Hunyuan3D-DiT
Generated Shape Generation

Hunyuan3D-Studio
Low Poly, Sketch to 3D, Animation…

Figure 1: An overall of Hunyuan3D 2.0 system.

2
1 Introduction

Digital 3D assets have woven themselves into the very fabric of modern life and production. In the
realms of gaming and film, these assets are vibrant expressions of creators’ imaginations, spreading
joy and crafting immersive experiences for players and audiences alike. In the fields of physical
simulation and embodied AI, 3D assets serve as essential building blocks, enabling machines and
robots to mimic and comprehend the real world. Yet, the journey of creating 3D assets is anything
but straightforward; it is often a complex, time-consuming, and costly endeavor. A typical production
pipeline may involve stages like sketch design, digital modeling, and 3D texture mapping, each
demanding high expertise and proficiency in digital content creation software. As a result, the
automated generation of high-resolution digital 3D assets has emerged as one of the most exciting
and sought-after topics in recent years.
Despite the importance of automated 3D generation and rapid development in image and video
generation fueled by the rise of diffusion models [33, 74, 24, 50, 43], the field of 3D generation
appears to be relatively stagnant in the era of large models and big data, with only a handful of works
making gradual progress [111, 118, 49]. Building on the 3DShape2Vectset [111], Michelangelo [118]
and CLAY [113] gradually enhance shape generation performance, where CLAY is the first work to
demonstrate the unprecedented potential of diffusion models in 3D asset generation. Nevertheless,
progress in the 3D domain remains limited. As evidenced in other fields [114, 4, 3], the prosperity of
a domain in the era of large models usually relies on a strong open-source foundational model, such
as Stable Diffusion [74, 69, 24] for image generation, LLaMA [90, 91, 22] for language models, and
HunyuanVideo [43] for video generation. To this end, we present Hunyuan3D 2.0, a 3D asset creation
system with two strong open-sourced 3D foundation models: Hunyuan3D-DiT for generative shape
creation and Hunyuan3D-Paint for generative texture synthesis.
Hunyuan3D 2.0 features a two-stage generation pipeline, starting with the creation of a bare mesh,
followed by the synthesis of a texture map for that mesh. This strategy is effective for decoupling
the difficulties of shape and texture generation [34, 106, 46, 47] and also provides flexibility for
texturing either generated or handcrafted meshes. With this architecture, our shape creation model
– Hunyuan3D-DiT, is designed as a large-scale flow-based diffusion model. As a prerequisite, we
first train an autoencoder – Hunyuan3D-ShapeVAE using advanced techniques such as mesh surface
importance sampling and variational token length to capture fine-grained details on the meshes.
Then, we build up a dual-single stream transformer [45] on the latent space of our VAE with the
flow-matching [53, 24] objective. Our texture generation model – Hunyuan3D-Paint is made of a
novel mesh-conditioned multi-view generation pipeline and a number of sophisticated techniques for
preprocessing and baking multi-view images into high-resolution texture maps.
We performed an in-depth comparison of Hunyuan3D 2.0 in relation to leading 3D generation models
worldwide, including three commercial closed-source end-to-end products, an end-to-end open-
sourced model Trellis [100], and several separate models [9, 37, 98, 110, 55, 59] for shape and texture
generation. We report visual and quantitative evaluation results across three dimensions: generated
textured mesh, bare mesh, and texture map. We also provided user study results on 300 test cases
involving 50 participants. The comparison shows the superiority of Hunyuan3D 2.0 in alignment
between conditional images and generated meshes, generation of fine-grained details, and human
preference ratings.

2 Hunyuan3D 2.0 Architecture

In this section, we elaborate on the model architecture of Hunyuan3D 2.0, focusing on two main
components: the shape generation model and the texture generation model. Fig. 2 illustrates the
pipeline of Hunyuan3D 2.0 for creating a high-resolution textured 3D asset. Given an input image,
Hunyuan3D-DiT initially generates a high-fidelity bare mesh via the shape generation model. This
model comprises a Hunyuan3D-ShapeVAE and a Hunyuan3D-DiT, which will be discussed in Sec. 3.
Subsequently, by leveraging strong geometric priors and the input image, we introduce Hunyuan3D-
Paint as our texture generation model in Sec. 4. This model produces self-consistent multi-view
outputs, which are used for baking high-definition texture maps.

3
Shape Generation Texture Synthesis
Hunyuan3D-ShapeVAE Hunyuan3D-Paint

···

Baking
FPS
Q
KV
Cross Attention
Single Image Super-Resolution
Uniform
···
Self Attention x8

···
Latent
Importance Tokens
Self Attention x 16
Reference
Marching
Cube Branch

Delighted
Image
Cross Attention
Grid Queries Multi-Task
Decoded Mesh Attention
···

···

Image Generation
Hunyuan3D-DiT Delighting Branch
Image

Denoising
··· Transformer

Latent Tokens Steps x N

Input Image Normal Position

Figure 2: An overall of Hunyuan3D 2.0 architecture for 3D generation. It consists of two main
components: Hunyuan3D-DiT for generating bare mesh from a given input image and Hunyuan3D-
Paint for generating a textured map for the generated bare mesh. Hunyuan3D-Paint takes geometry
conditions – normal maps and position maps of generated mesh as inputs and generates multi-view
images for texture baking.

3 Generative 3D Shape Generation

Hunyuan3D 2.0 employs the architecture of the latent diffusion model [74, 111, 118, 113] for shape
generation. This design is driven by the successful applications of latent diffusion models in image
and video generation. Specifically, our shape generation model consists of (1) an autoencoder –
Hunyuan3D-ShapeVAE (Sec. 3.1) that compresses the shape of a 3D asset represented by polygon
mesh into a sequence of continuous tokens in the latent space; (2) a flow-based diffusion model –
Hunyuan3D-DiT (Sec. 3.2), trained on the latent space of ShapeVAE for predicting object token
sequences from a user-provided image. The predicted tokens are further decoded into a polygon
mesh with VAE decoder. The details of these models are illustrated below.

3.1 Hunyuan3D-ShapeVAE

Hunyuan3D-ShapeVAE employs vector sets, a compact neural representation for 3D shapes proposed
by 3DShape2VecSet [111], which has also been leveraged in the recent work Dora [11]. Followed by
Michelangelo [118], we use a variational encoder-decoder transformer for shape compression and
decoding. Besides, we choose 3D coordinates and the normal vector of point cloud sampled from the
surface of 3D shapes as inputs for the encoder and instruct the decoder to predict the Signed Distance
Function (SDF) of the 3D shape, which can be further decoded into triangle mesh via the marching
cube algorithm. The overall network architecture is illustrated in Fig. 3.
Importance Sampled Point-Query Encoder. The encoder Es aims to extract representative features
to characterize 3D shapes. To achieve this, our first design utilizes an attention-based encoder to
encode point clouds uniformly sampled from the surface of a 3D shape. However, this design usually
fails to reconstruct the details of complex objects. We attribute this difficulty to the variations in the
complexity of regions on the shape surface. Therefore, in addition to uniformly sampled point clouds,
we designed an importance sampling method∗ that samples more points on the edges and corners of
the mesh, which provides more complete information for describing complex regions.
∗
Concurrent work [11] also proposes similar importance sampling to improve the VAE reconstruction
performance based on similar observations.

4
In detail, for an input mesh, we first collect uniformly sampled surface point clouds Pu ∈ RM ×3 ,
and importance sampled surface point clouds Pi ∈ RN ×3 . We use a layer of cross attention, to
compress the input point clouds into a set of continuous tokens via a set of point queries [111]. To
obtain point queries, we apply Farthest Point Sampling (FPS) separately to Pu and Pi to obtain the
′ ′
uniform point query Qu ∈ RM ×3 and the importance point query Qi ∈ RN ×3 . The final point
′ ′
cloud P ∈ R(M +N )×3 and point query Q ∈ R(M +N )×3 for the cross attention are constructed
by concatenating both sources. Then, we encode the point clouds P and point queries Q with
Fourier positional encoding followed by a linear projection, resulting Xp ∈ R(M +N )×d and Xq ∈
′ ′
R(M +N )×d , where d is the width of the transformer. The encoded point cloud and point query are
sent to the cross attention followed by a number of self-attention layers, which helps improve the
′ ′
feature representation, to obtain the hidden shape representation Hs ∈ R(M +N )×d . Since we adopt
the design of variational autoencoder [41], an additional linear projection is applied on Hs to predict
′ ′ ′ ′
the mean E(Zs ) ∈ R(M +N )×d0 and variance Var(Zs ) ∈ R(M +N )×d0 of the final latent shape
embedding in a token sequence, where d0 is the dimension of latent shape embedding.
Decoder. The decoder Ds reconstructs the 3D neu-
ral field from the latent shape embedding Zs from
the encoder. Ds starts from a projection layer to
transform the latent embedding from dimension d0 ···
back to the width of transformer d. Then, a number
Uniform Importance
of self-attention layers further process the hidden
···
embeddings, after which is another point perceiver
that takes 3D grid Qg ∈ R(H×W ×D)×3 as queries
to obtain 3D neural field Fg ∈ R(Fn ×W ×D)×d from FPS
Q
FPS
Grid Decoded
Queries Mesh
the hidden embeddings. We use another linear pro- KV

