1. Bibliographic Information

1.1. Title

The title of the paper is "Hunyuan3D 2.5: Towards High-Fidelity 3D Assets Generation with Ultimate Details". The central topic is the advancement of 3D diffusion models for generating highly detailed and realistic textured 3D assets.

1.2. Authors

The authors are listed under the affiliation "Tencent Hunyuan3D". The core contributors are further detailed in Section 6, including individuals for Shape Generation (Zeqiang Lai, Yunfei Zhao, Jingwei Huang, Haolin Liu, Zibo Zhao, Qingxiang Lin, Huiwen Shi, Xianghui Yang) and Texture Generation (Mingxin Yang, Shuhui Yang, Yifei Feng, Sheng Zhang, Xin Huang). Project leaders are Chunchao Guo, Jingwei Huang, Zeqiang Lai, and project sponsors are Jie Jiang, Linus.

1.3. Journal/Conference

This report is published as a preprint on arXiv. arXiv is an open-access archive for scholarly articles, primarily in physics, mathematics, computer science, quantitative biology, quantitative finance, statistics, electrical engineering and systems science, and economics. As a preprint server, it allows researchers to disseminate their work quickly before formal peer review and publication in a traditional journal or conference. While widely used and respected for early dissemination, it implies the work has not yet undergone or completed formal peer review.

1.4. Publication Year

The paper was published on 2025-06-19T17:57:40.000Z, indicating a publication year of 2025.

1.5. Abstract

The report introduces Hunyuan3D 2.5, a suite of 3D diffusion models designed for generating high-fidelity and detailed textured 3D assets. It builds upon its predecessor, Hunyuan3D 2.0, by significantly improving both shape and texture generation. For shape generation, a new foundation model called LATTICE is presented, trained on large-scale, high-quality datasets, with increased model size (up to 10B parameters) and computational resources. This model generates sharp, detailed 3D shapes with precise image-to-3D following, while maintaining clean and smooth mesh surfaces, thereby narrowing the gap between generated and handcrafted 3D assets. For texture generation, Hunyuan3D 2.5 incorporates physical-based rendering (PBR) through a novel multi-view architecture, extending the Hunyuan3D 2.0 Paint model. Extensive evaluations demonstrate that Hunyuan3D 2.5 significantly surpasses prior methods in both shape and end-to-end texture generation.

1.6. Original Source Link

Official Source: https://arxiv.org/abs/2506.16504 PDF Link: https://arxiv.org/pdf/2506.16504v1.pdf Publication Status: This is an arXiv preprint.

2. Executive Summary

2.1. Background & Motivation

The core problem the paper aims to solve is the generation of high-fidelity and detailed textured 3D assets. This problem is crucial for various industries, including game development, embodied AI, film special effects, and virtual reality, where the demand for realistic and efficient 3D model creation is constantly growing.

Existing 3D shape diffusion models, while revolutionary, still struggle with generating complex objects that exhibit fine-grained details while maintaining smooth surfaces and sharp edges. Furthermore, high-quality textures are essential for visual realism, but current methods face challenges with global consistency across multiple views, leading to artifacts. A significant gap exists in the open-source community for physical-based rendering (PBR) material generation solutions, which are vital for photorealistic 3D assets. The paper explicitly points out that existing methods fail at detail generation and produce incorrect PBR, as illustrated in Figure 2.

The paper's entry point is to build upon the successful two-stage pipeline of its predecessor, Hunyuan3D 2.0, and introduce substantial advancements in both the shape and texture generation stages, specifically targeting these identified gaps in detail, surface quality, and PBR realism.

2.2. Main Contributions / Findings

The paper makes several primary contributions:

• Novel Shape Foundation Model (LATTICE): Introduction of LATTICE, a new large-scale diffusion model for 3D shape generation. It is trained with scaled high-quality datasets, model-size (up to 10B parameters), and compute. This scaling enables the generation of sharp and detailed 3D shapes with precise image-3D following while ensuring clean and smooth mesh surfaces. This significantly closes the gap between automatically generated and professionally handcrafted 3D shapes.

• Upgraded PBR Texture Generation: Enhancement of the texture generation stage with physical-based rendering (PBR) capabilities. This is achieved through a novel multi-view architecture that extends the Hunyuan3D 2.0 Paint model, allowing for the simultaneous production of albedo, roughness, and metallic maps. This advancement enables more realistic and detailed rendering by accurately describing surface reflection properties and simulating geometric microsurface distributions.

• Dual-Phase Resolution Enhancement for Texture: Introduction of a dual-phase resolution enhancement strategy to strengthen texture-geometry coordination and improve visual quality, especially for high-resolution details, while managing computational feasibility.

• Significant Performance Outperformance: Extensive quantitative and qualitative evaluations, including user studies, demonstrate that Hunyuan3D 2.5 substantially outperforms previous state-of-the-art open-source and commercial methods in both shape generation and end-to-end texture generation.

  The key findings are that by combining large-scale model training with innovative architectural designs for PBR and multi-view consistency, it is possible to generate 3D assets that rival handcrafted quality in terms of detail, surface integrity, and photorealistic texturing. These findings address the limitations of prior methods in producing complex, fine-grained 3D content suitable for demanding applications.

3. Prerequisite Knowledge & Related Work

3.1. Foundational Concepts

To understand Hunyuan3D 2.5, a reader should be familiar with several core concepts in 3D computer graphics and generative AI:

