FILIGREE3D: Scaling Sparse Latent Flow Matching for Ultra-High-Resolution Image-to-3D Generation

Hongjie Li; Xinran Yang; Xiuchao Wu; Jiangjing Lyu; Chengfei Lv

cs.CVSep 28, 2026

FILIGREE3D: Scaling Sparse Latent Flow Matching for Ultra-High-Resolution Image-to-3D Generation

Authors: Hongjie Li, Xinran Yang, Xiuchao Wu, Jiangjing Lyu, Chengfei Lv

Organizations: AlibabaGroup

Abstract

Scaling image-to-3D generation to ultra-high resolutions requires controlling rapidly growing computational costs without sacrificing fine geometric detail. We present \textbf{Filigree3D}, a sparse latent flow-matching framework that generates 3D geometry from a single image at voxel resolutions up to $2048^3$ , with straightforward extensibility to $4096^3$ . To make training tractable, we introduce Structure-Aware Sparse Scaling, which combines spatial bounding with alternating local-global attention to constrain token growth while preserving both fine-scale details and long-range structural context. To enhance detail reconstruction, we curate training samples based on their high-resolution geometric gains and inject multi-scale image features into a sparse 3D DiT, effectively coupling structural semantics with fine-grained visual cues. Furthermore, a visibility-aware voxel regularization strategy improves robustness against sparse perturbations and facilitates the completion of unobserved geometry. Under our default configuration, Filigree3D maintains peak GPU memory consumption within practical limits for contemporary hardware, enabling the generation of highly intricate 3D geometry in approximately one minute. Extensive experiments demonstrate that our method yields substantial improvements in overall geometric fidelity and fine-detail preservation compared to existing baselines, validating practical, detail-preserving 3D generation at unprecedented resolutions.

Figures & tables

Figure 2: Overview of Filigree3D. The framework combines (A) structure-aware sparse scaling for budgeted processing, (B) block-specific multi-scale condition injection for hierarchical image guidance, and (C) resolution-gain curation for resolution-sensitive samples.

Figure 3: Qualitative comparison on Toys4K dataset.

Method	T (s) $\downarrow$	Toys4K					FineGeo3D
Method	T (s) $\downarrow$	CD $\downarrow$	F1-0.01 $\uparrow$	F1-0.002 $\uparrow$	ULIP-2 $\uparrow$	Uni3D $\uparrow$	CD $\downarrow$	F1-0.01 $\uparrow$	F1-0.002 $\uparrow$	ULIP-2 $\uparrow$	Uni3D $\uparrow$
Direct3D-S2	69.2	0.0339	0.5062	0.0596	0.2293	0.2785	0.0326	0.4917	0.0499	0.2184	0.2783
Hy3D-2.1	17.1	0.0233	0.6683	0.1011	0.2390	0.2797	0.0278	0.5809	0.0674	0.2330	0.2879
TRELLIS.2	59.5	0.0270	0.6315	0.0947	0.2275	0.2807	0.0274	0.5847	0.0639	0.2239	0.2716
Pixal3D	26.2	0.0360	0.5542	0.0888	0.2080	0.2775	0.0345	0.5448	0.0674	0.2115	0.2717
Ours-1536	50.9	0.0058	0.9744	0.4616	0.2624	0.2965	0.0054	0.9963	0.3679	0.2665	0.3042

Table 1: Quantitative comparison on Toys4K and FineGeo3D dataset. T means inference time.

Figure 4: Qualitative comparison on FineGeo3D dataset.

Method	CD $\downarrow$	F1-0.01 $\uparrow$	F1-0.002 $\uparrow$	ULIP-2 $\uparrow$	Uni3D $\uparrow$
w/o GIF	0.0053	0.9865	0.4121	0.2605	0.3077
w/o VR	0.0052	0.9863	0.4062	0.2610	0.3092
w/o MSF	0.0072	0.9662	0.3845	0.2407	0.2712
w/o DC	0.0060	0.9765	0.3941	0.2509	0.2915
Full Pipeline	0.0049	0.9969	0.4309	0.2760	0.3126

Table 2: Ablation study on FineGeo3D dataset. GIF, VR, MSF, and DC denote global information flow, voxel regularization, multi-scale features and data curation, respectively.