···
Marching
jection on the neural field to obtain the Sign Distance Q Cube
Function (SDF) Fsdf ∈ R(Fo ×W ×D)×1 , which can
Cross Attention Cross Attention
be decoded into triangle mesh with marching cube
KV
algorithms.
Self Attention x8 Self Attention x 16
Training Strategy & Implementation. We employ
multiple losses to supervise the model training, in-
···

cluding (1) the reconstruction loss that computes Latent Tokens

MSE loss between predicted SDF Ds (x|Zs ) and

ground truth SDF(x), and (2) the KL-divergence Figure 3: The overall architecture of
loss LKL to make the latent space compact and con- Hunyuan3D-ShapeVAE. Instead of only using
tinuous, which facilitates the training of diffusion uniform sampling on mesh surface, We have
models. Due to the dense computation required by developed an importance sampling strategy to
complete SDF, the reconstruction loss is calculated extract high-frequency detail information from
as the expectation of losses on randomly sampled the input mesh surface, such as edges and cor-
points in the space and shape surface. The overall ners. This allows the model to better capture
training loss Lr can be written as, and represent the intricate details of 3D shapes.
Note that during the point query construction,
Lr = Ex∈R3 [MSE(Ds (x|Zs ), SDF(x))] + γLKL the Farthest Point Sampling (FPS) operation
(1) is performed separately for the uniform point
where γ is the loss weight of KL loss. During train- cloud and the importance sampling point cloud.
ing, we also utilize a multi-resolution strategy to
speed up model convergence, where the length of the latent token sequence is randomly sampled from
a predefined set. A shorter sequence reduces the computation cost, and a longer sequence facilitates
the reconstruction quality. The longest sequence length is 3072 in our released version, which could
support high-resolution shape generation with fine-grained and sharp details.

3.2 Hunyuan3D-DiT

Hunyuan3D-DiT is a flow-based diffusion model aimed at producing high-fidelity and high-resolution

3D shapes according to given image prompts.
Network Structure. Inspired by FLUX [45], we adopt a dual- and single-stream network structure
as in Fig. 4. In dual-stream blocks, latent tokens and condition tokens are processed with separate
QKV projections, MLP, and etc., but interact within an attention operation. In single-stream blocks,

5
Hunyuan3D-DiT Double Stream Block Single Stream Block
Timestep Input Image Noisy Shape Input
y Shape Tokens Image Tokens y
Latent Tokens

PE DINO Mod LayerNorm LayerNorm Mod y Mod LayerNorm

MLP Linear Linear Scale & Shift Scale & Shift Scale & Shift

Linear Linear
Double Stream ×16
Linear Linear
Block QK-Norm QK-Norm

QK-Norm GELU
Attention
Single Stream × 32 Attention
Block Linear Linear

Gate Gate
Linear
y LayerNorm
LayerNorm LayerNorm
Gate
Mod Scale & Shift
Scale & Shift Scale & Shift
Output
Linear MLP MLP

Gate Gate
Output
Shape Output Image Output

Figure 4: Overview of Hunyuan3D-DiT. It adopts a transformer architecture with both double- and
single-stream blocks. This design benefits the interaction between modalities of shape and image,
helping our model to generate bare meshes with exceptional quality. (Note that the orange blocks
have no learnable parameters, the blue blocks contain trainable parameters, and the gray blocks
indicate a module composed of more details.)

latent tokens and condition tokens are concatenated and processed by spatial attention and channel
attention in parallel. We only use the embedding of the timestep for the modulation modules. Besides,
we omit the positional embedding of the latent sequence as the specific latent token of our ShapeVAE
in the sequence does not correspond to a fixed location in the 3D grid. Instead, the content of our 3D
latent tokens themselves is responsible for figuring out the position/occupancy of the generated shape
in the 3D grid, which is different from image/video generation where their tokens are responsible for
predicting content in a specific location in 2D/spatial-temporal grid.
Condition Injection. We employ a pre-trained image encoder to extract conditional image tokens
of the patch sequence including the head token at the last layer. To capture the fine-grained details
in the image, we utilize a large image encoder – DINOv2 Gaint [64] and large input image size –
518 × 518. Besides, we also remove the background of the input image, resize the object to a unified
size, reposition the object to the center, and fill the background with white, which helps to remove
the negative impact of the background and increase the effective resolution of the input image.
Training & Inference. We utilize flow matching objective [53, 24] for training our model. Specifi-
cally, flow matching first defines a probability density path between Gaussian distribution and data
distribution, then, the model is trained to predict the velocity field ut = xdtt that drifting sample xt
towards data x1 . In our case, we adopt the affine path with the conditional optimal transport schedule
specified in [54], where xt = (1 − t) × x0 + t × x1 , ut = x1 − x0 . Therefore, the training loss is
formulated as,
L = Et,x0 ,x1 [∥ uθ (xt , c, t) − ut ∥22 ], (2)
where t ∼ U(0, 1) and c denotes model condition. During the inference phase, we first randomly
sample a start point x0 ∼ N(0, 1) and employ a first-order Euler ordinary differential equation (ODE)
solver to solve x1 with our diffusion model uθ (xt , c, t).

4 Generative Texture Map Synthesis

Given a 3D mesh without texture and an image prompt, we aim to generate a high-resolution and
seamless texture map. The texture map should closely conform to the image prompt in the visible
region, exhibit multi-view consistency, and maintain harmonious with the input mesh.

6
Hunyuan3D-Paint
Training & Inference Pipeline Multi-Task Attention
Delighted Image Dense View Camera Embeddings Other Views Latent
Inference

Conv-In
Reference ···
Branch
Multiview Attention

ResBlock

ResBlock
Multi-Task
Attention Self Attention

Conv-Out
Image Generation
Delighting Branch Reference Attention

Generation Block
Conv-In

ResBlock
Cross
Viewpoint
Self Attention
Selection
Attention

Timestep=0 Reference Block

Input Image Normal Position Input Mesh

Figure 5: Overview of Hunyuan3D-Paint. We leverage an image delighting module to convert the

input image to an unlit state to produce light-invariant texture maps. The system features a double-
stream image conditioning reference-net, which provides faithfully conditional image features to the
model. Furthermore, it facilitates the production of texture maps that conform closely to the input
image. The multi-task attention module ensures that the model synthesizes multi-view consistent
images. This module maintains the coherence of all generated images while adhering to the input.

To achieve these objectives, we employ a three-stage framework, including a pre-processing stage

(Sec. 4.1), a multi-view image synthesis stage (Sec. 4.2, Hunyuan3D-Paint), and a texture baking
stage based on dense multi-view inference (Sec. 4.3). Other details related to model training and text-
and image-to-texture pipeline are provided in Sec. 4.4. Fig. 5 illustrates the complete pipeline of our
texture map synthesis method.

4.1 Pre-processing

Image Delighting Module. The reference image typically exhibits pronounced and varied illumi-
nation and shadow, whether collected by the user or generated by T2I models. Directly inputting
such images into the multi-view generation framework can cause illumination and shadows to be
baked into the texture maps. To address this issue, we leverage a delighting procedure on the input
image via an image-to-image approach [6] before multi-view generation. Specifically, to train such
an image delighting model, we collect a large-scale 3D dataset and render it under the illumination
of a random HDRI environmental map and an even white light to form the corresponding pair-wise
image data. Benefiting from this image delighting model, our multi-view generation model can be
fully trained on white-light illuminated images, enabling an illumination-invariant texture synthesis.
View Selection Strategy. In practical applications, to reduce the costs of texture generation (i.e.,
generate the largest area of texture with the minimum number of viewpoints), we employ a geometry-
aware viewpoint selection strategy to support effective texture synthesis. By considering the coverage
of the geometric surface, we heuristically select 8 to 12 viewpoints for inference. Initially, we fix
4 orthogonal viewpoints as basis since they cover most parts of the geometry. Subsequently, we
iteratively add novel viewpoints using a greedy search approach. The specific process is illustrated in
Algorithm 1, and the coverage function in the algorithm is defined as:
( " !#)
[
F(vi , Vs , M) = Aarea UV cover (vi , M) \ UV cover (vi , M) ∩ UV cover (vs , M) (3)
s∈Vs

where UV cover (v, M) is a function that returns the set of covering texels in UV space based on the
input view v and mesh geometry M, and Aarea (· · · ) is a function that calculates the coverage area
according to the given set of covering texels. This approach encourages the multi-view generation

7
model to focus on viewpoints with more unseen regions, together with the dense-view inference,
alleviating the burden of post-processing (i.e., texture inpainting).