• 3D Diffusion Models: At the heart of Hunyuan3D 2.5 are diffusion models. These are a class of generative models that learn to generate data by reversing a gradual diffusion process. In the forward diffusion process, noise is progressively added to data (e.g., an image or 3D representation) until it becomes pure noise. The model is then trained to learn the reverse process, predicting and removing the noise at each step to reconstruct the original data from pure noise. For 3D generation, this means learning to generate 3D shapes or textures from noisy latent representations.
• Two-Stage Pipeline: This refers to a common approach in complex generative tasks where the overall problem is broken down into sequential steps. In Hunyuan3D 2.5, this involves generating the 3D shape first (the geometry of the object), and then generating the texture (the visual appearance, color, and material properties) for that generated shape. This modularity allows for specialized models at each stage.
• Physical-Based Rendering (PBR): PBR is a modern rendering technique that aims to simulate how light interacts with surfaces in a physically accurate way, resulting in more realistic visuals. Instead of simply providing an RGB color map, PBR uses multiple material maps to define a surface's properties:
  • Albedo Map (Base Color): The intrinsic color of the surface, representing how much light it reflects when lit directly, excluding any lighting information.
  • Metallic Map: A grayscale map indicating how metallic a surface is. Pure black means non-metallic (dielectric), pure white means metallic. Metallic surfaces reflect light differently than non-metallic ones.
  • Roughness Map: A grayscale map indicating the microscopic surface irregularities. Black represents a perfectly smooth, shiny surface (like polished chrome), while white represents a rough, matte surface (like sandpaper). This influences how sharp or blurry reflections and highlights appear. Hunyuan3D 2.5 generates these PBR maps, particularly albedo and metallic/roughness (MR), to achieve high realism.
• Mesh: A 3D mesh is a collection of vertices, edges, and faces that define the shape of a 3D object. Vertices are points in space, edges connect vertices, and faces (typically triangles or quadrilaterals) are enclosed by edges, forming the surface of the object. Hunyuan3D 2.5 outputs 3D meshes as its shape representation.
• Multi-view Consistency: In 3D generation, especially for textures, multi-view consistency is crucial. This means that if you render an object from different camera angles, the textures and details should appear coherent and seamlessly wrap around the object without visible seams or distortions. Multi-view diffusion models are designed to achieve this by considering information from various viewpoints during the generation process.
• Latent Diffusion Models (LDMs): Many modern diffusion models operate in a latent space rather than directly on pixel data. An autoencoder compresses high-dimensional data (like images) into a lower-dimensional latent representation. The diffusion process then happens in this compressed latent space, making training and inference more efficient. The generated latent is then decoded back into the high-dimensional data. This is often the underlying mechanism for generating images or textures from noise.
• Attention Mechanisms: Attention mechanisms, particularly cross-attention, are fundamental components in transformer-based models and are used here. Cross-attention allows a model to weigh the importance of different parts of one sequence (e.g., a reference image or geometry conditions) when processing another sequence (e.g., the latent representation being diffused). In Hunyuan3D 2.5, it's used to inject learned embeddings into material channels and ensure spatial alignment.

3.2. Previous Works

The paper contextualizes Hunyuan3D 2.5 within the rapidly evolving landscape of 3D generation, referencing several key prior studies:

• 3D Shape Diffusion Models based on 3dshape2vecset: These models (Zhang et al., 2023) revolutionized 3D shape generation by introducing a novel 3D shape representation.
  • CLAY (Zhang et al., 2024b)
  • Hunyuan3D 2.0 (Zhao et al., 2025): The direct predecessor to Hunyuan3D 2.5, establishing the two-stage pipeline.
  • TripoSG (Li et al., 2025)
  • These methods typically encode 3D shapes into a compact latent space, allowing diffusion models to operate on these representations.
• Triplane-based Representation:
  • Direct3D (Wu et al., 2024b): Explored compressing and generating shapes using triplanes, which are three orthogonal 2D planes that encode 3D information.
• Structured 3D Latents:
  • Trellis (Xiang et al., 2024): A promising pipeline for high-quality textured 3D generation that uses an invented structured 3D latents representation.
• Multiview Diffusion for Texture Generation: A significant class of methods for generating textures, addressing global consistency issues.
  • Hunyuan3D 2.0 Paint model (Zhao et al., 2025): The basis for the texture generation in Hunyuan3D 2.5.
  • $Zero123++$ (Shi et al., 2023a): Uses spatial concatenation of multi-view images and self-attention for cross-view interaction.
  • Other works like MVDream (Shi et al., 2023b), and methods by Huang et al., 2024b; Vainer et al., 2024b; Li et al., 2024a; Tang et al., 2025; Long et al., 2024; Wang & Shi, 2023; Liu et al., 2023, which inject view constraints into attention blocks using diverse attention masks.
• Inpainting-based Methods: An alternative approach for texture generation that fills in missing texture regions. Hunyuan3D 2.5 notes these methods suffer from global consistency issues. Examples include Huang et al., 2024a; Wu et al., 2024a; Zhang et al., 2024c; Ceylan et al., 2024; Zeng et al., 2024a; Chen et al., 2023a; Richardson et al., 2023.
• Synchronization Techniques: Other texture generation approaches (Liu et al., 2025; Gao et al., 2024; Liu et al., 2024a; Zhang et al., 2024a).
• PBR Material Generation: The paper highlights the lack of open-source solutions for PBR material generation. It references existing methods mainly categorized into:
  • Generation-based (Vainer et al., 2024a; Sartor & Peers, 2023; Vecchio et al., 2024; Chen et al., 2024a; Zeng et al., 2024b): Leverage diffusion models to learn material priors.
  • Retrieval-based (Zhang et al., 2024c; Fang et al., 2024): Adapt pre-built material graphs.
  • Optimization-based (Chen et al., 2023b; Zhang et al., 2024d; Wu et al., 2023; Xu et al., 2023; Yeh et al., 2024; Youwang et al., 2024; Liu et al., 2024b): Generate initial textures and refine them using techniques like Score-Distillation Sampling (SDS).
• Other Shape Generation Models:
  • Michelangelo (Zhao et al., 2024), Craftsman 1.5 (Li et al., 2024b), LRM (Hong et al., 2023), Hunyuan3D 1.0 (Yang et al., 2024), LGM (Tang et al., 2024), FlashVDM (Lai et al., 2025).
  • Autoregressive models like MeshGPT (Siddiqui et al., 2024), BPT (Weng et al., 2024), and Meshtron (Hao et al., 2024).

3.3. Technological Evolution

3D generation has evolved dramatically:

1. Early Generative Models (2010s): Initial attempts used Generative Adversarial Networks (GANs) (Goodfellow et al., 2014) and Variational Autoencoders (VAEs) (Kingma, 2013) to generate simple 3D shapes or point clouds (e.g., Wu et al., 2016). These were often category-specific and lacked high detail.

2. Score-Distillation and Text-to-3D (Early 2020s): The rise of diffusion models (Rombach et al., 2022; Ho et al., 2020) combined with Score-Distillation Sampling (SDS) (Poole et al., 2023) enabled text-to-3D generation by leveraging powerful 2D text-to-image models. This marked a shift towards broader, more flexible generation.

