InstantHand: Feed-Forward 3D Gaussian Splatting for Animatable Hand Avatar Reconstruction from a Single Image

Chanho Kim; Muhammad Salman Ali; Kyeonghwan Gwak; Muhammad Shaheryar; Incheol Park; Jeongwan On; Seungryul Baek

cs.CVNov 5, 2026

InstantHand: Feed-Forward 3D Gaussian Splatting for Animatable Hand Avatar Reconstruction from a Single Image

Authors: Chanho Kim, Muhammad Salman Ali, Kyeonghwan Gwak, Muhammad Shaheryar, Incheol Park, Jeongwan On, Seungryul Baek

Organizations: UNIST, South Korea

Abstract

Hand avatar reconstruction typically relies on multi-view capture or per-subject optimization, limiting its scalability. We introduce InstantHand, a feed-forward framework that constructs an animatable 3D Gaussian hand avatar from a single image using off-the-shelf MANO, camera, and mask estimates, without test-time optimization. Image-averaged losses provide insufficient supervision for alpha leakage, which thickens thin finger contours. To address this issue, we propose a Soft Alpha Band regularizer that concentrates supervision on silhouette boundaries. Furthermore, colors regressed from image features can drift toward a learned prior even in observed regions. Our Reliability-gated Appearance Transfer preserves observed appearance by transferring source colors according to each Gaussian's rendering contribution, while completing unobserved regions using the learned prior conditioned on global source identity. InstantHand constructs an avatar in 2.84 seconds and renders at 161 FPS, matching per-subject-optimized baselines on InterHand2.6M and remaining competitive in zero-shot evaluation on HanCo.

Figures & tables

Figure 2: Overview of InstantHand. A frozen DINOv3 encoder provides image features

\mathbf{F}_{s}

that serve as keys and values for a conditional transformer, whose query tokens are point-embedded MANO-surface anchors obtained from the MANO layer

(\boldsymbol{\beta},\boldsymbol{\theta}_{\mathrm{mean}})

. The resulting geometry and appearance tokens are decoded by scale, offset, and color heads into a canonical Gaussian hand

\mathbf{G}^{can}

with fixed rotation

\mathbf{r}_{0}

and opacity

\alpha_{0}

. Reliability-gated Appearance Transfer then composes local source evidence and a global source identity under the source mask

M_{s}

to update only the Gaussian colors (

\mathbf{c}^{0}_{i}\!\to\!\mathbf{c}^{*}_{i}

), and the avatar is animated by MANO-driven deformation

\mathcal{W}

and rendered by

\mathcal{R}

Figure 3: Reliability-gated Appearance Transfer. The Stage-1 appearance prior is refined with Gaussian footprint-aligned local evidence (Prior Refinement) and conditioned on the global source identity, which also generates a source-conditioned prior bank (Identity Conditioning). The composed representation is decoded into an observed color, an appearance gate, and a source-level shift (Appearance Prediction), and Appearance Composition blends these with the base colors under the reliability

\gamma_{i}

to produce the final Gaussian colors

\mathbf{c}^{*}_{i}

. Dashed arrows denote conditioning. Three further inputs of Eq. 11 are omitted for clarity. The source identity also enters Prior Refinement, and the local evidence and base colors also enter Identity Conditioning.

Method	Protocol	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
HARP [ 22 ]	Per-subj. opt.	19.82	0.761	0.225
LiveHand [ 21 ]	Per-subj. opt.	23.01	0.763	0.186
HandAvatar [ 20 ]	Per-subj. opt.	23.79	0.820	0.178
Handy [ 46 ]	Per-subj. opt.	25.56	0.794	0.150
OHTA [ 23 ]	Per-subj. opt.	26.11	0.864	0.129
BIGS ∗ [ 47 ]	Per-subj. opt.	20.46	0.749	0.212

Table 1: Quantitative comparison on InterHand2.6M. Best, second-best, and third-best are marked bold green , underlined blue , and italic orange .

Method	Protocol	Reconstruction			Efficiency
		PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Time (s) $\downarrow$	FPS $\uparrow$	Mem. (GB) $\downarrow$
HandAvatar	Per-subj. opt.	20.32	0.841	0.198	5579	0.32	14.80
OHTA	Per-subj. opt.	20.06	0.843	0.186	3360	0.26	47.80
BIGS ∗	Per-subj. opt.	17.71	0.731	0.246	1380	43.11	1.50
Ours	Zero-shot feed-forward	20.18	0.836	0.188	2.84	160.71	1.79

Table 2: Single-image reconstruction and efficiency on HanCo. Time is the avatar-acquisition cost; FPS and memory are measured for

256\times 256

rendering (protocol in Appendix C.6 ). Best in bold .