Algorithm 1 View Selection Algorithm

1: Input:
2: Initialize a selected viewpoint set Vs , a reference viewpoint set Vr , a coverage function F,
and a mesh M.
3: Set Nmax = 12 as the maximum number of iterative searches, and Nf ixed = 4 as the initial
fixed number of viewpoints.
4: for i = Nf ixed to Nmax − 1 do
5: Set cmax = −1
6: Set vmax = −1
7: for vi ∈ Vr do
8: Set ccur = F(vi , Vs , M)
9: if ccur > cmax then
10: cmax = ccur
11: vmax = vi
12: end if
13: end for
14: Vr ← Vr \ {vmax }
15: Vs ← Vs ∪ {vmax }
16: end for

4.2 Hunyuan3D-Paint

Geometry-conditioned multi-view image generation is a key component in the texture synthesis

framework. In the context of image-guided texture synthesis, the design of the multi-view image
generation must be meticulously crafted to achieve image alignment, geometric following, and
multi-view consistency. These functionalities are realized in Hunyuan3D-Paint according to a bunch
of techniques containing a double-stream image conditioning reference-net, a multi-task attention
mechanism, and a strategy for geometry and view conditioning.
Double-stream Image Conditioning Reference-Net. In the aim of the image following, Hunyuan3D-
Paint implements a reference-net conditioning approach [115], upon which we develop a series of
alternative solutions. Specifically, rather than a noisy feature synchronized with the generation
branch, we directly feed the original VAE feature of the reference image into the reference branch
to maintain image details as much as possible. Since the feature is noiseless, we set the timestep of
the reference branch to 0 to maintain the input image information faithfully. On the other hand, to
regularize potential style bias introduced by the 3D rendering dataset, we abandon the shared-weight
reference-net as in [88, 78] and freeze the weights of the original SD2.1 weights (which serves as the
base model for our multi-view generation). We have found that the fixed-weights reference-net serves
as a soft regularization that anchors the generated image distribution, preventing it from drifting
away towards the rendered image distribution, significantly improving the performance on real-world
image conditioning. Together, these two schemes form the double-stream image conditioning strategy.
We leverage the zero-noised double-stream image conditioning reference-net by capturing a feature
cache prior to each self-attention module. This cache is then fed into the multi-view diffusion model
via a reference attention module.
Multi-task Attention Mechanism. We introduce two additional attention modules alongside the
original self-attention to enable the image diffusion model to generate multi-view images guided
by a reference image. The reference attention module integrates the reference image into the multi-
view diffusion process. In contrast, the multi-view attention module ensures consistency across the
generated views. To mitigate potential conflicts arising from these multi-functionalities, we design
the additional two distinct attention modules in a parallel structure (illustrated in Fig. 5), which can
be expressed as:
!
T T
Qref Kref
 
Qmv Kmv
ZM V A = ZSA + λref · Softmax √ Vref + λmv · Softmax √ Vmv (4)
d d

8
where ZSA represents the feature calculated by the original frozen-weight self-attention, and
Qref , Kref , Vref and Qmv , Kmv , Vmv are the Query, Key, and Value projected features of reference
attention and multi-view attention, respectively.
Geometry and View Conditioning. Following geometry is another unique feature in texture map
synthesis. To enable effective training, we opt for an easy implementation of directly concatenating
the geometry conditions with noise. Specifically, we first input the multi-view canonical normal maps
and canonical coordinate maps (CCM)—two view-invariant geometry conditions we utilize—into
a pre-trained Variational Autoencoder (VAE) to obtain geometric features. These features are then
concatenated with latent noise and fed into the channel-extended input convolution layer of the
diffusion model.
Other than geometry conditioning, we adopt a learnable camera embedding in our pipeline to boost
the viewpoint clue for the multi-view diffusion model. Specifically, we assign a unique unsigned
integer to each pre-defined viewpoint and set up a learnable view embedding layer to map the integer
to a feature vector, which is then injected into the multi-view Diffusion model. We have found in our
experiments that combining the geometry conditioning with a learnable camera embedding yields the
best performance.

4.3 Texture Baking

Dense-view inference. Potential self-occlusion is a significant challenge in the context of texture

synthesis within a multi-view image generation framework, particularly when handling irregular
geometries produced by shape generative models. This issue necessitates modeling texture synthesis
as a two-stage framework. The first stage focuses on multi-view image generation, while the second
stage involves non-trivial inpainting to fill the holes caused by self-occlusion. In our texture-map
synthesis framework, we alleviate the burden on the second stage of inpainting by facilitating dense-
view inference during the multi-view generation stage. To enable effective dense-view inference, a
view dropout strategy is introduced as a flexible training mechanism that allows the model to encounter
all of the pre-set viewpoints, thereby enhancing its 3D perception capabilities and generalization.
Specifically, we randomly select 6 viewpoints from a total of 44 pre-set viewpoints to serve as a batch
input to the multi-view diffusion backbone network. During the inference phase, our framework is
able to output the images of any specified viewpoints, supporting dense-view inference.
Single Image Super-resolution. To enhance texture quality, we apply a pre-trained single-image
super-resolution model [95] to each generated image from different viewpoints. Experiments have
demonstrated that this single-image super-resolution approach maintains consistency among multi-
views, as it does not introduce significant variations to the images.
Texture Inpainting. After unwrapping the synthesized dense multi-view images into a texture map,
a small set of patches in the UV texture remain that are not fully covered. To address this issue,
we employ an intuitive inpainting approach. First, we project the existing UV texture into vertex
texture. Then, we query each UV texel’s texture by computing a weighted sum of the textures from
the connected, textured vertices. The weights are set to be the reciprocal of the geometric distances
between the texels and the vertices.

4.4 Implementation Details

Model Training. For training the multi-view image generation framework, we start by inheriting
the ZSNR checkpoint of the Stable Diffusion 2 v-model [52]. We train our multi-view diffusion
model using a self-collected large-scale 3D dataset. Multi-view images are rendered under the
illumination of an even white light to accommodate our delighting model. Specifically, we render the
reference image with a random azimuth and a fixed range of elevation from -20 to 20 degrees. This
variation disrupts the consistency between the reference and generated images, thereby increasing the
robustness of our texture generation framework. We directly train on 512 × 512 resolution with a
total of 80,000 steps, a batch size of 48, and a learning rate of 5 × 10−5 . We use 1000 warm-up steps
and the "trailing" scheduler proposed by ZSNR.
General Text- and Image-to-Texture. It is worth noting that Hunyuan3D-Paint not only generates
high-quality texture maps for generated meshes but also supports arbitrary texture generation guided
by any text or image input provided by the user for any geometric model. To achieve this, we leverage

9
 3DShape2VecSet [111] Michelangelo [118] Direct3D [98] Hunyuan3D-ShapeVAE (Ours)
V-IoU(↑) 87.88% 84.93% 88.43% 93.6%
S-IoU(↑) 80.66% 76.27% 81.55% 89.16%
Table 1: Numerical comparisons. We evaluate the reconstruction performance of Hunyuan3D-
ShapeVAE and baselines based on volume IoU (V-IoU) and Surface (S-IoU). The results indicate
Hunyuan3D-ShapeVAE overwhelms all baselines in the reconstruction performance.

advanced T2I models and corresponding conditional generation modules, such as ControlNet [114]
and IP-Adapter [107], to generate input images that align with geometric shapes based on user-
provided text or image prompts. Benefitting from this paradigm, we are capable of texturing any
specified geometry with arbitrary images, whether they are matched or mismatched. An application of
using different images to texture the same geometry, dubbed as re-skinning, is illustrated in Fig. 9.

5 Evaluations

To thoroughly evaluate the performance of Hunyuan3D 2.0, we conducted experiments from three
perspectives: (1) 3D Shape Generation (including Shape Reconstruction and Shape Generation), (2)
Texture map synthesis, and (3) Textured 3D assets generation.