3. Feedforward and Native 3D Diffusion (Mid 2020s): Feedforward methods (e.g., LRM, Hunyuan3D 1.0, LGM) aimed for single-step 3D asset generation. Concurrently, native 3D diffusion models emerged, operating directly on 3D data representations (like vecset in Zhang et al., 2023), significantly improving generation quality (e.g., Michelangelo, CLAY, Trellis, Hunyuan3D 2.0, TripoSG). These methods, while often multi-step, could be accelerated (e.g., FlashVDM).

4. Autoregressive Mesh Generation: Another parallel line of research focused on generating meshes with human-like topology using autoregressive models (e.g., MeshGPT, BPT, Meshtron).

5. High-Fidelity, Detailed, and PBR-enabled Generation (Current): The current frontier, addressed by Hunyuan3D 2.5, is pushing towards unprecedented detail, sharp edges, smooth surfaces, and photorealistic textures using PBR materials. This involves scaling up models and datasets, and developing sophisticated multi-view architectures for consistency and realism.

   Hunyuan3D 2.5 fits into this timeline by representing a significant step forward in the native 3D diffusion model paradigm, specifically focusing on overcoming the fine-grained detail and PBR texture limitations of its predecessors and other state-of-the-art models.

3.4. Differentiation Analysis

Compared to the main methods in related work, Hunyuan3D 2.5 offers several core differences and innovations:

• Enhanced Shape Fidelity via LATTICE: While models like CLAY, Hunyuan3D 2.0, and TripoSG advanced shape generation using 3dshape2vecset, Hunyuan3D 2.5 introduces LATTICE. The key differentiation is the unprecedented scaling of datasets, model-size (10B parameters), and compute, leading to a qualitative leap in shape generation. LATTICE is specifically designed to produce extreme detail, sharp edges, and smooth surfaces simultaneously, which is a common struggle for existing methods (as shown in Figure 4 and Figure 6). This makes generated shapes much closer to handcrafted quality.

• True PBR Material Generation: Many texture generation methods focus on RGB textures or struggle to accurately decouple material properties from illumination. Hunyuan3D 2.5 explicitly upgrades its texture generation to PBR, producing albedo, metallic, and roughness maps based on a principled BRDF model. This is a significant improvement over methods that might infer PBR from RGB or lack a dedicated PBR output, directly addressing the incorrect PBR issue highlighted in Figure 2. The paper notes that PBR material generation solutions are largely unavailable in the open-source community, positioning Hunyuan3D 2.5 as a leader in this aspect.

• Novel Multi-view PBR Architecture with Dual-Channel Attention: Building on Hunyuan3D 2.0's multi-view approach, Hunyuan3D 2.5 introduces a specific multi-view architecture with learnable embeddings for each PBR channel (albedo, MR). Crucially, the dual-channel attention mechanism ensures spatial coherence between albedo and MR maps by sharing the attention mask, which is a subtle yet effective way to prevent misalignment artifacts that commonly plague multi-channel texture generation.

• Robust Geometric Alignment with Dual-Phase Resolution Enhancement: Addressing the challenge of integrating high-resolution textures with complex geometry while managing memory, Hunyuan3D 2.5 proposes a dual-phase resolution enhancement strategy. This zoom-in training approach allows the model to learn fine-grained details without the prohibitive memory costs of full high-resolution multi-view training, offering a practical solution that improves texture-geometry coordination.

• Superior End-to-End Performance: Through comprehensive quantitative metrics and user studies, Hunyuan3D 2.5 demonstrates superior performance across both shape and textured 3D asset generation compared to a wide range of open-source and commercial models (Tables 1, 2, and Figure 8). This indicates that the combination of LATTICE for shapes and the PBR-enabled multi-view texture model leads to a more robust and higher-quality end-to-end pipeline.

  In essence, Hunyuan3D 2.5 differentiates itself by pushing the boundaries of detail fidelity through scaling, embracing physically accurate rendering with a dedicated PBR pipeline, and innovating architectural components to ensure multi-channel consistency and efficient high-resolution learning.

4. Methodology

4.1. Principles

The core idea behind Hunyuan3D 2.5 is to generate high-fidelity and detailed textured 3D assets by employing a two-stage pipeline. This pipeline separates the complex problem into manageable sub-problems: first, generating the 3D shape, and then texturing that shape. The theoretical basis for this approach lies in the power of diffusion models to learn complex data distributions, scaled up to generate unprecedented geometric detail and combined with principles of physical-based rendering (PBR) for photorealistic textures. Intuition suggests that a modular approach allows each stage to specialize: the shape generation model focuses solely on geometric accuracy and detail, while the texture generation model focuses on material realism and multi-view consistency, leveraging geometric information provided by the first stage.

4.2. Core Methodology In-depth (Layer by Layer)

Hunyuan3D 2.5 is an image-to-3D generation model that adheres to the overall architecture of Hunyuan3D 2.0. The process can be broken down into the following stages:

4.2.1. Overall Pipeline

The Hunyuan3D 2.5 pipeline, as shown in Figure 3, starts with an input image and concludes with a fully textured 3D model.

The following figure (Figure 3 from the original paper) shows the overview of Hunyuan3D 2.5 pipeline:

该图像是论文中图3的示意图，展示了Hunyuan3D 2.5的整体流程。该流程将3D资产生成分为两个阶段：先进行形状生成，经过网格后处理，再进行纹理生成，最终输出带纹理的3D模型。

Figure 3: Overview of Hunyuan3D 2.5 pipeline. It separates the 3D asset generation into two stages: first, it generates the shape, and then it creates the texture based on that shape.

1. Input Image Processing: The process begins with an input image. This image is first processed by an image processor. The role of this processor is to remove the background from the image and perform proper resizing to prepare it for the subsequent stages.
2. Shape Generation: The pre-processed image is then fed into a shape generation model. This model is conditioned on the input image to generate a 3D mesh. Crucially, at this stage, the mesh is generated without any texture. The shape generation model used here is the newly introduced LATTICE.
3. Mesh Post-processing: Once the 3D mesh is generated, it undergoes further processing. This step involves extracting essential geometric information from the mesh, such as its normal map and UV map.
   • A normal map is a texture that stores information about the direction of surface normals, which allows for detailed lighting effects without increasing polygon count.
   • A UV map is a 2D representation of the 3D mesh's surface, used to project 2D textures onto the 3D object.
4. Texture Generation: The extracted mesh information (normal map, UV map, etc.) and the original input image (or reference image derived from it) are then passed to a texture generation model. This model is responsible for creating the detailed textures.
5. Final Textured 3D Model: The generated textures are then applied to the 3D mesh, resulting in the final textured 3D model.