Panel of Figure 4 — Figure 4: Qualitative comparison with prior methods on InterHand2.6M [ 48 ] (top three rows) and HanCo [ 55 ] (bottom two rows). Columns: GT, HandAvatar, OHTA, BIGS ∗ , Ours; the black box marks the input image.

Figure 5: Avatars reconstructed from fully estimated inputs (SAM 3 mask, WiLoR MANO and camera) for original and edited source images. Top: an unseen InterHand2.6M test subject, with hand-drawn tattoos composited onto the source image. Bottom: an in-the-wild image released with OHTA [ 23 ] , with its skin tone darkened by an image generation model. Edited rows reuse the mask, MANO, and camera of the original row, and each column shows the same pose and view.

Setting	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
w/o SAB ( $\mathcal{L}_{\mathrm{band}}$ )	23.89	0.825	0.166
w/o reliability gating ( $\gamma$ )	23.18	0.829	0.171
w/o appearance gate ( $g_{i}$ )	24.52	0.833	0.152
w/o appearance transfer	26.01	0.865	0.134
w/o synthetic data	26.13	0.864	0.134
Ours (GT cues)	26.21	0.866	0.133

Table 3: Ablation and source-cue robustness on InterHand2.6M. Ablation rows remove one component and retrain both stages. Estimated-cue rows keep the trained full model and replace its source inputs with estimates.

Configuration	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Bound. F $\uparrow$	Leakage $\downarrow$	Int. cov. $\uparrow$
Ours (constrained + SAB)	26.21	0.866	0.133	0.698	0.234	0.999
Constrained, w/o SAB	23.89	0.825	0.166	0.366	0.555	1.000
Head-recipe port [ 14 ]	23.60	0.809	0.182	0.299	0.608	1.000
Body-recipe port [ 17 ]	23.36	0.805	0.182	0.267	0.627	1.000

Table 4: Controlled comparison with feed-forward avatar recipes on InterHand2.6M. Each recipe swaps its template (FLAME, SMPL-X) for MANO and keeps its published parameterization under otherwise matched training (Appendix E.2 ); boundary metrics, distinct from the training bands, follow Appendix C.4 .

Appendix figures & tables27 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1: Examples of real (top) and synthetic (bottom) training data. Synthetic samples are used only in the source-conditioned appearance stage (Sec. B.2 ).

Stage	Dataset	Weight	Source stride	Targets
Canonical	PALM	0.45	1	1 / 1 / 4
Canonical	InterHand2.6M	0.30	1	1 / 2 / 3
Canonical	HO3Dv3	0.10	10	1 / 3 / 2
Canonical	Goliath	0.10	5	1 / 5 / 0
Canonical	HARP	0.05	5	1 / 5 / 0
Appearance	PALM	0.35	1	1 / 1 / 4

Appendix

Table S1: Dataset sampling policies used for training. Targets per source image are reported as self / same-group / cross-group.

Dataset	Setting	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
InterHand2.6M	w/o transfer	26.58	0.869	0.133
	Ours	28.54	0.903	0.108
HanCo (zero-shot)	w/o transfer	20.47	0.839	0.187
	Ours	23.88	0.898	0.153

Appendix

Table S2: Self-reconstruction with geometry fixed (Sec. C.5 ). w/o transfer renders the Stage-1 base colors, and Ours renders the composed colors.

Stage	Component	Time
Upstream estimation	SAM 3 foreground mask [ 54 ]	0.38 s
	WiLoR MANO $+$ camera [ 37 ]	0.08 s
	subtotal	0.46 s
Avatar construction	Canonical Gaussian construction	1.38 s
	Source-conditioned appearance composition	1.46 s
	subtotal (Table 2 of the main paper)	2.84 s

Appendix

Table S3: Raw-image-to-avatar runtime on a single NVIDIA A100 (80 GB) under the protocol of Sec. C.6 .

Method	Original input	Input here
HARP	Monocular video	Single frame
LiveHand	Multi-view video	Single frame
HandAvatar	Monocular video	Single frame
Handy	Single image	Single frame
OHTA	Single image	Single frame
BIGS ∗	Monocular video	Single frame

Appendix

Table S4: Input protocol for every compared method.