5.1 3D Shape Generation

Shape generation is crucial for 3D generation, as high-fidelity and high-resolution bare meshes
form the foundation for downstream tasks. In this section, we compare and evaluate the capability
of 3D shape generation in Hunyuan3D 2.0 from two perspectives: shape reconstruction and shape
generation.
Baselines. We compare the reconstruction performance of Hunyuan3D-ShapeVAE with
3DShape2VecSet [111], Michelangelo [118], and Direct3D [98]. The mentioned methods rep-
resent the state-of-the-art ShapeVAE architecture, and the core differences are neural representations,
where 3DShape2VecSet uses a downsampled vector set, point query; Michelangelo utilizes a learnable
vector set, learnable query; Direct3D leverages learnable triplane; Hunyuan3D-ShapeVAE employees
point query with importance sampling. Note that, except Direct3D, which requires a 3072 token
length (suffering significant performance degeneration when reducing token length), all VAE models
compare by 1024 token length. Hunyuan3D-DiT is compared with several state-of-the-art baselines.
Open-source baselines are Michelangelo [118], Craftsman 1.5 [49], and Trellis [100]. Closed-source
baselines are Shape Model 1, Shape Model 2, and Shape Model 3.
Metrics. We employ the Intersection of Union (IoU) to measure the reconstruction performance.
Specifically, we compute randomly sampled volume points IoU (V-IoU) and near-surface region IoU
(S-IoU) to reflect the reconstruction performance comprehensively. To evaluate shape generative
performance, we employ ULIP [104] and Uni3D [121] to compute the similarity between the
generated mesh and input images (ULIP-I and Uni3D-I) and the similarity between the generated
mesh and images prompt synthesizing by the vision language model [13] (ULIP-T and Uni3D-T).
Shape Reconstruction Comparisons. The Numerical comparison of shape reconstruction is shown
in Tab. 1. According to the table, Hunyuan3D-ShapeVAE overwhelms all baselines. Comparisons
among Hunyuan3D-ShapeVAE, 3DShape2VecSet, and Michelangelo demonstrate the effectiveness
of the importance sampling strategies. Fig. 6 illustrates the visual comparison of shape reconstruction,
which shows that Hunyuan3D-ShapeVAE could faithfully recover the shape with fine-grained details
and produce neat space without any floaters.
Shape Generation Comparisons. Tab. 2 shows the numerical comparison between Hunyuan3D-DiT
and competing methods, which indicates that Hunyuan3D-DiT produces the most condition following
results. Furthermore, according to the visual comparison in Fig. 7, results from Hunyuan3D-DiT
follow the image prompt most, including clear human faces, surface bumps, logo texts, and layouts.
Meanwhile, the generated bare mesh is holeless, which supports a solid basis for downstream tasks.

10
 Ground Truth Michelangelo 3DShape2Vectset Direct3D Hunyuan3D-ShapeVAE

Figure 6: Visual comparisons. We illustrate the reconstructed mesh (blue paint aims to show more
details) in the figure, which showcases that only Hunyuan3D-ShapeVAE reconstructs mesh with
fine-grained surface details and neat space. (Better viewed by zooming in.)

11
 ULIP-T(↑) ULIP-I(↑) Uni3D-T(↑) Uni3D-I(↑)
Michelangelo [118] 0.0752 0.1152 0.2133 0.2611
Craftsman 1.5 [49] 0.0745 0.1296 0.2375 0.2987
Trellis [100] 0.0769 0.1267 0.2496 0.3116
Shape Model 1 0.0799 0.1181 0.2469 0.3064
Shape Model 2 0.0741 0.1308 0.2464 0.3106
Shape Model 3 0.0746 0.1284 0.2516 0.3131
Hunyuan3D-DiT (Ours) 0.0771 0.1303 0.2519 0.3151

Table 2: Numerical comparisons. By evaluating the shape generation performance on ULIP-T/I,

Uni3D-T/I, demostrating Hunyuan3D-DiT could produce the most condition followed results.

Input Michelangelo Craftsman Trellis Shape Shape Shape Hunyuan3D-

1.5 Model 1 Model 2 Model 3 DiT

Figure 7: Visual comparisons. We display the input image and the generated bare mesh (blue paint
aims to show more details) from all methods in the figure. The human faces and piano keys show
that Hunyuan3D-DiT could synthesize detailed surface bumps, maintaining completeness. Several
scenes or logos demonstrate that Hunyuan3D-DiT could generate intricate details. (Better viewed by
zooming in.)

12
 CMMD(↓) FIDCLIP (↓) CLIP-score(↑) LPIPS(↓)
TEXTure [73] 3.047 35.75 0.8499 0.0076
Text2Tex [9] 2.811 31.72 0.8680 0.0071
SyncMVD [59] 2.584 29.93 0.8751 0.0063
Paint3D [110] 2.810 30.29 0.8724 0.0063
TexPainter [112] 2.483 28.83 0.8789 0.0062
Hunyuan3D-Paint (Ours) 2.318 26.44 0.8893 0.0059

Table 3: Numerical comparisons. We compare Hunyuan3D-Paint with baselines on various metrics,

and the results indicate that our model could produce the most condition-conforming texture maps.

5.2 Texture Map Synthesis

As texture maps directly influence the visual appeal of textured 3D assets, we conduct comprehensive
text-conditioned texture map synthesis experiments to validate the performance of Hunyuan3D-Paint.
Baselines. We compare Hunyuan3D-Paint with the following texture generation methods, including
TEXTure [73], Text2Tex [9], SyncMVD [59], Paint3D [110], and TexPainter [112]. All the baselines
leverage geometric and diffusion priors to facilitate the overall generation quality of texture maps.
Metrics. We apply several frequently used image-level metrics to enable a fair comparison of
texture map generation. Specifically, we leverage a CLIP-version of Fréchet Inception Distance
F IDCLIP to compute the distance between the rendering of the generated textured map in semantic
perspectives. We use the implementation of Clean-FID [66]. Besides, the recently introduced CLIP
Maximum-Mean Discrepancy (CMMD) [38] is utilized to serve as another important criterion, which
is a more accurate measurement of images with rich details. In addition to these two metrics, we also
use CLIP-score [71] to validate semantic alignment between renderings of the generated texture map
and given prompt and LPIPS [116] to estimate the consistency between renderings of the generated
texture map and ground-truth images.
Comparisons. The numerical comparison of text-to-texturing is shown in Tab. 3, showcasing that
Hunyuan3D-Paint achieves the best generative quality and semantic following. The visual comparison
refers to Fig. 8. The fish and rabbit show that our model produces the most condition-following
results. And the football demonstrates the ability of Hunyuan3D-Paint to produce clear texture maps.
The texture map of the castle and bear contains rich texture patterns, showcasing that our model can
produce intricate details.
Applications. All generated texture maps are seamless and lighting-invariant. Moreover, Hunyuan3D-
Paint is flexible to produce various texture maps for bare mesh or hand-crafted mesh according to
different prompts. As shown in the Fig. 9, our model produces different texture maps for a mesh with
seamless and intricate details.

5.3 Textured 3D Assets Generation

In this section, we evaluate the generated textured 3D assets for reflecting the end-to-end generation
capabilities of Hunyuan3D 2.0.
Baselines. We compare Hunyuan3D 2.0 against leading models in the field, including open-source
model Trellis [100] and closed-source models Model 1, Model 2, and Model 3.
Metrics. We mainly measure the generative quality of textured 3D assets by their renderings. Similar
to Sec. 5.2, we employ F IDCLIP to compute the image content distance, CLIP-score to reflect
semantic alignment, CMMD to measure the similarity in the image details, and LPIPS to evaluate the
consistency between rendering from generated textured 3D assets and given image prompts.
Comparisons. The numerical results reported in the Tab. 4 indicate that Hunyuan3D 2.0 surpasses
all baselines in the quality of generated textured 3D assets and the condition following ability. The
illustration in the Fig. 11 demonstrates that Hunyuan3D 2.0 produces the textured 3D assets with the
highest quality. Even for the text in the image prompt, our model can produce the correct bumps on
the shape surface and an accurate texture map according to the geometric conditions. The rest of the
cases demonstrate the ability of our model to generate high-resolution and high-fidelity results with
complex actions or scenes.

13
 A fish with orange and pink scales

A bunny crafted from colorful, hand-painted ceramic tiles

A futuristic dragon with sleek, metallic blue and silver surfaces

A photo of a soccer ball showing a pentagon-shaped black section surrounded by white panels

A castle with a central tower and four turrets, featuring a mix of dark and light stone textures

A teddy bear wearing a striped scarf, standing on a wooden base with grass trim

TEXTure Text2Tex SyncMVD Paint3D TexPainter Ours

Figure 8: Visual comparisons. We demonstrate several generated texture maps on different bare
meshes. The fish and rabbit texture map showcases that Hunyuan3D-Paint produces the most text-
conforming results. The football indicates that our model could synthesize seamless and clean texture
maps. Moreover, Hunyuan3D-Paint could generate complex texture maps, like the castle and bear.
(Better viewed by zooming in.)
14
Figure 9: Visual results. We generate different texture maps for two meshes, and the results validate
the performance of Hunyuan3D-Paint on texture reskinning. (Better viewed by zooming in.)

