NOFACE3D: NOise-scheduled Focused Anchoring and Controlled Enhancement for 3D Generation

James Rose; Jingwen Liang; Wenchao Ma; Oren Jacob; Haomiao Jiang

cs.CVOct 9, 2026

NOFACE3D: NOise-scheduled Focused Anchoring and Controlled Enhancement for 3D Generation

Authors: James Rose, Jingwen Liang, Wenchao Ma, Oren Jacob, Haomiao Jiang

Organizations: Roblox · The Pennsylvania State University

Abstract

Recent image-to-3D models generate complete, high-quality assets from a single image, with global shape and appearance that are already convincing; what remains is the fine scale, where small, thin, elongated, or detail-rich regions are reconstructed softly and under-resolved. This gap can arise in generators with fixed-resolution spatial grids: the same spatial sampling density is used for a large, smooth, low-curvature surface and for a thin, high-curvature structure that carries most of the perceived detail. To this end, we introduce NOFACE3D, a training-free framework that refines these detail-critical regions while preserving the rest of the asset. Given a generated mesh, its input image, and a refinement box, NOFACE3D extracts the selected submesh, scales it up to fill the generator's input volume, and re-encodes it into the full-resolution latent grid. Conditioned on the image crop, it refines the region through flow-based sampling with a spatially varying noise schedule that anchors boundary latents. The refined region is then integrated back into the original asset, preserving the surrounding geometry and appearance. We implement NOFACE3D on TRELLIS.2 and its sparse voxel representation; all of it operates purely at inference time on the frozen pretrained generator. We evaluate NOFACE3D on PAniC-3D, TexVerse and BigDetail, a new benchmark of 339 assets with annotated detail-critical regions. On all three, it generates sharper geometry and finer texture detail than the baselines. Project page: https://idkcouscous.github.io/NoFACE3D/

Figures & tables

Figure 2: Overview of NOFACE3D . Given an asset generated by an off-the-shelf image-to-3D model and a user-marked region, we scale up the region to fill the generator’s input volume and re-encode it into the full-resolution latent grid, assign each token a release level $\tau$ from a boundary-distance field (anchored near attachment boundaries, with stronger regeneration farther away), and resample geometry and texture with spatially varying SDEdit ( Meng et al., 2022 ) , re-anchoring unreleased tokens to the base asset at every step. The refined mesh replaces the base faces inside the crop box.

	Geometry					Appearance					GPT-5.6
	CD	F1 0.25%	F1 1%	IoU	SNE	PSNR	SSIM	LPIPS	DINO	FID	Ovr.	Geo.
Method	$\downarrow$	$\uparrow$	$\uparrow$	$\uparrow$	$\downarrow$	$\uparrow$	$\uparrow$	$\downarrow$	$\uparrow$	$\downarrow$	$\uparrow$	$\uparrow$
PAniC-3D ( $N{=}329$ )
GT	reference										$8.52$	$6.83$
TRELLIS.2	$\mathbf{2.938}$	$0.138$	$0.389$	$0.383$	$1.12$	$15.80$	$0.882$	$0.118$	$0.792$	$43.17$	$6.48$	$6.82$
Easy3E	$3.955$	$0.074$	$0.265$	$0.294$	$1.24$	$13.17$	$0.878$	$0.128$	$0.588$	$98.74$	$2.44$	$3.02$

Table 1: Comparison with refinement and editing methods. Whole-asset metrics against GT; CD and SNE

{\times}100

; IoU at

128^{3}

; FID is set-level. GPT-5.6: pairwise win share (1–10, chance

5.0

) for overall quality and geometry detail.

Figure 3: Qualitative comparison on PAniC-3D (top) and TexVerse (bottom): texture and normal map per method (DetailGen3D is untextured).

Table 4

Figure 4: Qualitative results on BigDetail . Top: input image with the labeled region and its enlargement. Below: TRELLIS.2 and NOFACE3D , texture and normal map.