Method	Input	Feed-forward	Repr.
LiveHand [ 21 ]	Multi-view video	✗	NeRF-like
GaussianHand § [ 24 ]	Multi-view video	✗	3DGS
HandAvatar [ 20 ]	Monocular video	✗	Implicit
HARP [ 22 ]	Monocular video	✗	Mesh
BIGS ∗ [ 47 ]	Monocular video	✗	3DGS
OHTA [ 23 ]	Single image	✗	Implicit

Appendix

Table S5: Comparison of representative 3D hand avatar reconstruction methods under their original input settings. BIGS ∗ : our single-hand adaptation of BIGS (Sec. E.1 ). § GaussianHand reports results from as few as 5–20 camera views.

Figure S2: Qualitative effect of the number of Gaussians. Columns: ground truth (GT) and the 3K, 12K, and 49K settings of Table S6 . Red circles mark regions magnified in the corner insets.

$k$	#Gaussians	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
1	3,093	26.11	0.859	0.168
2	12,337	25.92	0.860	0.147
3	49,281	26.21	0.866	0.133

Appendix

Table S6: Ablation on the number of Gaussians (MANO subdivision level). All variants are full models differing only in the subdivision level.

Figure S3: Artifacts under per-axis Gaussian scales. Ground truth, the isotropic model (ours), and the per-axis-scale-freed variant of Table S7 on identical frames. Magenta circles mark interior regions and red circles silhouette regions, each magnified in the matching corner inset.

Configuration	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Ours (isotropic; rotation, opacity fixed)	26.21	0.866	0.133
Per-axis scale freed	25.58	0.856	0.142
Rotation and opacity freed	25.74	0.862	0.133
All three freed	25.50	0.856	0.148

Appendix

Table S7: Effect of relaxing the Gaussian attribute constraints (InterHand2.6M). Boundary supervision is retained in every row.

(a) Controlled perturbations around ground truth
Source perturbation	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Bound. F $\uparrow$	B-IoU $\uparrow$	Leakage $\downarrow$
Clean GT inputs	26.21	0.866	0.133	0.698	0.562	0.234
Pose noise ( $\sigma_{\mathrm{pose}}=5^{\circ}$ )	26.12	0.864	0.134	0.698	0.562	0.234
Pose noise ( $\sigma_{\mathrm{pose}}=10^{\circ}$ )	26.05	0.863	0.134	0.698	0.562	0.234
Mask erosion (2 px)	26.28	0.867	0.132	0.704	0.565	0.229
Mask dilation (2 px)	26.24	0.864	0.134	0.697	0.561	0.237

Appendix

Table S8: Sensitivity to source-side MANO and mask errors on InterHand2.6M. (a) Controlled perturbations around the ground-truth source cues. (b) Ground-truth source cues replaced by SAM 3 and WiLoR outputs. The last row also replaces the source camera.

$\kappa$	$\tau_{m}$	$\zeta$	PSNR $\uparrow$	LPIPS $\downarrow$
(a) InterHand2.6M (novel-pose/novel-view)
0.25	0.5	95	26.21	0.133
0.5	0.5	95	26.21	0.133
1.0	0.5	95	26.20	0.133
0.5	0.3	95	26.22	0.133
0.5	0.7	95	26.19	0.133

Appendix

Table S9: Inference-time sensitivity to observation-reliability hyperparameters. One parameter is varied at a time with all weights fixed, and bold marks the default.

Source frames	Coverage	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
1	0.45	26.21	0.866	0.133
2	0.68	26.53	0.867	0.131
4	0.71	26.52	0.867	0.132
8	0.96	26.57	0.868	0.130

Appendix

Table S10: Multi-frame evidence integration at inference time (InterHand2.6M). The same trained model is used without retraining or additional parameters, and frames beyond the first are selected to maximize the union coverage of observed Gaussians (

\gamma_{i}>0

Figure S4: Qualitative effect of the Soft Alpha Band (SAB). Columns: ground truth (GT), without SAB (w/o SAB), and with SAB (Ours). The black box marks the input frame, and the remaining rows are novel poses. Red circles mark boundary regions magnified in the corner insets.

Figure S5: Qualitative comparison with the recipe ports of Table 4 of the main paper. InterHand2.6M novel-pose/novel-view frames, with the source camera in the top row. Red circles mark inter-finger gaps and contours magnified in the corner insets.

Figure S6: Qualitative source-view appearance diagnostic with geometry fixed (protocol in Sec. C.5 ). Columns: source observation (Source), Stage-1 base colors (w/o transfer), and composed colors (Ours). Red circles mark regions magnified in the insets.

Figure S7: Branch-level analysis of source-conditioned local detail transfer. Columns: source observation (Source), Stage-1 base colors (w/o transfer), the local-only variant (Local-only), and the full model (Ours) under the same source pose and view. The red circle marks the region magnified below.

Figure S8: Qualitative generalization to unseen FreiHAND [ 61 ] test images under fully estimated inputs. Rows: input frame (Input), WiLoR [ 37 ] pose overlay (MANO), SAM 3 [ 54 ] source crop (Source), and our render (Render). No ground-truth annotations or test-time optimization are used.

Figure S9: Novel-pose animation from a single image. In each block, the top row shows the driving MANO poses and each subject row the source image followed by our renderings in those poses. All inputs are fully estimated. Best viewed zoomed in.

Figure S10: Representative appearance failure cases. Columns: Input, MANO, Source, and Render (the row labels of Fig. S8 ). Red circles mark the affected regions in the third and bottom rows.

	InterHand2.6M					HanCo (zero-shot)
Setting	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Bnd. F $\uparrow$	Leak. $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Bnd. F $\uparrow$	Leak. $\downarrow$
$k=3$ (Tables 1 , 2 )	26.21	0.866	0.133	0.698	0.234	20.18	0.836	0.188	0.617	0.314
$k=4$ input	25.98	0.862	0.137	0.671	0.247	19.75	0.829	0.190	0.559	0.373
$k=4$ , $s_{i}$	25.86	0.862	0.136	0.661	0.293	19.68	0.831	0.189	0.542	0.381
$k=4$ , $0.85\,s_{i}$	26.06	0.864	0.132	0.697	0.244	20.04	0.837	0.184	0.613	0.321
$k=4$ , $0.7\,s_{i}$	26.15	0.865	0.130	0.711	0.205	20.30	0.840	0.181	0.668	0.269

Appendix

Table S11: Inference-time densification (Sec. I ). The same trained

k=3

network is used throughout. “

k=4

input” feeds the

k=4

anchors to the network. The other

k=4

rows interpolate the

k=3

prediction onto the

k=4

anchors and multiply every scale

s_{i}

by the stated factor, and “

k=4

s_{i}

” keeps the scales unchanged. Boundary metrics follow Sec. C.4 . Bold marks the setting used for Fig. S11 and the supplementary video.

Figure S11: Inference-time densification. Columns: ground truth, the

k=3

avatar used throughout the paper (49K Gaussians), and the same prediction densified onto the

k=4

anchors (197K, scales

0.6\,s_{i}

, Table S11 ). Rows: three HanCo evaluation frames. Magenta circles mark knuckles and red circles the wrist, each magnified in the matching corner inset.

	InterHand2.6M				HanCo (zero-shot)
Tercile	Range	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Range	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Palm-orientation change
Easy	0–79	26.37	0.863	0.133	0–82	20.67	0.842	0.169
Medium	80–115	25.67	0.858	0.140	82–128	19.70	0.826	0.212
Extreme	116–179	26.61	0.876	0.126	129–180	20.15	0.840	0.183
Finger-articulation change

Appendix

Table S12: Pose-stratified evaluation of our method. Terciles follow Sec. C.8 , and ranges are in degrees.

	PSNR $\uparrow$				LPIPS $\downarrow$
Tercile	Ours	OHTA	HandAvatar	BIGS ∗	Ours	OHTA	HandAvatar	BIGS ∗
InterHand2.6M
Easy	27.37	27.23	24.90	21.30	0.106	0.103	0.150	0.168
Medium	26.56	26.43	24.21	20.38	0.119	0.115	0.163	0.191
Extreme	24.68	24.66	22.24	19.67	0.175	0.169	0.223	0.277
HanCo (ours zero-shot, baselines fitted to the source frame)

Appendix

Table S13: Articulation-stratified comparison on the finger-articulation terciles of Table S12 .

Camera	GT change	Ours	OHTA	HandAvatar	BIGS ∗
cam3	0.0362	$+1.1\%$	$-5.1\%$	$-17.6\%$	$+37.8\%$
cam5	0.0319	$+7.3\%$	$+5.0\%$	$-8.7\%$	$+50.5\%$
cam6 (source view)	0.0360	$+2.6\%$	$-2.2\%$	$-23.3\%$	$+30.7\%$
cam7	0.0412	$+8.0\%$	$+2.2\%$	$-15.6\%$	$+34.9\%$
Mean of $\lvert\cdot\rvert$	0.0363	$4.7\%$	$3.6\%$	$16.3\%$	$38.5\%$

Appendix

Table S14: Excess temporal variation on HanCo (Eq. S6 , 75 consecutive-frame pairs per camera). Ours is zero-shot, and the baselines are fitted to the source frame. GT change: mean change between consecutive ground-truth frames. Method columns give

E_{t}

relative to it, where 0 means as much change as the footage and positive values mean additional variation. cam6 is the source view, and the last row is the mean of the absolute per-camera values.