CMMD(↓) FIDCLIP (↓) FIDIncept (↓) CLIP-score(↑)

Trellis [100] 3.591 54.639 289.287 0.787
Model 1 3.600 55.866 305.922 0.779
Model 2 3.368 49.744 294.628 0.806
Model 3 3.218 51.574 295.691 0.799
Hunyuan3D 2.0 (Ours) 3.193 49.165 282.429 0.809

Table 4: Numerical comparison. According to the results, Hunyuan3D-Paint produces the most
condition-following texture maps.

User Study. In addition, we conducted a user study by randomly inviting 50 volunteers to evaluate
300 unselected results generated by Hunyuan3D 2.0 subjectively. The evaluation criteria included 1)
overall visual quality, 2) adherence to image conditions, and 3) overall satisfaction (dissatisfaction
in either 1 or 2 results in overall dissatisfaction). The user study results in Fig. 10 indicate that
Hunyuan3D 2.0 outperforms comparative methods, particularly in its ability to adhere to image
conditions.

User Study
Method
80 trellis
Model 1
Model 2
70
Model 3
Hunyuan3D 2.0
60
Percentage (%)

50

40

30

20

10

0
Overall Satisfy 3D Assets Quality Image Following

Figure 10: The results of user study.

15
 Trellis Model 1 Model 2 Model 3 Ours

Figure 11: Visual comparisons. The first case reflects that Hunyuan3D 2.0 could synthesize detailed
surface bumps and correct texture maps. The second penguin showcases our model’s ability to handle
complex actions. The last mountain demonstrates that Hunyuan3D-DiT could produce intricate
structures, and Hunyuan3D-Paint can synthesize vivid texture maps. (Better viewed by zooming in.)

16
6 Hunyuan3D-Studio

We have developed Hunyuan3D-Studio. This platform includes a comprehensive set of tools for the
3D production pipeline, as illustrated in Fig. 1. Hunyuan3D-Studio aims to provide experts and
novices with a no-frills way to engage in 3D generation production and research. In this section, we
highlight several features of Hunyuan3D-Studio, including Sketch-to-3D, Low-polygon Stylization,
and Autonomous Character Animator. These features aim to streamline the 3D creation process and
make it accessible to a broader audience.

6.1 Sketch-to-3D

In game development and content creation, converting 2D sketches into 3D assets is a crucial
technology that significantly enhances digital artistry design efficiency and flexibility. Previous
methods [1, 120, 117, 32] suffer from the lack of the generative foundation model. They tackle this
by training a small-scale generative or reconstruction model with sketch images as input directly on a
limited dataset, significantly limiting the model’s capabilities.
Benefitting from Hunyuan3D 2.0, we could convert sketches to images with rich details as input
to the foundation 3D generative model. Specifically, the Hunyuan3D-Studio has developed the
Sketch-to-3D module, which first converts sketches to images with rich details, maintaining original
contours. Then, synthesize high-resolution and high-fidelity textured 3D assets, significantly reducing
the barrier for users to engage in content creation.
As shown in Fig. 1, the Sketch-to-3D module can generate highly detailed and realistic 3D assets
while maintaining close consistency with the original sketches. With this technology, users can
synthesize 3D content with a simple sketch, providing a powerful tool for game developers and digital
artists and a low-barrier creation platform for ordinary users.

6.2 Low-polygon Stylization

Low-polygon stylization is critical in many computer graphics (CG) pipelines, as the face count of a
mesh significantly impacts the application of 3D assets. Low-polygon stylization can significantly
reduce computational costs, making it an essential process in 3D asset management. To address this,
we have established a low-polygon stylization module that efficiently converts the dense meshes
generated by Hunyuan3D 2.0 into low-polygon meshes. This module operates in two steps: geometric
editing and texture preserving.
For geometric editing, we employ a faster and more robust traditional method [28, 35], despite
the recent auto-regressive transformer-based polygon-generation approaches [96, 103, 86, 12]. By
setting an optimization criterion, we merge the vertices of the mesh to transform the dense mesh
into a low-polygon mesh. As shown at the top-right of Fig. 1, each 3D model can be represented by
only dozens of triangles after geometric editing. The change in the face count of the mesh causes
significant deviations in the vertices and faces of the low-polygon mesh compared to the dense
mesh. Therefore, to preserve the texture patterns of the textured 3D assets, we construct a KD-tree
for the input dense mesh. We then use the nearest-neighbor search within the KD tree to query
the texture colors for the vertices of the low-polygon mesh. Finally, we obtain the texture map for
the low-polygon mesh by performing texture baking on the low-polygon mesh with vertex colors.
This process ensures that the visual quality of the textures is maintained while optimizing the mesh
structure for production-level textured 3D assets.

6.3 3D Character Animation

Hunyuan3D 2.0 generates static 3D assets with high-resolution shapes and texture maps. However,
drivable 3D models yet have broad requirements [82, 81, 84, 83, 57], such as game development
and animation production. To extend the range of applications of Hunyuan3D 2.0, we develop a 3D
character animation function in Hunyuan3D-Studio. The animation algorithm inputs the generated
character and extracts features from mesh vertices and edges. Then, we utilize the Graph Neural
Network (GNN) to detect skeleton key points and assign skinning weights to the mesh surface.
Finally, based on the predicted skeleton skinning and motion templates, the algorithm utilizes motion

17
retargeting to drive the character. Some frames are displayed in Fig. 1. With 3D character animation,
the generated results from Hunyuan3D 2.0 can come to life.

7 Related Work

7.1 3D Shape Generation

Representations. The field of shape generation has undergone significant advancements, driven
by the unique challenges associated with the 3D modality. Unlike other modalities, 3D data lacks
a universal storage and representation format, leading to diverse approaches in shape generation
research. The primary 3D representations include voxels, point clouds, polygon meshes, and implicit
functions. With the advent of seminal works such as 3D ShapeNets [99], IF-Net [14], and 3D
Gaussian Splatting [40], implicit functions, and 3D Gaussian Splitting have become prevalent in
shape generation. However, even lightweight and flexible representations like implicit functions
impose substantial modeling and computational burdens on deep neural networks. As a result,
neural representations of 3D shapes have emerged as a new research focus, aiming to enhance
efficiency. Pioneering methods such as 3DShape2VecSet [111], Michelangelo [118], CLAY [113],
and Dora [11] represent 3D shapes using vector sets (one-dimensional latent token sequences proposed
by 3DShape2VecSet [111]), significantly improving representation efficiency. Another approach
involves structured representations (e.g., triplane [68, 7, 26] or sparse volume [63, 119, 72]) to encode
3D shapes, which better preserve spatial priors but are less efficient than vector sets. Inspired by
recent advances in Latent Diffusion Models, Hunyuan3D 2.0 employs vector sets to represent 3D
shapes’ implicit functions, alleviating the compression and fitting demands on neural networks and
achieving a breakthrough in shape generation performance.
Shape Generative models. The evolution of generative model paradigms has continually influenced
shape generation research. Early works [97, 76, 105, 109] based on Variational Auto-encoder [41],
generative adversarial networks (GANs) [29], normalizing flow [65], and auto-regressive model-
ing [31] demonstrated strong generative capabilities within several specific categories. The success
of diffusion models [33] and their variants [58, 53] in text-conditioned image [74, 45] generation
has spurred the popularity of diffusion-based shape generative models, with notable works [16, 118]
achieving stable cross-category shape generation. Additionally, advancements in network architec-
tures have propelled shape-generation research. The transition from early 3D Convolutional Neural
Networks [30] to the now-common Transformer architectures [93, 67] has led to the development
of classic shape generation networks, enhancing performance. Building on these advancements,
Hunyuan3D 2.0 employs a flow-based scalable transformer, further improving the model’s shape
generation capabilities.
Large-scale Dataset. Large-scale datasets are the cornerstone of scaling laws. However, the scale
of 3D data is much smaller than that in large language models and image generation fields. From
3Dscanrep [18, 44, 92] to ShapeNet [8], the growth of 3D datasets has been gradual [85, 122,
25, 77, 102, 21, 17]. The release of objaverse [20] and objaverse-xl [19] has been a significant
driver in realizing the scaling law for shape generation. Leveraging these open-source 3D datasets,
Hunyuan3D-2.0 can generate high-fidelity and high-resolution 3D assets.
Benefiting from these open-source algorithms and 3D datasets, Hunyuan3D 2.0 is capable of gen-
erating high-fidelity and high-resolution 3D assets. Therefore, we have released Hunyuan3D 2.0
to contribute to the open-source 3D generation community and further advanced 3D generation
algorithms.