Figure 5: Resolution Scaling and Commercial Comparison. Left: Filigree3D recovers increasingly fine geometry from $1536^{3}$ to $4096^{3}$ . Right: comparison with leading commercial systems at their highest-quality settings. Red boxes highlight fine geometric details.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Value	Setting	Value
Input image resolution	$1024\times 1024$	Local window width $w$	32
Image encoder	DINOv2 ViT-L/16	Global communication interval $K$	6 blocks
Feature dimension per layer	1,024	Coarse pooling stride $p$	8
DINOv2 feature layers	5, 7, 11, 23	Positional encoding	RoPE
SLat channels	32	Attention backend	FlashAttention-3
SLat / geometry resolution	$128^{3}$ / $2048^{3}$	Optimizer	AdamW

Appendix

Table 3: Configuration of the sparse latent flow-matching denoiser.

Figure 6: More qualitative comparisons on FineGeo3D dataset. Our method reconstructs finer geometric structures and preserves intricate details more faithfully than existing methods.

Figure 7: Additional $2048^{3}$ generation results. Our FineGeo3D recovers fine-grained geometric details while preserving coherent global structures across diverse objects.

Explore similar work

FLUX3D: High-Fidelity 3D Gaussian Generation with Diffusion-Aligned Sparse Representation
Jun 23, 2026Haorui Ji, Weizhe Liu, Hongdong Li +13D Gaussian 3D Generative Models
The Australian National University · ByteDance
SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation
Oct 1, 2026Tianjiao Yu, Xinzhuo Li, Yifan Shen +43D Generation Video Latents
University of Illinois Urbana-Champaign
GeoWorld: Providing Full-frame Geometry Features to Facilitate 3D Scene Generation
Nov 28, 2025Yuhao Wan, Lijuan Liu, Jingzhi Zhou +63D Scene Generation Generative Video Models
1VCIP & AAIS, Nankai University · 2ByteDance Inc. · 3Renmin University of China +2

Jun 23, 2026, cs.CVJ/K move · Enter open · S save

FLUX3D: High-Fidelity 3D Gaussian Generation with Diffusion-Aligned Sparse Representation

Haorui Ji, Weizhe Liu, Hongdong Li, Hengkai Guo

The Australian National University · ByteDance

3D Gaussian 3D Generative Models Voxel 3D Generation Diffusion Alignment Diffusion Transformers+1

Open paper AskPDF arXiv

Sparse voxel representation has emerged as a scalable foundation for image-to-3D Gaussian Splatting (3DGS) generation, yet current methods struggle to preserve high-frequency visual details of input images due to two structural bottlenecks. First, they adopt discriminative 2D features optimized for semantic abstraction to construct sparse voxel latents, which suppress reconstructive cues and induce a representation bottleneck. Second, in the generation stage, standard diffusion transformers lack effective mechanisms to align dense 2D image tokens with sparse 3D voxel latents, resulting in a cross-modal correspondence bottleneck. To address these issues, we propose FLUX3D, a scalable image-to-3DGS framework that boosts both representation learning and cross-modal alignment during generation. We first revisit 2D feature selection for sparse-voxel-based 3D representation learning, propose Diffusion-Aligned Structured Latents (DA-SLAT) and couple it with a decoder-only architecture to improve 3DGS reconstruction fidelity. We also design a sparse-structure-aware diffusion framework, which integrates the Sparse-structure Multimodal Diffusion Transformer (SMDiT) and Modal-Aware Rotary Positional Embedding (MARoPE) to achieve geometry-agnostic 2D-3D alignment. Extensive benchmark experiments demonstrate that FLUX3D yields substantial improvements in appearance fidelity and significantly outperforms all state-of-the-art (SOTA) methods in generating high-quality 3DGS assets.

Abstract

Explore similar work

FLUX3D: High-Fidelity 3D Gaussian Generation with Diffusion-Aligned Sparse Representation

SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation

GeoWorld: Providing Full-frame Geometry Features to Facilitate 3D Scene Generation