4.2.2. Detailed Shape Generation (LATTICE)

The first stage focuses on Detailed Shape Generation using LATTICE, a new shape generation model specifically developed for Hunyuan3D 2.5.

• Model Description: LATTICE is a large-scale diffusion model. It is designed to produce high-fidelity, detailed shapes with desirable properties like sharp edges and smooth surfaces. It can be conditioned on either a single image or four multi-view images as input.

• Training Strategy: LATTICE is trained on an extensive and high-quality 3D dataset that features complex objects. The paper emphasizes the importance of scaling up the dataset quality and size, model-size (with the largest model reaching 10B parameters), and computational resources. This scaling is identified as the key factor contributing to the model's stable improvement and exceptional detail generation.

• Efficiency: To ensure efficient inference, the model incorporates guidance and step distillation techniques.

  • Guidance typically involves using a classifier or a classifier-free guidance mechanism to steer the diffusion process towards desired outputs (e.g., matching the input image more closely).
  • Step distillation aims to reduce the number of sampling steps required during inference, thereby speeding up the generation process significantly without a substantial loss in quality.

• Major Features (Figure 4): The following figure (Figure 4 from the original paper) shows major features of the new shape generation model in Hunyuan3D 2.5:

  
  该图像是论文中图4的示意图，展示了Hunyuan3D 2.5新形状生成模型的主要特征，包括极致细节、锐利边缘和平滑表面，突出表现了复杂小资产、清晰边缘和干净网格表面。

  Figure 4: Illustration of major features of the new shape generation model in Hunyuan3D 2.5.

  • Extreme Detail: Hunyuan3D 2.5 is capable of generating fine-grained details at an unprecedented level. Examples include correctly rendered fingers on a hand, intricate patterns on a bicycle wheel, and even small objects like a bowl within a larger scene. This level of detail is attributed to the model's scaling-up strategy, allowing it to approach the accuracy of handcrafted designs.
  • Smooth Surfaces & Sharp Edges: A challenge for many existing models is simultaneously achieving sharp edges (e.g., corners of objects, distinct features) while maintaining smooth, clean surfaces (e.g., the body of a character, the side of a vehicle). Hunyuan3D 2.5 is highlighted for striking an excellent balance between these two often conflicting requirements, especially for complex objects.

4.2.3. Realistic Texture Generation

The second stage, Realistic Texture Generation, extends the Hunyuan3D 2.0 and 2.1 texture models into a high-fidelity material generation framework, focusing on PBR.

The following figure (Figure 5 from the original paper) shows the overview of the material generation framework:

该图像是论文中描述材质生成框架的示意图，展示了训练与推理流程及多任务注意力模块的结构，包含输入图像、参考分支、生成分支、视点选择及多视角注意力机制等关键部分。

Figure 5: Overview of material generation framework.

• Architecture (Figure 5): The model takes normal map and CCM (likely meaning Color-Coded Mesh or similar geometric conditions) rendered from the 3D mesh as geometric conditions. It also uses a reference image as guidance. The output is high-quality PBR material maps.
• 3D-aware RoPE: The framework inherits 3D-aware Rotary Positional Embeddings (RoPE) from (Feng et al., 2025). RoPE (Su et al., 2022) is a method to encode positional information into attention mechanisms by rotating query and key vectors. 3D-aware RoPE likely extends this concept to 3D contexts, enhancing cross-view consistency by providing a robust spatial understanding across different viewpoints, which is crucial for seamless texture map generation.

4.2.3.1. Multi-Channel Material Generation

The model is designed to generate multiple PBR material maps simultaneously, specifically albedo, metallic, and roughness.

• Learnable Embeddings: To differentiate and guide the generation of distinct material properties, the model introduces learnable embeddings for each material map. These are:

  • $\mathbf { E _ { a l b e d o } }$ for albedo.
  • $\mathbf { E } _ { \mathrm { m r } }$ for MR (a combined expression of metallic and roughness).
  • $\mathbf { E _ { n o r m a l } }$ (though the normal map is typically an input condition, this embedding might be for a generated normal map or related feature). These embeddings, all $\in \mathbb { R } ^ { 1 6 \times 1024 }$, are initialized and then injected into their respective channels using cross-attention layers. The embeddings and attention modules are trainable, allowing the network to effectively model the distribution of these three materials separately.

• Dual-Channel Attention Mechanism: A key innovation for maintaining spatial correspondence across different material channels. Despite the domain gaps between material channels (e.g., color albedo vs. grayscale metallic/roughness), spatial alignment from semantic to pixel level is critical. The paper identifies misaligned attention masks as a main cause of multi-channel misalignment in existing methods. To counter this, a dual-channel attention mechanism is proposed where the attention mask is intentionally shared among multiple channels, while the value computation varies for each output. Specifically, the basecolor (albedo) branch is chosen as the reference because it contains information most semantically similar to the reference image (both in RGB color space). The attention mask calculated from the basecolor channel is then used to guide the reference attention in the other two branches (MR and potentially normal if generated).

  The mechanism is formulated as follows: $$\begin{array} { r } { \mathbf { M } _ { a t t n } = \mathrm { Softmax } \left( \frac { Q _ { a l b edo } K _ { r e f } ^ { T } } { \sqrt { d } } \right) \qquad } \\ { z _ { a l b edo } ^ { n e w } = z _ { a l b edo } + \mathbf { M L P } _ { a l b e do } \left[ \mathbf { M } _ { a t t n } \cdot V _ { a l b e do } \right] , } \\ { z _ { M R } ^ { n e w } = z _ { M R } + \mathbf { M L P } _ { M R } \left[ \mathbf { M } _ { a t t n } \cdot V _ { M R } \right] \qquad } \end{array}$$ Where:

  • $\mathbf { M } _ { a t t n }$: The computed attention mask. This mask quantifies the relevance of different parts of the reference information ($K_{ref}$) to the query from the albedo channel ($Q_{albedo}$).

  • $\mathrm { Softmax } (\cdot)$: The softmax function, which normalizes the attention scores into a probability distribution, ensuring that the weights sum to 1.

  • Q _ { a l b e d o }: The query vector derived from the albedo channel's current feature representation. It asks "what information am I looking for?"

  • K _ { r e f }: The key vector derived from the reference image's features. It represents "what information is available?"

  • $T$: Denotes the transpose operation.

  • $\sqrt { d }$: A scaling factor (where $d$ is the dimension of the key vectors) used to prevent the dot product from becoming too large, which can push the softmax function into regions with very small gradients.

  • $z _ { a l b e d o } ^ { n e w }$: The updated (new) feature representation for the albedo channel after incorporating reference information.

  • z _ { a l b e d o }: The original (current) feature representation for the albedo channel.

  • $\mathbf { M L P } _ { a l b e d o } [\cdot]$: A Multi-Layer Perceptron (MLP) specific to the albedo channel, used to process the attended value information.

  • V _ { a l b e d o }: The value vector derived from the albedo channel's features. It represents "what information is being passed on?" when combined with the attention mask.

  • $z _ { M R } ^ { n e w }$: The updated (new) feature representation for the MR channel.

  • z _ { M R }: The original (current) feature representation for the MR channel.

  • $\mathbf { M L P } _ { M R } [\cdot]$: An MLP specific to the MR channel.

  • V _ { M R }: The value vector derived from the MR channel's features.

    This design ensures that the generated albedo and MR features maintain spatial coherence because they are both guided by the same attention mask $\mathbf { M } _ { a t t n }$, which is derived from the albedo channel's interaction with the reference image.

• Illumination-Invariant Consistency Loss: During training, an illumination-invariant consistency loss (from He et al., 2025) is incorporated. This loss aims to disentangle material properties from illumination components. This is crucial for PBR, as albedo should represent the intrinsic color of an object regardless of how it's lit.

4.2.3.2. Geometric Alignment

Achieving precise texture-geometry alignment is vital, especially for complex, high-polygon geometry. The challenge is that higher-resolution images preserve richer high-frequency geometric details and mitigate VAE compression losses, but training with them for multi-view tasks is memory-intensive.

• Dual-Phase Resolution Enhancement Strategy: To address the memory constraints while still achieving high-quality texture-geometry alignment, Hunyuan3D 2.5 proposes a dual-phase resolution enhancement strategy:
  1. First Phase (Conventional Multi-view Training): The model undergoes initial training using a conventional multi-view training approach. This involves 6-view images at a resolution of $512 \times 512$ pixels, following the methodology of Hunyuan3D 2.0. This phase establishes a strong foundation for multi-view consistency and basic texture-geometry correspondence.
  2. Second Phase (Zoom-in Training Strategy): This phase focuses on capturing high-quality details while preserving the multi-view benefits from the first phase. During this phase, the model employs a zoom-in training strategy. This means that both the reference image and the multi-view generated images are randomly zoomed into during training. This technique allows the model to learn fine-grained texture details without the need for training from scratch with full high-resolution inputs, thus circumventing the memory constraints associated with direct high-resolution multi-view training.
• Inference: During inference, the model leverages multi-view images at resolutions up to $768 \times 768$. This process is further accelerated by the UniPC sampler (Zhao et al., 2023), enabling efficient generation of high-quality results.

5. Experimental Setup

5.1. Datasets

The paper describes its use of datasets for both training and evaluation, though specific dataset names for training are not provided directly within the main text.

• Training Datasets: For the shape generation model (LATTICE), the paper states it is "trained on an extensive and high-quality 3D dataset featuring complex objects". The emphasis is on the scaled high-quality datasets, model-size, and compute. The exact composition or name of this proprietary dataset is not disclosed, which is common for models developed by large research labs like Tencent.
• Evaluation Datasets: For evaluation, the paper mentions:

  • Shape Generation: The evaluation compares generated meshes against input images and image prompts synthesized by vision-language models. The results are presented numerically and visually.

  • Texture Generation: Quantitative comparisons were performed, implying a dataset of ground truth textures or images for comparison. For the user study, the paper states, "The testset included a diverse range of real-world images from various categories." This indicates the use of a varied collection of in-the-wild input images to test image-to-3D and text-to-3D tasks.

    Given the descriptions, the datasets are chosen to validate the method's ability to generate complex objects with fine-grained details and photorealistic textures, covering a wide range of categories encountered in real-world scenarios.

5.2. Evaluation Metrics

The paper employs a comprehensive set of metrics to evaluate both shape generation and texture generation performance.

5.2.1. Shape Generation Metrics

• ULIP (Xue et al., 2023): ULIP (Unified Language-Image-Point Cloud) is a model designed for learning a unified representation of language, images, and point clouds for 3D understanding. In this context, it's used to calculate the similarity between the generated 3D mesh and input cues.

  • Conceptual Definition: ULIP quantifies the semantic alignment between different modalities (text, image, 3D). When used for evaluation, it measures how well the generated 3D shape semantically matches a given input image or text prompt. A higher ULIP score indicates better semantic consistency.
  • ULIP-I: Measures the similarity between the generated mesh and the input images.
  • ULIP-T: Measures the similarity between the generated mesh and image prompts synthesized by a vision-language model (Chen et al., 2024b).
  • Mathematical Formula: The ULIP score is typically based on the cosine similarity between the embeddings of the different modalities in a shared latent space. $ \text{Similarity}(E_A, E_B) = \frac{E_A \cdot E_B}{|E_A| |E_B|} $
  • Symbol Explanation:
    • $E_A$: The embedding vector for modality A (e.g., generated mesh).
    • $E_B$: The embedding vector for modality B (e.g., input image or text prompt).
    • $\cdot$: The dot product operation.
    • $\| \cdot \|$: The Euclidean norm (magnitude) of a vector. The ULIP model provides these embeddings, and the formula calculates their cosine similarity, ranging from -1 (completely dissimilar) to 1 (perfectly similar). Higher values are better.

• Uni3D (Zhou et al., 2023): Uni3D aims to explore unified 3D representations at scale. Similar to ULIP, it provides a means to measure semantic similarity across modalities involving 3D data.

  • Conceptual Definition: Uni3D evaluates the semantic alignment between generated 3D models and their corresponding input conditions (images or text). It reflects how accurately the generated 3D model captures the semantic content described by the input. A higher Uni3D score indicates better alignment.
  • Uni3D-I: Measures the similarity between the generated mesh and the input images.
  • Uni3D-T: Measures the similarity between the generated mesh and image prompts synthesized by a vision-language model.
  • Mathematical Formula: Similar to ULIP, Uni3D scores are based on the cosine similarity of embeddings from different modalities, computed within the Uni3D model's unified embedding space. $ \text{Similarity}(E_A, E_B) = \frac{E_A \cdot E_B}{|E_A| |E_B|} $
  • Symbol Explanation: The symbols are the same as for ULIP, referring to the respective embeddings generated by the Uni3D model. Higher values are better.

5.2.2. Texture Generation Metrics

• Fréchet Inception Distance (FID):

  • Conceptual Definition: FID measures the similarity between the feature distributions of generated images and real (ground truth) images. It uses features extracted from a pre-trained Inception network. A lower FID score indicates that the generated images are closer to the real images in terms of perceptual quality and diversity, implying higher realism.
  • Mathematical Formula: $ \text{FID} = |\mu_1 - \mu_2|^2 + \text{Tr}(\Sigma_1 + \Sigma_2 - 2(\Sigma_1 \Sigma_2)^{1/2}) $
  • Symbol Explanation:
    • $\mu_1$: The mean of the feature vectors for the real images.
    • $\mu_2$: The mean of the feature vectors for the generated images.
    • $\Sigma_1$: The covariance matrix of the feature vectors for the real images.
    • $\Sigma_2$: The covariance matrix of the feature vectors for the generated images.
    • $\|\cdot\|^2$: The squared Euclidean distance.
    • $\text{Tr}(\cdot)$: The trace of a matrix.
    • $(\Sigma_1 \Sigma_2)^{1/2}$: The matrix square root of the product of the covariance matrices. Lower FID values are better.

• CLIP-based FID (CLIP-FID):

  • Conceptual Definition: CLIP-FID is similar to FID but uses features extracted from the CLIP (Contrastive Language-Image Pre-training) model instead of InceptionV3. CLIP embeddings capture higher-level semantic information. CLIP-FID assesses the perceptual quality and semantic realism of generated images based on CLIP features. A lower CLIP-FID indicates better alignment with real image distributions in CLIP's semantic space.
  • Mathematical Formula: The formula structure is identical to FID, but the $\mu$ and $\Sigma$ are derived from CLIP embeddings rather than InceptionV3 embeddings. $ \text{CLIP-FID} = |\mu_1 - \mu_2|^2 + \text{Tr}(\Sigma_1 + \Sigma_2 - 2(\Sigma_1 \Sigma_2)^{1/2}) $
  • Symbol Explanation: The symbols are the same as for FID, but they refer to the mean and covariance of CLIP feature vectors for real and generated images. Lower CLIP-FID values are better.

• Learned Perceptual Image Patch Similarity (LPIPS):

  • Conceptual Definition: LPIPS measures the perceptual similarity between two images, often a generated image and a ground truth image. It uses features from a pre-trained deep convolutional neural network (like AlexNet, VGG, or ResNet) to compute a weighted distance between feature maps, aiming to correlate better with human judgment of similarity than pixel-wise metrics like MSE. A lower LPIPS score means the images are perceptually more similar.
  • Mathematical Formula: $ \text{LPIPS}(x, x_0) = \sum_l \frac{1}{H_l W_l} \sum_{h,w} |w_l \odot (\phi_l(x){h,w} - \phi_l(x_0){h,w})|^2_2 $
  • Symbol Explanation:
    • $x$: The generated image.
    • $x_0$: The ground truth image.
    • $l$: Index representing different layers of the pre-trained network.
    • $H_l, W_l$: Height and width of the feature maps at layer $l$.
    • $w_l$: A learnable vector of weights for channel $l$.
    • $\odot$: Element-wise multiplication.
    • $\phi_l(\cdot)$: The feature map extracted from layer $l$ of the pre-trained network.
    • $\|\cdot\|^2_2$: The squared L2 norm (Euclidean distance). Lower LPIPS values are better.

• CLIP Maximum-Mean Discrepancy (CMMD):

  • Conceptual Definition: CMMD is used to assess the diversity and richness of generated texture details. It measures the discrepancy between the distributions of CLIP embeddings of generated images and ground truth images. A lower CMMD suggests that the generated samples cover a similar range of variations and details as the real samples.
  • Mathematical Formula: MMD (Maximum Mean Discrepancy) quantifies the distance between two probability distributions. When CLIP features are used, it becomes CMMD. The general form of MMD for distributions $P$ and $Q$ is: $ \text{MMD}^2(P, Q) = \mathbb{E}_P[k(x, x')] + \mathbb{E}Q[k(y, y')] - 2\mathbb{E}{P,Q}[k(x, y)] $
  • Symbol Explanation:
    • P, Q: The two distributions being compared (e.g., CLIP features of real vs. generated images).
    • $\mathbb{E}[\cdot]$: Expectation operator.
    • k(a, b): A kernel function (e.g., Gaussian kernel) that measures the similarity between two samples.
    • x, x': Samples drawn from distribution $P$.
    • y, y': Samples drawn from distribution $Q$. Lower CMMD values are better.

• CLIP-Image Similarity (CLIP-I):

  • Conceptual Definition: CLIP-I measures how well the generated textures semantically align with the input images (or image prompt methods). It uses the CLIP model to embed both the generated texture and the input image (or reference image) into a shared semantic space and then calculates their cosine similarity. A higher CLIP-I score indicates better semantic consistency between the input and the generated output.
  • Mathematical Formula: This is a direct application of CLIP's image-to-image similarity, often calculated as cosine similarity between their respective CLIP image embeddings. $ \text{CLIP-I}(I_{\text{gen}}, I_{\text{input}}) = \frac{E_{\text{CLIP}}(I_{\text{gen}}) \cdot E_{\text{CLIP}}(I_{\text{input}})}{|E_{\text{CLIP}}(I_{\text{gen}})| |E_{\text{CLIP}}(I_{\text{input}})| } $
  • Symbol Explanation:
    • $I_{\text{gen}}$: The generated texture image.
    • $I_{\text{input}}$: The input (or reference) image.
    • $E_{\text{CLIP}}(\cdot)$: The CLIP model's embedding function for images.
    • $\cdot$: Dot product.
    • $\| \cdot \|$: Euclidean norm. Higher CLIP-I values are better.

5.3. Baselines

The proposed Hunyuan3D 2.5 method is compared against several state-of-the-art open-source and closed-source commercial models for both shape generation and texture generation.

5.3.1. Shape Generation Baselines

• Open-source baselines:
  • Michelangelo (Zhao et al., 2024)
  • Craftsman 1.5 (Li et al., 2024b)
  • Trellis (Xiang et al., 2024)
  • Hunyuan3D 2.0 (Zhao et al., 2025): The direct predecessor, providing a clear comparison point for the advancements in Hunyuan3D 2.5.
• Closed-source baselines:
  • Commercial Model 1
  • Commercial Model 2 These baselines are representative as they include recent and widely recognized models in the 3D generation research community, alongside commercial solutions that often set high standards for quality.

5.3.2. Texture Generation Baselines

• Quantitative comparison (Text- and Image-conditioned methods):
  • Text2Tex (Chen et al., 2023a)
  • SyncMVD (Liu et al., 2024a)
  • Paint-it (Youwang et al., 2024)
  • Paint3D (Zeng et al., 2024a)
  • TexGen (Yu et al., 2024) These baselines represent state-of-the-art approaches in 3D texture synthesis, covering both text- and image-conditioned paradigms.
• Qualitative comparison and User Study:
  • Three different commercial models (unnamed specifically for texture generation, but likely the same or similar to those used for shape generation comparison). These commercial models serve as strong benchmarks for evaluating the practical quality and user preference of the generated textures.

6. Results & Analysis

6.1. Core Results Analysis

The experimental results demonstrate that Hunyuan3D 2.5 consistently outperforms baseline methods in both shape and texture generation, validating the effectiveness of its scaled LATTICE model and PBR-enabled multi-view texture architecture.

6.1.1. Shape Generation

The numerical comparison for shape generation is provided in Table 1.

The following are the results from Table 1 of the original paper: Table 1: Numerical comparisons of different shape generation models on ULIP-T/I, Uni3D-T/I.

• Numerical Superiority: Hunyuan3D 2.5 achieves the highest scores in ULIP-T (0.07853) and Uni3D-T (0.2542), indicating superior text-to-shape similarity. It also matches the best performance in Uni3D-I (0.3151) and is competitive in ULIP-I (0.1306). These results suggest that the LATTICE model is highly effective at interpreting semantic information from text and images to generate relevant 3D shapes.

• Comparison to Predecessor: When compared to its predecessor Hunyuan3D 2.0, Hunyuan3D 2.5 shows improvements in ULIP-T (0.07853 vs 0.0771) and Uni3D-T (0.2542 vs 0.2519), and maintains similar performance on image-based metrics, highlighting the impact of the LATTICE model's enhancements.

• Visual Comparison (Figure 6): The paper notes that numerical metrics might not fully reflect model capabilities. Figure 6 provides a visual comparison, which often captures qualitative aspects better. The following figure (Figure 6 from the original paper) shows a visual comparison of different methods in terms of shape generation:

  
  该图像是论文中图6的插图，展示了多种方法在3D形状生成上的视觉对比。包含输入彩色图片及基于不同模型（Trellis、Hunyuan3D 2.0、两款商业模型及Hunyuan3D 2.5）的灰度雕塑形态，直观体现形状细节和复杂度的差异。

  Figure 6: Visual comparison of different methods in terms of shape generation. Figure 6 visually demonstrates the superior detail, sharpness, and smoothness of shapes generated by Hunyuan3D 2.5 compared to Trellis, Hunyuan3D 2.0, and two commercial models. For example, the generated character models by Hunyuan3D 2.5 show more intricate features and cleaner surfaces. This visual evidence strongly supports the claims of improved fidelity and detail generation.

6.1.2. Texture Generation

The quantitative comparison for texture generation is presented in Table 2.

The following are the results from Table 2 of the original paper: Table 2: Quantitative comparison with state-of-the-art methods. We compare with two classes of methods, one conditioned on text only, and the other one based on image. Our method achieves the best performance compared with both classes.

• Overall Superiority: Hunyuan3D 2.5 (labeled "Ours") demonstrates clear superiority across all evaluated metrics: CLIP-FID (23.97), FID (165.8), CMMD (2.064), CLIP-I (0.9281), and LPIPS (0.1231). Lower values are better for CLIP-FID, FID, CMMD, LPIPS, and higher values are better for CLIP-I. This indicates that Hunyuan3D 2.5 generates textures that are more perceptually realistic, diverse, and semantically aligned with input images than all other compared methods.

• PBR Material Generation Challenges: The paper highlights that competing models struggle with PBR material generation, specifically in accurately estimating MR (metallic and roughness) values and decoupling illumination effects from the albedo component. This suggests that the PBR material framework and illumination-invariant consistency loss in Hunyuan3D 2.5 effectively address these issues.

• Visual Comparison (Figure 7): The following figure (Figure 7 from the original paper) shows a visual comparison of different methods in terms of texture generation:

  
  该图像是图7，展示了不同方法在纹理生成方面的视觉对比，包括模型正反面及对应的完整材质贴图和反照率贴图的效果。

  Figure 7: Visual comparison of different methods in terms of texture generation. We compared the front and back of models generated by different methods, as well as the effects of the corresponding complete material maps and albedo maps. Figure 7 provides visual evidence for these claims, showcasing the front and back of generated models, as well as their material maps and albedo maps. Hunyuan3D 2.5 textures appear more coherent and realistic, especially regarding metallic and roughness properties, contributing to a more photorealistic final render.

6.1.3. User Study

A user study was conducted to evaluate human preference for end-to-end textured models generated by Hunyuan3D 2.5 against three latest commercial models. Participants were asked to rank methods for each sample.

The following figure (Figure 8 from the original paper) shows user study against three latest commerical models in terms of end-to-end textured results:

该图像是图表，展示了用户研究中Hunyuan3D 2.5与三款最新商业模型在端到端有纹理3D生成任务（Image to 3D和Text to 3D）中的对比结果，分别以“劣于”“相同”“优于”三类比例表示。

Figure 8: User study against three latest commerical models in terms of end-to-end textured results.

