Hunyuan3D 2.5: Towards High-Fidelity 3D Assets Generation with Ultimate Details

1 Introduction

• 3D generation is increasingly used across industries, with automated modeling and texturing helping simplify creative work and improve production efficiency.Source: §1 · p. 2
• Existing 3D shape models struggle to generate fine-grained detail while maintaining smooth surfaces and sharp edges.The introduction frames this as an open problem for high-fidelity shape generation.Source: §1 · p. 2
• Multiview texture generation still faces consistency problems that can cause artifacts and seams, while open-source PBR material generation remains unavailable.The introduction also notes that traditional RGB textures do not meet photorealistic 3D asset demands.Source: §1 · p. 2

SHAPE GENERATION

• LATTICE, the new shape foundation model, scales across high-quality datasets, model size, and compute, with its largest version producing sharp, detailed, image-aligned shapes with clean, smooth surfaces.The report says these properties significantly narrow the gap between generated and handcrafted 3D shapes.Source: SHAPE GENERATION · p. 2
• The upgraded texture framework extends Hunyuan3D 2.0 (Zhao et al., 2025) and 2.1 (Hunyuan3D et al., 2025), simultaneously generating multi-view albedo, roughness, and metallic maps under a principled BRDF model.The maps describe surface reflection properties and simulate geometric microsurface distributions.Source: SHAPE GENERATION · p. 2
• The two-stage pipeline first generates a 3D shape, then creates its texture based on that shape.Source: SHAPE GENERATION · p. 2, Figure 3 · p. 3
• A dual-phase resolution enhancement strategy improves texture-geometry coordination and end-to-end visual quality.Source: SHAPE GENERATION · p. 3
• Evaluations across diverse in-the-wild images show Hunyuan3D 2.5 outperforms state-of-the-art open-source and closed-source commercial models in shape and end-to-end texture generation.Source: SHAPE GENERATION · p. 3

2 Method

• The pipeline processes an input image, generates an untextured mesh, extracts geometry features such as normals and UVs, then generates texture from those outputs.It follows the overall architecture of Hunyuan3D 2.0.Source: §2 · p. 3

2.1 Detailed Shape Generation

• LATTICE generates detailed shapes from a single image or four multi-view images, balancing sharp edges and smooth surfaces while approaching handcrafted detail.The model is trained on an extensive, high-quality 3D dataset; guidance and step distillation reduce inference time. Examples include accurate fingers, bicycle-wheel patterns, and a bowl within a larger scene.Source: §2.1 · pp. 3–4

2.2 Realistic Texture Generation

• The material framework uses mesh-rendered normal maps and CCM plus a reference image to generate PBR maps, with 3D-aware RoPE supporting cross-view consistency.The framework extends the multi-view PBR architecture of Hunyuan3D 2.1 (Hunyuan3D et al., 2025) and inherits 3D-aware RoPE from (Feng et al., 2025).Source: §2.2 · p. 4
• Separate learnable embeddings model albedo, metallic-roughness, and normal maps, while shared attention masks align albedo and metallic-roughness features spatially.The basecolor branch supplies the shared mask for the other branches, and an illumination-invariant consistency loss encourages disentanglement of material properties from illumination.Source: §2.2 · pp. 4–5
• A two-phase resolution strategy first trains on six-view 512×512 images, then uses random zoom-ins to learn fine texture details without full-resolution multi-view training.The zoom-in phase addresses memory constraints while preserving the multi-view benefits established in phase one.Source: §2.2 · p. 5

3.1 Shape Generation

• Shape evaluation compares open-source methods Michelangelo (Zhao et al., 2024), Craftsman 1.5, Trellis (Xiang et al., 2024), and Hunyuan3D-2 (Zhao et al., 2025), plus two commercial models.ULIP and Uni3D measure similarity to input images (ULIP-I, Uni3D-I) and vision-language-model-synthesized image prompts (ULIP-T, Uni3D-T).Source: §3.1 · p. 5

3.2 Texture Generation

• Texture comparisons include Text2Tex (Chen et al., 2023a), Paint3D (Zeng et al., 2024a), Paint-it (Youwang et al., 2024), SyncMVD (Liu et al., 2024a), and TexGen (Yu et al., 2024).Texture metrics are FID, CLIP-FID, LPIPS, CMMD for diversity and detail richness, and CLIP-I for semantic alignment with input images.Source: §3.2 · pp. 7–8

3.1 Shape Generation

• Hunyuan3D 2.5 achieved the best shape similarities on ULIP-T, Uni3D-T, and Uni3D-I, and was judged visually better than the other compared models.The paper cautions that the reported metrics do not fully reflect model capabilities.Source: §3.1 · p. 5, Figure 6 · p. 6, Table 1 · p. 8

3.2 Texture Generation

• Table 2 reports that Hunyuan3D 2.5 performs best against both text-conditioned and image-conditioned texture-generation methods.End-to-end visual comparisons highlight competing models’ difficulty estimating metallic and roughness values and separating input-image illumination from albedo.Source: Figure 7 · p. 7, Table 2 · p. 8, §3.2 · p. 8
• 72% was Hunyuan3D 2.5’s win rate in the image-to-3D user study, 9 times Commercial Model 1’s rate.Participants ranked methods across a diverse test set of real-world images; the study compared Hunyuan3D 2.5 with three commercial models.Source: §3.2 · p. 8

Shape Generation

• Shape-generation work spans early category-specific models, score-distillation text-to-3D, single-step feedforward generation, and more recent native 3D diffusion using 3D data.Examples include early generators (Wu et al., 2016; Sanghi et al., 2022; Yan et al., 2022; Yin et al., 2023), score distillation (Poole et al., 2023), feedforward LRM (Hong et al., 2023), and native models Michelangelo (Zhao et al., 2024) and CLAY (Zhang et al., 2024b).Source: §4 · p. 8

Texture Synthesis

• Texture synthesis uses cross-view attention to address global consistency, while PBR methods generate material priors, retrieve material graphs, or optimize initial textures.Zero123++ spatially concatenates multiview images for self-attention, while other methods use attention masks; PBR approaches recover properties through rendering, adapt library material graphs, or refine textures with techniques such as Score-Distillation Sampling (Poole et al., 2023).Source: §4 · p. 9

• Hunyuan3D 2.5 is an advanced suite of 3D diffusion models for generating high-quality, detailed 3D assets.Source: §5 · p. 9
• A new shape foundation model and physical-based rendering (PBR) for texture generation improve shape fidelity and texture realism.Source: §5 · p. 9
• Extensive evaluations report that Hunyuan3D 2.5 outperforms current state-of-the-art models in shape detail, surface smoothness, and texture consistency.Source: §5 · p. 9