Figure 5: Noise scheduling ablation. Top: regeneration without anchoring or noise scheduling. Bottom: with noise scheduling. Texture and normal views use matched cameras and detail crops. Orange boxes mark the regions enlarged below each full-asset view.

Figure 6: Localized 3D editing. A Genshin character gains a hat (head region), then robotic legs (leg region). Dashed boxes mark the regions.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Metric	TRELLIS.2	NOFACE3D
Chamfer $\downarrow$	$\mathbf{3.359\pm 2.764}$	$3.370\pm 2.773$
F1 @ $0.25\%$ $\uparrow$	$0.141\pm 0.070$	$\mathbf{0.144\pm 0.070}$
F1 @ $1\%$ $\uparrow$	$0.37219\pm 0.15536$	$0.37222\pm 0.15548$
Surface IoU $128^{3}$ $\uparrow$	$0.373\pm 0.140$	$\mathbf{0.381\pm 0.142}$
SNE $\downarrow$	$\mathbf{1.312\pm 0.610}$	$1.314\pm 0.620$
PSNR $\uparrow$	$15.70\pm 2.17$	$\mathbf{16.53\pm 2.76}$

Appendix

Table 4: TRELLIS.2 versus NOFACE3D on

4{,}892

PAniC-3D assets against artist GT. Mean

\pm

population s.d.; Chamfer and SNE

{\times}100

; FID is set-level.

	GPT-4o			GPT-5.6
Criterion	GT	TRELLIS.2	NOFACE3D	GT	TRELLIS.2	NOFACE3D
Fidelity / completeness	$5.68$	$4.41$	$\mathbf{4.91}$	$6.38$	$3.61$	$\mathbf{5.01}$
3D plausibility	$5.63$	$4.64$	$\mathbf{4.73}$	$5.87$	$4.33$	$\mathbf{4.80}$
Geo.–tex. alignment	$5.74$	$4.44$	$\mathbf{4.82}$	$4.94$	$4.67$	$\mathbf{5.40}$
Texture details	$6.05$	$4.08$	$\mathbf{4.86}$	$7.46$	$2.42$	$\mathbf{5.12}$
Geometry details	$5.64$	$4.63$	$\mathbf{4.73}$	$5.29$	$4.75$	$\mathbf{4.96}$

Appendix

Table 5: Monadic judged comparison on

4{,}892

PAniC-3D assets: each mesh is rated alone, and win shares (1–10) are derived from the ratings.

Metric	TRELLIS.2	Easy3E	Elevate3D	DetailGen3D	NOFACE3D
Chamfer $\downarrow$	$\mathbf{2.938\pm 1.884}$	$3.955\pm 2.071$	$3.059\pm 1.896$	$2.939\pm 1.900$	$2.939\pm 1.878$
F1 @ $0.25\%$ $\uparrow$	$0.138\pm 0.073$	$0.074\pm 0.034$	$0.115\pm 0.051$	$0.127\pm 0.056$	$\mathbf{0.142\pm 0.074}$
F1 @ $1\%$ $\uparrow$	$0.389\pm 0.166$	$0.265\pm 0.111$	$0.372\pm 0.146$	$\mathbf{0.393\pm 0.153}$	$0.389\pm 0.165$
Surface IoU $128^{3}$ $\uparrow$	$0.383\pm 0.144$	$0.294\pm 0.113$	$0.377\pm 0.135$	$0.386\pm 0.141$	$\mathbf{0.391\pm 0.147}$
SNE $\downarrow$	$1.12\pm 0.40$	$1.24\pm 0.37$	$1.26\pm 0.38$	$\mathbf{1.04\pm 0.40}$	$1.12\pm 0.40$
PSNR $\uparrow$	$15.80\pm 1.90$	$13.17\pm 1.08$	$16.59\pm 1.76$	–	$\mathbf{16.78\pm 2.44}$

Appendix

Table 6: Objective comparison on PAniC-3D (

N{=}329

). Mean

\pm

s.d.; Chamfer and SNE

{\times}100

; FID is set-level.

