Recent advances in Gaussian splatting have enabled efficient and photorealistic head avatar reconstruction, yet monocular video-based or image-based methods still suffer from under-constrained geometry and texture, leading to artifacts such as blurring, stretching, and temporal instability. We present EDGAR (Efficient Diffusion-Driven Gaussian Avatar Refinement), a conditional diffusion framework designed to enhance the quality of Gaussian avatars by correcting artifacts while preserving identity fidelity. To train EDGAR, we construct EDGAR-SET, a hybrid triplet dataset composed of real-captured and synthetic-rendered samples, which simulates realistic Gaussian avatar failure cases by pairing degraded renderings with ground-truth and reference images. EDGAR refines degraded frames through a fusion block that injects identity features extracted by the self-supervised Sapiens model into a one-step diffusion backbone, enabling effective artifact removal without sacrificing efficiency. Our EDGAR model incorporates a Mixture-of-Experts (MoE) architecture to address diverse degradation types and severity levels. The refined multi-view renderings are subsequently used as pseudo-targets to optimize a single underlying Gaussian avatar, so that the final avatar can be rendered without applying EDGAR to every output frame. Experiments on three different Gaussian avatar reconstruction methods, each evaluated under three different settings, show improvements in both perceptual and reconstruction measures over the raw outputs and an SDEdit-style refinement baseline. In addition, cross-view evaluation demonstrates improved calibrated consistency, while tests under large yaw and open-mouth expressions verify robustness under extreme settings. These results collectively indicate a useful operating regime for recoverable appearance artifacts. The code is available at https://kwanyun.github.io/edgar/
Figures & tables
Figure 1. From diverse failure cases of Gaussian avatars, EDGAR learns to refine arbitrary avatars trained with a video or an image, acting as both an artifact remover and a quality enhancer.
Figure 2. Examples from EDGAR-SET and discarded samples. We filter out cases where the geometry or expression is misaligned, making the recovery target ambiguous. Samples are removed when the combination of If and Ir from novel views or expressions cannot reliably reproduce Igt .
Figure 3. Configuration of EDGAR-SET. The dataset is constructed under diverse conditions to capture a wide range of Gaussian avatar failure cases. The view-count groups (one, two to four, and at least five) are mutually exclusive and account for 23%, 51%, and 26% of retained examples, respectively.
Figure 4. Network architecture of EDGAR. EDGAR learns to refine rendered images of Gaussian avatars using EDGAR-SET. The fuse block in the Upblock of Diffusion UNet is simplified.
Figure 5. Avatar refinement process using EDGAR. Multi-view renderings of the source Gaussian avatar are refined by EDGAR and used as supervision targets for Gaussian avatar refinement.
Figure 6. