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

Di Yang; Zhihao Li; Yanhai Xiong; Yufei Wang

cs.CVOct 11, 2026

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

Authors: Di Yang, Zhihao Li, Yanhai Xiong, Yufei Wang

Organizations: College of William & Mary · SparcAI Inc.

Abstract

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.

Figures & tables

Figure 1: The few-step distribution shift and its correction by CM . From left: input image; the untextured mesh and PBR render of our two-step student without Channel Match, where the sampling-induced shifted shape latent drives the frozen PBR texture model to produce faulty surface materials (dark/blotchy patches); and the mesh and render with CM . For both few-step teacher and student sampling, calibrating the latent removes the texture artifacts, while the geometry remains unchanged. (We consider samples with simple geometry here, where both the two-step teacher and our two-step student already recover structures close to the many-step reference. This controlled setting isolates the effect of latent distribution shift rather than geometry reconstruction errors, showing that the shift is a general issue of few-step sampling and can be effectively corrected by CM .)

Figure 2: Qualitative comparison on TRELLIS.2 demonstration assets. From left: input image, the 12 and 2-step teacher, SCM @2, MDT @2, and ours @2; each block shows the PBR render and untextured mesh. MDT fractures the geometry, SCM leaves surface artifacts, and ours is closest to the teacher.

Prior methods (cited; Toys4K subset, each method’s own protocol)
Method	Steps	Time (s)	FD-Incep $\downarrow$	FD-DINOv2 $\downarrow$	ULIP-I $\uparrow$
LGM ∗ ( Tang et al., 2024 )	–	5	26.31	322.71	–
3DTopia-XL ∗ ( Chen et al., 2025c )	25	5	37.68	437.37	–
Ln3Diff ∗ ( Lan et al., 2025 )	250	8	26.61	357.93	–
TRELLIS ∗ ( Xiang et al., 2025 )	25 $\times$ 2	–	9.35	67.21	–
Our pipeline (TRELLIS.2 backbone; full Toys4K; fixed textured-mesh protocol)

Table 1: Main comparison on Toys4K. ULIP-I is

\times 100

; Fréchet distances are set-level. Starred rows are cited from prior work under each method’s own protocol and are not directly comparable to the lower block. Times are A800-80GB end-to-end image-to-shape latency where measured.

Stage	Teacher	Ours @2
Shape sampling (s/asset)	5.29	1.18 ( $4.5\times$ )
End-to-end image $\to$ shape (s/asset)	11.51	7.55 ( $1.5\times$ )

Table 2: Latency on a single A800-80GB GPU.

Fraction of assets moved toward teacher ( $\uparrow$ )
Calibration	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$
None (two-step student)	0.0802	0.8819	24.44
Std only	0.0762	0.8860	24.72
Full (mean $+$ std)	0.0747	0.8885	24.85
Std only	73%	75%	67%
Full (mean $+$ std)	73%	88%	73%

Table 3: Channel Match on the TRELLIS.2 demonstration assets (

n{=}72

), scored against the twelve-step teacher’s renders under the four-view protocol. Both forms move the two-step student toward the teacher; the full match gives the larger gain on these artifact-heavy assets.

Student @2
Calibration	ULIP-I $\uparrow$	FD-Incep $\downarrow$	FD-DINOv2 $\downarrow$
None	38.48	14.09	127.44
Std only	38.46	13.88	126.88
Full (mean $+$ std)	38.38	14.16	127.88
Teacher @2
None	38.23	15.67	155.40

Table 4: Calibration form at two steps on the full Toys4K set, for both the distilled student and the naively teacher.

$\tau_{S}$	applied	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$
$-\infty$ (always)	100%	0.0747	0.8885	24.85
$0.20$ (default)	77%	0.0742	0.8884	24.89
$0.27$	50%	0.0761	0.8862	24.72
$+\infty$ (never)	0%	0.0802	0.8819	24.44

Table 5: Sensitivity to the gating threshold

\tau_{S}

on the demonstration assets (

n{=}72

, full mean

+

std match), scored by paired LPIPS/SSIM/PSNR against the twelve-step teacher under the four-view protocol.

Config	Steps	Collapse $S$ $\downarrow$	Collapsed chans. $\downarrow$
Teacher	12	0.150	16.8%
Student	4	0.146	17.5%
Student	2	0.174	22.4%
Student	1	0.207	28.4%

Table 6: Latent drift versus sampling steps on Toys4K. Collapse score

S

(Eq. 10 ) is the mean per-channel std deficit; it and the fraction of channels with sampled std below

0.8

Method	Lat. (ms) $\downarrow$	VRAM (GB) $\downarrow$
FlashAttn foward (no JVP) †	4.8	—
Naive foward	114.4	27.8
Naive foward $+$ JVP	454.4	67.6
Triton JVP (ours)	41.7	$\sim$ 0.2

Table 7: Fused JVP attention vs. the naive PyTorch baseline (A800-80GB, bf16,

N{=}15{,}000

). † FlashAttention (forward only, no tangent) is a speed ceiling , not a JVP baseline.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: Additional qualitative comparison on TRELLIS.2 demonstration assets. From left to right: input image, the 12-step teacher, SCM @2, MDT @2, and ours @2. Each block displays the PBR render and the untextured mesh. MDT fractures the geometry, SCM leaves surface artifacts, and our method produces results closest to the teacher.

Figure 4: The effect of Channel Match on the TRELLIS.1 architecture (variable voxel resolution). From left to right: input image; the untextured mesh and render of the two-step student without CM; and the mesh and render with CM. Although TRELLIS.1 does not exhibit the severe dark artifacts seen in TRELLIS.2, the few-step sampling still degrades the overall texture quality. Applying CM effectively restores the texture fidelity by calibrating the shifted shape latent.

Figure 5: High-resolution ( $1536^{3}$ ) shape generation on TRELLIS.2. The figure demonstrates the few-step distribution shift and its correction by Channel Match (CM). From left to right: input image; the untextured mesh and PBR render of the two-step student without CM, where the shifted shape latent causes the frozen texture model to produce faulty surface materials (e.g., dark patches); and the corresponding mesh and render with CM applied. Calibrating the latent removes the texture artifacts while maintaining the high-resolution geometry. We consider samples with relatively simple geometry to isolate the effect of latent distribution shift from geometry reconstruction errors, demonstrating that CM effectively corrects the shift at high resolutions.

3D Texture Generation

See all 19 papers

NOFACE3D: NOise-scheduled Focused Anchoring and Controlled Enhancement for 3D Generation
Oct 9, 2026James Rose, Jingwen Liang, Wenchao Ma +23D Asset Generation
Roblox · The Pennsylvania State University
Flow3D-OPD: Multi-Teacher On-Policy Distillation for 3D Geometry Generation with Flow-Matching Diffusion Transformer
Sep 7, 2026Zhiwei Ning, Zhen Zhou, Puhua Jiang +73D Mesh Generation Diffusion Model Distillation
Shanghai Jiao Tong University · Tencent Hunyuan3D · Zhejiang Normal University +1
Probability-Flow Distillation: Exact Wasserstein Gradient Flow for High-Fidelity 3D Generation
May 9, 2026Rohith Ramanan, A. N. RajagopalanDiffusion Model Distillation Wasserstein Gradient Flows
Department of Physics Indian Institute of Technology Madras · Department of Electrical Engineering Indian Institute of Technology Madras

Oct 9, 2026J/K move · Enter open · S save

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

James Rose, Jingwen Liang, Wenchao Ma, Oren Jacob+1

Roblox · The Pennsylvania State University

3D Asset Generation

Open paper AskPDF arXiv

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/

Abstract

3D Texture Generation

Explore similar work

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

Flow3D-OPD: Multi-Teacher On-Policy Distillation for 3D Geometry Generation with Flow-Matching Diffusion Transformer

Probability-Flow Distillation: Exact Wasserstein Gradient Flow for High-Fidelity 3D Generation