Figure S12: Canonical color and observation reliability. Final colors

\mathbf{c}^{*}

(top) and reliability

\gamma

(bottom) of each canonical avatar, rendered from the source view and from an orbit of the source camera about the hand, with the same viewpoints and crops in both rows.

Figure S13: Rendered opacity (accumulated alpha $A$ ) of the canonical avatars from the views of Fig. S12 . The straight cut in the $180^{\circ}$ view is the wrist opening of the MANO template.

3D Hand Reconstruction

See all 41 papers

OASIS: Occlusion-aware Single-image Hand Avatar Reconstruction via 3D Gaussian Splatting
Jul 31, 2026Zhisheng Han, Shiyao Wu, Jiayan Qiu +63D Gaussian Splatting 3D Reconstruction
University of Leicester Leicester, UK · University of Exeter Exeter, UK · University of Birmingham Birminghan, UK +1
EgoHand: Monocular 3D Hand Reconstruction from Egocentric Video
Oct 21, 2026Chao-Yeh Chen, Tiffany Chen, Lin Duan +123D Hand Reconstruction 3D Hand Pose Estimation
NVIDIA
WildGHand: Learning Anti-Perturbation Gaussian Hand Avatars from Monocular In-the-Wild Videos
Feb 24, 2026Hanhui Li, Xuan Huang, Wanquan Liu +53D Hand Reconstruction 3D Reconstruction
School of Intelligent Systems Engineering, Shenzhen Campus of Sun Yat-sen University, P.R. China, 518107. · Lenovo Research Group, Shenzhen, P.R. China, 518038.

Jul 31, 2026J/K move · Enter open · S save

OASIS: Occlusion-aware Single-image Hand Avatar Reconstruction via 3D Gaussian Splatting

Zhisheng Han, Shiyao Wu, Jiayan Qiu, Yakun Ju+5

University of Leicester Leicester, UK · University of Exeter Exeter, UK · University of Birmingham Birminghan, UK+1

3D Gaussian Splatting 3D Reconstruction 3D Hand Reconstruction 3D Avatar Generation Deformable Shape Reconstruction Single-View 3D Reconstruction

Open paper AskPDF arXiv

Single-image 3D hand avatar reconstruction is fundamentally ill-posed and particularly challenging due to limited visual evidence under severe self-occlusion and the complex pose-dependent deformation of highly articulated hands. Existing methods predominantly rely on implicit NeRF-style representations, whose volumetric fitting is computationally expensive and often struggles to preserve fine-grained hand details. In this work, we present OASIS, a tailored 3D Gaussian Splatting framework for single-image hand avatar reconstruction. To faithfully encode sparse image-specific appearance cues in single-view reconstruction, we construct geometry-aligned visual evidence tokens by explicitly aligning input image observations with 3D hand geometry and context-adaptively tokenizing the resulting visual evidence. Since severe self-occlusion makes the reliability of image evidence inherently visibility-dependent, we introduce a visibility-conditioned point-image attention to reliably transfer visual evidence to geometric tokens, yielding occlusion-aware Gaussian features for faithful and robust reconstruction. To further capture non-rigid deformation of articulated hands, we introduce a Feature-on-Mesh representation to enable Gaussian deformation to be guided by local surface stretching. Under this framework, we adopt a one-shot adaptation scheme that learns a shared hand prior from multi-identity training data and then fits it to a target image for target-specific reconstruction. Extensive experiments show that OASIS outperforms existing baselines in both visual fidelity and efficiency across challenging poses and in-the-wild scenarios, and further demonstrates strong versatility in downstream applications such as text-to-avatar generation and texture editing.

Abstract

3D Hand Reconstruction

Explore similar work

OASIS: Occlusion-aware Single-image Hand Avatar Reconstruction via 3D Gaussian Splatting

EgoHand: Monocular 3D Hand Reconstruction from Egocentric Video

WildGHand: Learning Anti-Perturbation Gaussian Hand Avatars from Monocular In-the-Wild Videos