7.2 Texture Map Synthesis

High-quality texture-map synthesis has been a long-standing topic within the computer graphics
community. Its significance has only increased with the growing demand for end-to-end 3D generation
techniques, where it plays a crucial role for appearance modeling.
Text to Texture. Given a plain mesh, text/image to texture aims to generate a high-quality texture
that aligns well with the given geometry according to a guided text and image. Early attempts tried to
approach texture synthesis by harnessing the categorical information and train a generative model
on a specified dataset [10, 5, 23, 80, 26, 27]. While achieving plausible texturing results, these

18
methods failed to generalize to objects of other categories, limiting their applicability in production
environments.
More recently, Stable Diffusion [74], owing to its impressive text-guided image generation capability
and flexible structure, has spawned a plethora of text-to-texture research. To take full advantage of
pre-trained image diffusion models, most subsequent works have approached the texture synthesis
problem as a geometry-conditioned multi-view images generation problem.
Initially, score distillation was adopted to harness the generation power of image diffusion models
for 3D content (texture) synthesis [87, 51, 62, 70]. However, these methods are often limited by the
over-saturated colors and misalignment with geometry.
Subsequently, optimization-free approaches pioneered by TEXTure [73] have been introduced [101,
15, 56, 112, 110, 59, 9]. To ensure consistency across multi-view images, these methods either
adopt an inpainting framework by specifying viewpoint-related masks or employ a "synchronizing"
operation during the denoising process. However, since Stable Diffusion is trained on a dataset with
a noticeable forward-facing viewpoint bias [55], these training-free methods are limited and often
suffer from severe performance issues, such as the Janus problem and multi-view inconsistency,
which result in textures with significant artifacts.
With the development of extensive 3D datasets, training multi-view diffusion models has become a
prevailing direction for texture generation [61, 2], exhibiting more powerful capabilities on texture
consistency than the training-free approaches.
Image to Texture. In a related direction, image-guided texture generation has garnered attention
in recent months, aligning closely with our research focus. This relatively unexplored area of
image-guided texture synthesis demonstrates significant potential for further development since
images provide more diverse information than text prompts, and text-to-texture generation can
be fully replaced by a text-to-image and image-to-texture pipeline. Unfortunately, most of the
existing works focus on semantic alignment with the reference image rather than precise alignment.
FlexiTex [39] and EASI-Tex [42] both utilize an IP-Adapter [107] for image prompt injection. While
TextureDreamer [108] employs a DreamBooth-like [75] approach to facilitate texture transfer across
different objects.
However, we argue that there are two explicit advantages to exactly following every detail of the
reference image. First, as part of an end-to-end image-guided 3D generation process, the geometry
generated in the first stage strives to align with the reference image, while the appearance details are
left for texture synthesis stage. Thus, one of the main objectives of our texture generation framework
is to enhance the geometry with more detailed appearance features from the well-aligned reference
image. Second, with the rapid development of image diffusion techniques, more exquisite reference
images are now available. Carefully adhering to these details can significantly improve the quality
of the generated textures. Based on these advantages, Hunyuan3D-Paint is designed with a detailed
preserving image injection module according to the philosophy of aligning the reference image not
only semantically but also following the details as closely as possible.
Multi-view Images Generation. Due to the viewpoint bias and multi-view inconsistency inherent in
training-free image diffusion models, multi-view image diffusion was developed to alleviate these
issues by utilizing large-scale 3D datasets, such as objaverse and objaverse-xl [20, 19].
Most works force the multi-view generated latents to communicate with each other by manipulat-
ing the self-attention layers with 3D-aware masks [36, 48, 88, 60, 94, 79, 78, 55]. For example,
Zero123++ [78] first treats the multi-view attention as a self-attention on a large image, which is the
spatial concatenation of six multi-view images. MVDiffusion [89] applies a correspondence-aware
attention (CAA) to inform the model to focus only on the correlation among the spatially-close pixels.
MVAdapter [36], following Era3D [48] implements a simpler but effective row-wise and column-wise
attention to alleviate the computational burden of CAA and achieves comparable performance.
Inspired by these works, we propose a multi-view generation framework equipped with a multi-task
attention mechanism to achieve both multi-view consistency and image alignment simultaneously.
Benefiting from this careful design and being trained on a large 3D rendering dataset, Hunyuan3D-
Paint is able to achieve high-quality, consistent textures with strong alignment to the reference
image.

19
8 Conclusion
In this report, we introduce an open-source 3D creation system—Hunyuan3D 2.0 —for generating
textured meshes from images. We present Hunyuan3D-ShapeVAE, which is trained using a novel
importance sampling method. This approach compresses each 3D object into a few latent tokens
while minimizing reconstruction losses. Building on our VAE, we developed Hunyuan3D-DiT, an
advanced diffusion transformer capable of generating visually appealing shapes that align precisely
with input images. Besides, we introduce Hunyuan3D-Paint, another diffusion model designed to
create textures for both our generated meshes and user-crafted meshes. With several innovative
designs, our texture generation model, in conjunction with our shape generation model, can produce
high-resolution, high-fidelity textured 3D assets from a single image. As we continue to make
progress, we hope that Hunyuan3D 2.0 will serve as a robust baseline for large-scale 3D foundation
models within the open-source community and facilitate future research endeavors.

20
9 Contributors
• Project Sponsors: Jie Jiang, Yuhong Liu, Di Wang, Yong Yang, Tian Liu
• Project Leaders: Chunchao Guo, Jingwei Huang
• Core Contributors:
– Data: Lifu Wang, Jihong Zhang, Meng Chen, Liang Dong, Yiwen Jia, Yulin Cai, Jiaao Yu,
Yixuan Tang, Hao Zhang, Zheng Ye, Peng He, Runzhou Wu, Chao Zhang, Yonghao Tan
– Shape Generation: Zeqiang Lai, Qingxiang Lin, Yunfei Zhao, Haolin Liu, Zibo Zhao
– Texture Synthesis: Shuhui Yang, Yifei Feng, Mingxin Yang, Sheng Zhang
– Downstream Tasks: Xianghui Yang, Huiwen Shi, Sicong Liu, Junta Wu, Yihang Lian, Fan
Yang, Ruining Tang, Zebin He, Xinzhou Wang, Jian Liu, Xuhui Zuo
– Studio: Zhuo Chen, Biwen Lei, Haohan Weng, Jing Xu, Yiling Zhu, Xinhai Liu, Lixin Xu,
Changrong Hu, Tianyu Huang
• Contributors: Jie Xiao, Yangyu Tao, Jianchen Zhu, Jinbao Xue, Kai Liu, Chongqing Zhao,
Xinming Wu, Zhichao Hu, Lei Qin, Jianbing Peng, Zhan Li, Minghui Chen, Xipeng Zhang, Lin Niu,
Paige Wang, Yingkai Wang, Haozhao Kuang, Zhongyi Fan, Xu Zheng, Weihao Zhuang, YingPing
He

21
References
[1] Hmrishav Bandyopadhyay, Subhadeep Koley, Ayan Das, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath
Chowdhury, Tao Xiang, and Yi-Zhe Song. Doodle your 3d: From abstract freehand sketches to precise
3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,
pages 9795–9805, 2024.
[2] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and
Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects. arXiv preprint
arXiv:2407.02430, 2024.
[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang
Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science.
https://cdn. openai. com/papers/dall-e-3. pdf, 2(3):8, 2023.
[4] Xiao Bi, Deli Chen, Guanting Chen, Shanhuang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai
Dong, Qiushi Du, Zhe Fu, et al. Deepseek llm: Scaling open-source language models with longtermism.
arXiv preprint arXiv:2401.02954, 2024.
[5] Alexey Bokhovkin, Shubham Tulsiani, and Angela Dai. Mesh2tex: Generating mesh textures from image
queries. In IEEE International Conference on Computer Vision (ICCV), October 2023.
[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing
instructions. arXiv preprint arXiv:2211.09800, 2022.
[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo,
Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative
adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern
recognition, pages 16123–16133, 2022.
[8] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio
Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository.
arXiv preprint arXiv:1512.03012, 2015.
[9] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex:
Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International
Conference on Computer Vision, pages 18558–18568, 2023.
[10] Qimin Chen, Zhiqin Chen, Hang Zhou, and Hao Zhang. Shaddr: Interactive example-based geometry
and texture generation via 3d shape detailization and differentiable rendering. In SIGGRAPH Asia 2023
Conference Papers, 2023.
[11] Rui Chen, Jianfeng Zhang, Yixun Liang, Guan Luo, Weiyu Li, Jiarui Liu, Xiu Li, Xiaoxiao Long, Jiashi
Feng, and Ping Tan. Dora: Sampling and benchmarking for 3d shape variational auto-encoders. arXiv
preprint arXiv:2412.17808, 2024.
[12] Yiwen Chen, Yikai Wang, Yihao Luo, Zhengyi Wang, Zilong Chen, Jun Zhu, Chi Zhang, and Guosheng
Lin. Meshanything v2: Artist-created mesh generation with adjacent mesh tokenization. arXiv preprint
arXiv:2408.02555, 2024.
[13] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang,
Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic
visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition, pages 24185–24198, 2024.
[14] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of
the IEEE/CVF conference on computer vision and pattern recognition, pages 5939–5948, 2019.
[15] Wei Cheng, Juncheng Mu, Xianfang Zeng, Xin Chen, Anqi Pang, Chi Zhang, Zhibin Wang, Bin Fu, Gang
Yu, Ziwei Liu, et al. Mvpaint: Synchronized multi-view diffusion for painting anything 3d. arXiv preprint
arXiv:2411.02336, 2024.
[16] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion:
Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition, pages 4456–4465, 2023.
[17] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang,
Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik.
Abo: Dataset and benchmarks for real-world 3d object understanding. CVPR, 2022.