• Significant Win Rate: As shown in Figure 8, for the image-to-3D task, Hunyuan3D 2.5 achieved a remarkable 72% win rate. This is 9 times higher than Commercial Model 1, and significantly superior to the other commercial models. This result strongly validates the perceived quality and realism of Hunyuan3D 2.5's outputs from a human perspective, which is often the ultimate judge for creative assets. The user study underscores the practical value and competitive edge of Hunyuan3D 2.5 in generating high-quality 3D assets.

6.2. Ablation Studies / Parameter Analysis

The paper does not explicitly detail dedicated ablation studies or parameter analyses in a separate section. However, the discussion of the dual-phase resolution enhancement strategy implicitly acts as a form of ablation or design choice justification. The two phases—first, conventional multi-view training, then zoom-in training—demonstrate how the model progressively improves texture-geometry alignment. The choice of $512 \times 512$ for the first phase and $768 \times 768$ during inference (accelerated by UniPC sampler) are key parameters that show an optimized approach to managing computational resources versus detail capture. The discussion about the dual-channel attention mechanism for Multi-Channel Material Generation also highlights a design choice aimed at addressing specific issues (misalignment of attention masks), implying its contribution to the overall performance. The paper states that it "systematically examin[ed] the reference attention module," which suggests internal analysis that led to this design.

7. Conclusion & Reflections

7.1. Conclusion Summary

The paper successfully introduces Hunyuan3D 2.5, an advanced suite of 3D diffusion models, marking a significant step forward in the generation of high-fidelity and detailed textured 3D assets. Its main contributions include the novel LATTICE shape foundation model, which leverages unprecedented scaling in datasets, model size, and compute to generate exceptionally detailed and smooth 3D shapes. Furthermore, the texture generation module has been significantly upgraded to support physical-based rendering (PBR) through a novel multi-view architecture, effectively producing albedo, metallic, and roughness maps with high spatial coherence via a dual-channel attention mechanism. A dual-phase resolution enhancement strategy also contributes to superior texture-geometry alignment. Extensive quantitative evaluations and user studies unequivocally demonstrate that Hunyuan3D 2.5 outperforms state-of-the-art models, both open-source and commercial, in terms of shape fidelity, texture realism, and overall perceived quality. This work provides a powerful tool that significantly closes the gap between automatically generated and handcrafted 3D assets, holding immense value for various industries.

7.2. Limitations & Future Work

The paper itself does not explicitly dedicate a section to limitations or future work. However, based on the presented methodology and results, potential limitations and implicit future directions can be inferred:

• Computational Cost: Training a 10B parameter model like LATTICE on "scaled high-quality datasets" demands immense computational resources. This could be a limitation for researchers or smaller organizations seeking to reproduce or further develop the method without access to similar infrastructure.
• Dataset Availability: The reliance on "scaled high-quality datasets" for training, which are likely proprietary, makes replication and direct comparison challenging for the broader research community.
• Generalization to Novel Categories: While the model is trained on "complex objects," its generalization capabilities to extremely novel or abstract 3D asset categories beyond its training distribution might still be an area for further exploration.
• Real-time Generation/Editing: The paper mentions step distillation for faster inference, but further optimization for real-time applications or interactive 3D asset editing is always a desirable future direction in this field.
• Beyond PBR Maps: While PBR (albedo, metallic, roughness) is a significant step, advanced rendering often involves other maps like normal maps (which are inputs here), displacement maps, ambient occlusion maps, or even volumetric materials. Expanding the generative capabilities to these aspects could be a future enhancement.
• User Control and Editability: The paper focuses on generation from images. Future work could explore more nuanced user control, allowing artists to guide the generation process with sketches, semantic masks, or even direct manipulation of PBR parameters.
• Dynamic/Animated 3D Assets: The current focus is on static 3D assets. Extending high-fidelity generation to animated or deformable 3D models represents a complex but valuable future research avenue.

7.3. Personal Insights & Critique

This paper showcases a clear and impactful advancement in 3D asset generation. My personal insights include:

• The Power of Scaling: Hunyuan3D 2.5 is a strong testament to the "scaling laws" observed in large language models and vision models. The sheer scale of LATTICE (10B parameters) combined with high-quality data and compute directly translates into unprecedented levels of detail and fidelity in 3D shape generation. This reinforces the idea that for truly "ultimate details," brute-force scaling, when done intelligently, is a highly effective strategy.

• PBR as a Necessity for Realism: The dedicated focus on PBR material generation, beyond simple RGB textures, highlights its critical role in achieving photorealistic 3D assets. The illumination-invariant consistency loss and the dual-channel attention mechanism for albedo and MR maps are elegant solutions to ensure physical accuracy and coherence, indicating a deep understanding of rendering principles integrated into the generative process. This sets a new benchmark for what generative 3D models should output.

• Elegant Solutions to Practical Problems: The dual-phase resolution enhancement strategy for texture generation is a clever engineering solution to the practical memory constraints of training high-resolution multi-view models. Instead of simply demanding more VRAM, the zoom-in training approach allows the model to "focus" on details without overwhelming computational resources, making the high-resolution output feasible. This demonstrates a balance between theoretical ambition and practical implementation.

• Metrics vs. Perception: The observation that numerical metrics (ULIP, Uni3D) might not fully capture the model capabilities compared to visual results or user studies is insightful. This highlights the ongoing challenge in 3D generation research: developing quantitative metrics that truly align with human aesthetic and perceptual judgment. The strong performance in the user study, however, validates the practical success of Hunyuan3D 2.5.

• Potential for Democratization: While the training resources are vast, the existence of such a robust model, especially if it becomes accessible through APIs or optimized versions, has immense potential to democratize high-quality 3D content creation, making it available to a wider range of creators who previously relied on expensive and time-consuming manual modeling.

  A potential area for improvement or a critical thought is the lack of open-sourcing the LATTICE model itself, given its 10B parameters, which makes it challenging for external researchers to delve into its inner workings or build upon it directly. While understandable for commercial reasons, this does limit collaborative scientific advancement. Additionally, more details on the "CCM rendered by 3D mesh as geometry conditions" would have been beneficial for a complete understanding of the texture generation inputs. Overall, Hunyuan3D 2.5 represents a significant technical achievement, pushing the boundaries of generative AI for 3D content creation.