Metric	TRELLIS.2	Easy3E	Elevate3D	DetailGen3D	NOFACE3D
Chamfer $\downarrow$	$7.27\pm 14.9$	$11.7\pm 19.3$	$13.7\pm 18.7$	$35.4\pm 31.0$	$\mathbf{7.22\pm 14.8}$
F1 @ $0.25\%$ $\uparrow$	$0.123\pm 0.147$	$0.060\pm 0.059$	$0.058\pm 0.064$	$0.009\pm 0.010$	$\mathbf{0.140\pm 0.150}$
F1 @ $1\%$ $\uparrow$	$0.306\pm 0.230$	$0.219\pm 0.188$	$0.194\pm 0.178$	$0.032\pm 0.031$	$\mathbf{0.320\pm 0.234}$
Surface IoU $128^{3}$ $\uparrow$	$0.334\pm 0.247$	$0.250\pm 0.193$	$0.198\pm 0.169$	$0.055\pm 0.044$	$\mathbf{0.337\pm 0.244}$
SNE $\downarrow$	$1.97\pm 1.33$	$1.81\pm 1.37$	$2.16\pm 1.40$	$\mathbf{1.23\pm 0.84}$	$1.61\pm 1.18$
PSNR $\uparrow$	$18.02\pm 9.24$	$17.25\pm 9.10$	$16.97\pm 9.15$	–	$\mathbf{19.57\pm 9.29}$

Appendix

Table 7: Objective comparison on TexVerse (

N{=}151

). Mean

\pm

s.d.; Chamfer and SNE

{\times}100

Criterion	GT	TRELLIS.2	Easy3E	Elevate3D	DetailGen3D	NOFACE3D
GPT-4o
Fidelity / completeness	$7.84$	$6.41$	$2.17$	$4.73$	–	$\mathbf{7.57}$
3D plausibility	$7.17$	$5.77$	$2.76$	$2.96$	$4.54$	$\mathbf{6.81}$
Geo.–tex. alignment	$7.90$	$6.45$	$2.41$	$4.39$	–	$\mathbf{7.60}$
Texture details	$7.87$	$6.46$	$1.95$	$4.93$	–	$\mathbf{7.61}$
Geometry details	$7.36$	$5.88$	$2.23$	$3.04$	$4.50$	$\mathbf{6.98}$

Appendix

Table 8: Pairwise judged comparison on PAniC-3D (

N{=}329

). Win share (1–10, chance

5.0

Criterion	GT	TRELLIS.2	Easy3E	Elevate3D	DetailGen3D	NOFACE3D
GPT-4o
Fidelity / completeness	$5.58$	$5.24$	$5.34$	$5.07$	–	$\mathbf{5.72}$
3D plausibility	$5.61$	$5.21$	$5.41$	$5.01$	$3.12$	$\mathbf{5.64}$
Geo.–tex. alignment	$5.59$	$5.23$	$5.38$	$5.05$	–	$\mathbf{5.66}$
Texture details	$5.62$	$5.25$	$5.34$	$5.03$	–	$\mathbf{5.71}$
Geometry details	$5.58$	$5.22$	$5.39$	$5.04$	$3.12$	$\mathbf{5.65}$

Appendix

Table 9: Pairwise judged comparison on TexVerse (

N{=}151

). Win share (1–10, chance

5.0

	GPT-4o		GPT-5.6
Criterion	TRELLIS.2	NOFACE3D	TRELLIS.2	NOFACE3D
Fidelity / completeness	$4.40$	$\mathbf{5.60}$	$3.76$	$\mathbf{6.24}$
3D plausibility	$4.42$	$\mathbf{5.58}$	$4.60$	$\mathbf{5.40}$
Geo.–tex. alignment	$4.40$	$\mathbf{5.60}$	$3.89$	$\mathbf{6.11}$
Texture details	$4.57$	$\mathbf{5.43}$	$4.61$	$\mathbf{5.39}$
Geometry details	$4.60$	$\mathbf{5.40}$	$4.21$	$\mathbf{5.79}$

Appendix

Table 10: Pairwise judged comparison on BigDetail (

N{=}339

), TRELLIS.2 versus NOFACE3D . Head-to-head win share (1–10); each pair sums to

10

3D Asset Generation

3D asset generation creates and edits objects, characters, materials, and scenes from text, images, video, or geometric constraints.