22
[18] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images.
In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages
303–312, 1996.

[19] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan
Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl
Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objaverse-xl: A universe
of 10m+ 3d objects. arXiv preprint arXiv:2307.05663, 2023.

[20] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig
Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d
objects. arXiv preprint arXiv:2212.08051, 2022.

[21] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann,
Thomas B. McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d
scanned household items, 2022.

[22] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha
Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv
preprint arXiv:2407.21783, 2024.

[23] Aysegul Dundar, Jun Gao, Andrew Tao, and Bryan Catanzaro. Fine detailed texture learning for 3d
meshes with generative models. IEEE Trans. Pattern Anal. Mach. Intell., 2023.

[24] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi,
Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution
image synthesis. In Forty-first International Conference on Machine Learning, 2024.

[25] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue
Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In
Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10933–10942, 2021.

[26] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic,
and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images.
Advances In Neural Information Processing Systems, 35:31841–31854, 2022.

[27] Lin Gao, Tong Wu, Yu-Jie Yuan, Ming-Xian Lin, Yu-Kun Lai, and Hao Zhang. Tm-net: Deep generative
networks for textured meshes. ACM Trans. Graph., 40(6):1–15, 2021.

[28] Michael Garland and Paul S Heckbert. Surface simplification using quadric error metrics. In Proceedings
of the 24th annual conference on Computer graphics and interactive techniques, pages 209–216, 1997.

[29] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron
Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing
systems, 27, 2014.

[30] Ben Graham. Sparse 3d convolutional neural networks. arXiv preprint arXiv:1505.02890, 2015.

[31] Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive
networks. In International Conference on Machine Learning, pages 1242–1250. PMLR, 2014.

[32] Benoit Guillard, Edoardo Remelli, Pierre Yvernay, and Pascal Fua. Sketch2mesh: Reconstructing and
editing 3d shapes from sketches. In Proceedings of the IEEE/CVF International Conference on Computer
Vision, pages 13023–13032, 2021.

[33] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural
information processing systems, 33:6840–6851, 2020.

[34] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli,
Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3d. In The Twelfth
International Conference on Learning Representations, 2024.

[35] Hugues Hoppe. New quadric metric for simplifying meshes with appearance attributes. In Proceedings
Visualization’99 (Cat. No. 99CB37067), pages 59–510. IEEE, 1999.

[36] Zehuan Huang, Yuanchen Guo, Haoran Wang, Ran Yi, Lizhuang Ma, Yan-Pei Cao, and Lu Sheng.
Mv-adapter: Multi-view consistent image generation made easy. arXiv preprint arXiv:2412.03632, 2024.

23
[37] Ka-Hei Hui, Aditya Sanghi, Arianna Rampini, Kamal Rahimi Malekshan, Zhengzhe Liu, Hooman
Shayani, and Chi-Wing Fu. Make-a-shape: a ten-million-scale 3d shape model. In Forty-first International
Conference on Machine Learning, 2024.
[38] Sadeep Jayasumana, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv
Kumar. Rethinking fid: Towards a better evaluation metric for image generation. In IEEE Computer
Vision and Pattern Recognition (CVPR), pages 9307–9315, 2024.
[39] DaDong Jiang, Xianghui Yang, Zibo Zhao, Sheng Zhang, Jiaao Yu, Zeqiang Lai, Shaoxiong Yang,
Chunchao Guo, Xiaobo Zhou, and Zhihui Ke. Flexitex: Enhancing texture generation with visual
guidance. arXiv preprint arXiv:2409.12431, 2024.
[40] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for
real-time radiance field rendering. ACM Trans. Graph., 42(4):139–1, 2023.
[41] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013.
[42] Perla Sai Raj Kishore, Yizhi Wang, Ali Mahdavi-Amiri, and Hao Zhang. EASI-Tex: Edge-aware mesh
texturing from single-image. ACM Transactions on Graphics (Special Issue of SIGGRAPH), 43(4), 2024.
[43] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu,
Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv
preprint arXiv:2412.03603, 2024.
[44] Venkat Krishnamurthy and Marc Levoy. Fitting smooth surfaces to dense polygon meshes. In Proceedings
of the 23rd annual conference on Computer graphics and interactive techniques, pages 313–324, 1996.
[45] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.
[46] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and
Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3D generation. In European
Conference on Computer Vision (ECCV), 2024.
[47] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and
Chen Change Loy. Gaussiananything: Interactive point cloud latent diffusion for 3d generation. In ICLR,
2025.
[48] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang
Zhang, Wenhan Luo, Ping Tan, et al. Era3d: High-resolution multiview diffusion using efficient row-wise
attention. arXiv preprint arXiv:2405.11616, 2024.
[49] Weiyu Li, Jiarui Liu, Hongyu Yan, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long.
Craftsman: High-fidelity mesh generation with 3d native generation and interactive geometry refiner,
2024.
[50] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang,
Xingchao Liu, Minbin Huang, Zedong Xiao, Dayou Chen, Jiajun He, Jiahao Li, Wenyue Li, Chen Zhang,
Rongwei Quan, Jianxiang Lu, Jiabin Huang, Xiaoyan Yuan, Xiaoxiao Zheng, Yixuan Li, Jihong Zhang,
Chao Zhang, Meng Chen, Jie Liu, Zheng Fang, Weiyan Wang, Jinbao Xue, Yangyu Tao, Jianchen Zhu,
Kai Liu, Sihuan Lin, Yifu Sun, Yun Li, Dongdong Wang, Mingtao Chen, Zhichao Hu, Xiao Xiao, Yan
Chen, Yuhong Liu, Wei Liu, Di Wang, Yong Yang, Jie Jiang, and Qinglin Lu. Hunyuan-dit: A powerful
multi-resolution diffusion transformer with fine-grained chinese understanding, 2024.
[51] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis,
Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In
IEEE Computer Vision and Pattern Recognition (CVPR), 2023.
[52] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample
steps are flawed. In Proceedings of the IEEE/CVF winter conference on applications of computer vision,
pages 5404–5411, 2024.
[53] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for
generative modeling. arXiv preprint arXiv:2210.02747, 2022.
[54] Yaron Lipman, Marton Havasi, Peter Holderrieth, Neta Shaul, Matt Le, Brian Karrer, Ricky T. Q. Chen,
David Lopez-Paz, Heli Ben-Hamu, and Itai Gat. Flow matching guide and code, 2024.
[55] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick.
Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference
on computer vision, pages 9298–9309, 2023.

24
[56] Shang Liu, Chaohui Yu, Chenjie Cao, Wen Qian, and Fan Wang. Vcd-texture: Variance alignment
based 3d-2d co-denoising for text-guided texturing. In European Conference on Computer Vision, pages
373–389. Springer, 2025.

[57] Wen Liu, Zhixin Piao, Jie Min, Wenhan Luo, Lin Ma, and Shenghua Gao. Liquid warping gan: A unified
framework for human motion imitation, appearance transfer and novel view synthesis. In Proceedings of
the IEEE/CVF international conference on computer vision, pages 5904–5913, 2019.

[58] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer
data with rectified flow. arXiv preprint arXiv:2209.03003, 2022.

[59] Yuxin Liu, Minshan Xie, Hanyuan Liu, and Tien-Tsin Wong. Text-guided texturing by synchronized
multi-view diffusion. In SIGGRAPH Asia 2024 Conference Papers, pages 1–11, 2024.

[60] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai
Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain
diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,
pages 9970–9980, 2024.

[61] Jiawei Lu, Yingpeng Zhang, Zengjun Zhao, He Wang, Kun Zhou, and Tianjia Shao. Genesistex2: Stable,
consistent and high-quality text-to-texture generation. arXiv preprint arXiv:2409.18401, 2024.

[62] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-
guided generation of 3d shapes and textures. In IEEE Computer Vision and Pattern Recognition (CVPR),
pages 12663–12673, 2023.

[63] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d
completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition, pages 306–315, 2022.

[64] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre
Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu,
Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel
Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski.
Dinov2: Learning robust visual features without supervision, 2023.

[65] George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshmi-
narayanan. Normalizing flows for probabilistic modeling and inference. Journal of Machine Learning
Research, 22(57):1–64, 2021.

[66] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan
evaluation. In IEEE Computer Vision and Pattern Recognition (CVPR), 2022.

[67] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the
IEEE/CVF International Conference on Computer Vision, pages 4195–4205, 2023.

[68] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional
occupancy networks. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August
23–28, 2020, Proceedings, Part III 16, pages 523–540. Springer, 2020.

[69] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna,
and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv
preprint arXiv:2307.01952, 2023.

[70] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d
diffusion. In The Eleventh International Conference on Learning Representations, 2023.

[71] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish
Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from
natural language supervision. In International conference on machine learning, pages 8748–8763. PMLR,
2021.

[72] Xuanchi Ren, Jiahui Huang, Xiaohui Zeng, Ken Museth, Sanja Fidler, and Francis Williams. Xcube:
Large-scale 3d generative modeling using sparse voxel hierarchies. In Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition, pages 4209–4219, 2024.

[73] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided
texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023.

25
[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution
image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer
vision and pattern recognition, pages 10684–10695, 2022.

[75] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dream-
booth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the
IEEE/CVF conference on computer vision and pattern recognition, pages 22500–22510, 2023.

[76] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Ka-
mal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603–18613, 2022.

[77] Pratheba Selvaraju, Mohamed Nabail, Marios Loizou, Maria Maslioukova, Melinos Averkiou, Andreas
Andreou, Siddhartha Chaudhuri, and Evangelos Kalogerakis. Buildingnet: Learning to label 3d buildings.
In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10397–
10407, October 2021.

[78] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen,
Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model.
arXiv preprint arXiv:2310.15110, 2023.

[79] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view
diffusion for 3d generation. In The Twelfth International Conference on Learning Representations, 2023.

[80] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify:
Generating textures on 3d shape surfaces. In European Conference on Computer Vision (ECCV), pages
72–88. Springer, 2022.

[81] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion
phase manifolds. ACM Transactions on Graphics (TOG), 41(4):1–13, 2022.

[82] Sebastian Starke, Paul Starke, Nicky He, Taku Komura, and Yuting Ye. Categorical codebook matching
for embodied character controllers. ACM Transactions on Graphics (TOG), 43(4):1–14, 2024.

[83] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning
multi-contact character movements. ACM Transactions on Graphics (TOG), 39(4):54–1, 2020.

[84] Sebastian Starke, Yiwei Zhao, Fabio Zinno, and Taku Komura. Neural animation layering for synthesizing
martial arts movements. ACM Transactions on Graphics (TOG), 40(4):1–16, 2021.

[85] Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with
an explicit shape bias. In Proceedings of the IEEE/CVF conference on computer vision and pattern
recognition, pages 1798–1808, 2021.

[86] Jiaxiang Tang, Zhaoshuo Li, Zekun Hao, Xian Liu, Gang Zeng, Ming-Yu Liu, and Qinsheng Zhang.
Edgerunner: Auto-regressive auto-encoder for artistic mesh generation. arXiv preprint arXiv:2409.18114,
2024.

[87] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian
splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023.

[88] Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra,
Yasutaka Furukawa, and Rakesh Ranjan. Mvdiffusion++: A dense high-resolution multi-view diffusion
model for single or sparse-view 3d object reconstruction. In European Conference on Computer Vision,
pages 175–191. Springer, 2025.

[89] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling
holistic multi-view image generation with correspondence-aware diffusion. arXiv, 2023.

[90] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix,
Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard
Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint
arXiv:2302.13971, 2023.

[91] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay
Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and
fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023.

26
 [92] Greg Turk and Marc Levoy. Zippered polygon meshes from range images. In Proceedings of the 21st
annual conference on Computer graphics and interactive techniques, pages 311–318, 1994.

[93] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017.

[94] Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv
preprint arXiv:2312.02201, 2023.

[95] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy.
Esrgan: Enhanced super-resolution generative adversarial networks. In The European Conference on
Computer Vision Workshops (ECCVW), September 2018.

[96] Haohan Weng, Zibo Zhao, Biwen Lei, Xianghui Yang, Jian Liu, Zeqiang Lai, Zhuo Chen, Yuhong Liu,
Jie Jiang, Chunchao Guo, et al. Scaling mesh generation via compressive tokenization. arXiv preprint
arXiv:2411.07025, 2024.

[97] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic
latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information
processing systems, 29, 2016.

[98] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao
Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. arXiv preprint
arXiv:2405.14832, 2024.

[99] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaoou Tang, and Jianxiong Xiao.
3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pages 1912–1920, 2015.

[100] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin
Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint
arXiv:2412.01506, 2024.

[101] Xiaoyu Xiang, Liat Sless Gorelik, Yuchen Fan, Omri Armstrong, Forrest Iandola, Yilei Li, Ita Lifshitz,
and Rakesh Ranjan. Make-a-texture: Fast shape-aware texture generation in 3 seconds. arXiv preprint
arXiv:2412.07766, 2024.

[102] Jiacong Xu, Yi Zhang, Jiawei Peng, Wufei Ma, Artur Jesslen, Pengliang Ji, Qixin Hu, Jiehua Zhang,
Qihao Liu, Jiahao Wang, et al. Animal3d: A comprehensive dataset of 3d animal pose and shape. arXiv
preprint arXiv:2308.11737, 2023.

[103] Jingwei Xu, Chenyu Wang, Zibo Zhao, Wen Liu, Yi Ma, and Shenghua Gao. Cad-mllm: Unifying
multimodality-conditioned cad generation with mllm. arXiv preprint arXiv:2411.04954, 2024.

[104] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Car-
los Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point
clouds for 3d understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern
recognition, pages 1179–1189, 2023.

[105] Xingguang Yan, Liqiang Lin, Niloy J Mitra, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Shape-
former: Transformer-based shape completion via sparse representation. In Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition, pages 6239–6249, 2022.

[106] Xianghui Yang, Huiwen Shi, Bowen Zhang, Fan Yang, Jiacheng Wang, Hongxu Zhao, Xinhai Liu,
Xinzhou Wang, Qingxiang Lin, Jiaao Yu, et al. Hunyuan3d-1.0: A unified framework for text-to-3d and
image-to-3d generation. arXiv preprint arXiv:2411.02293, 2024.

[107] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter
for text-to-image diffusion models, 2023.

[108] Yu-Ying Yeh, Jia-Bin Huang, Changil Kim, Lei Xiao, Thu Nguyen-Phuoc, Numair Khan, Cheng Zhang,
Manmohan Chandraker, Carl S Marshall, Zhao Dong, et al. Texturedreamer: Image-guided texture
synthesis through geometry-aware diffusion. In Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition, pages 4304–4314, 2024.

[109] Fukun Yin, Xin Chen, Chi Zhang, Biao Jiang, Zibo Zhao, Jiayuan Fan, Gang Yu, Taihao Li, and Tao
Chen. Shapegpt: 3d shape generation with a unified multi-modal language model. arXiv preprint
arXiv:2311.17618, 2023.

27
[110] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and
Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4252–4262, 2024.

[111] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representa-
tion for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 42(4):1–16,
2023.

[112] Hongkun Zhang, Zherong Pan, Congyi Zhang, Lifeng Zhu, and Xifeng Gao. Texpainter: Generative mesh
texturing with multi-view consistency. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024.

[113] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu,
and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets.
ACM Transactions on Graphics (TOG), 43(4):1–20, 2024.

[114] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion
models, 2023.

[115] Lyumin Zhang. https://github.com/mikubill/sd-webui-controlnet/discussions/1236.

https://github.com/Mikubill/sd-webui-controlnet/discussions/1236, 2023.

[116] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable
effectiveness of deep features as a perceptual metric. In IEEE Computer Vision and Pattern Recognition
(CVPR), 2018.

[117] Song-Hai Zhang, Yuan-Chen Guo, and Qing-Wen Gu. Sketch2model: View-aware 3d modeling from
single free-hand sketches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition, pages 6012–6021, 2021.

[118] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, Bin Fu, Tao Chen, Gang Yu, and
Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent
representation. Advances in Neural Information Processing Systems, 36, 2024.

[119] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally
attentional sdf diffusion for controllable 3d shape generation. ACM Transactions on Graphics (ToG),
42(4):1–13, 2023.

[120] Jie Zhou, Zhongjin Luo, Qian Yu, Xiaoguang Han, and Hongbo Fu. Ga-sketching: Shape modeling
from multi-view sketching with geometry-aligned deep implicit functions. In Computer Graphics Forum.
Wiley Online Library, 2023.

[121] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d:
Exploring unified 3d representation at scale. arXiv preprint arXiv:2310.06773, 2023.
[122] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. arXiv preprint
arXiv:1605.04797, 2016.

28