See all 161 papers

Oct 11, 2026

Channel Match: Calibrating Shape Latents for Few-Step 3D Generation

Recent native 3D generators achieve high-quality asset creation by modeling structured 3D latents with flow-based transformers, but their high-resolution generation remains computationally expensive due to costly iterative sampling. In this work, we introduce a few-step distillation framework for native 3D generation, enabled by continuous self-consistency training with a custom JVP FlashAttention kernel and an analytic velocity regression objective for sparse structured 3D latents. To stabilize joint optimization, we further introduce an adaptive loss weighting strategy that automatically balances them without manual tuning. Beyond acceleration, we reveal a previously overlooked failure mode of few-step native 3D generation: sampling-induced latent distribution drift preserves geometric correctness but disrupts downstream PBR texture synthesis, causing texture artifacts on geometrically valid assets. To address this issue, we propose Channel Match, a training-free adaptive per-channel latent calibration method that restores the latent distribution without modifying either generator. We validate our approach on TRELLIS.2 at the

1024^3

resolution setting. Our two-step model achieves

4.5\times

faster geometry sampling while maintaining geometry quality comparable to the many-step teacher, and Channel Match further improves rendering quality by recovering the latent interface between geometry and appearance generation.

Di Yang, Zhihao Li, Yanhai Xiong +1

College of William & Mary · SparcAI Inc.

3D Texture Generation 3D Asset Generation 3D Shape Generation Consistency Distillation

Ask PDF

Sep 28, 2026

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

Scaling image-to-3D generation to ultra-high resolutions requires controlling rapidly growing computational costs without sacrificing fine geometric detail. We present 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.

Hongjie Li, Xinran Yang, Xiuchao Wu +2

AlibabaGroup

Flow Matching 3D ViTs 3D Shape Generation Single-View 3D Reconstruction

Ask PDF

Mar 31, 2026

Extend3D: Town-Scale 3D Generation

In this paper, we propose Extend3D, a training-free pipeline for 3D scene generation from a single image, built upon an object-centric 3D generative model. To overcome the limitations of fixed-size latent spaces in object-centric models for representing wide scenes, we extend the latent space in the

x

and

y

directions. Then, by dividing the extended latent space into overlapping patches, we apply the object-centric 3D generative model to each patch and couple them at each time step. Since patch-wise 3D generation with image conditioning requires strict spatial alignment between image and latent patches, we initialize the scene using a point cloud prior from a monocular depth estimator and iteratively refine occluded regions through SDEdit. We discovered that treating the incompleteness of 3D structure as noise during 3D refinement enables 3D completion via a concept, which we term under-noising. Furthermore, to address the sub-optimality of object-centric models for sub-scene generation, we optimize the extended latent during denoising, ensuring that the denoising trajectories remain consistent with the sub-scene dynamics. To this end, we introduce 3D-aware optimization objectives for improved geometric structure and texture fidelity. We demonstrate that our method yields better results than prior methods, as evidenced by human preference and quantitative experiments.

Seungwoo Yoon, Jinmo Kim, Jaesik Park

Seoul National University

Image-to-3D Scene Generation 3D Scene Generation 3D Diffusion Models 3D Scene Completion

Ask PDF

Abstract

3D Asset Generation

Explore similar work

Channel Match: Calibrating Shape Latents for Few-Step 3D Generation

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

Extend3D: Town-Scale 3D Generation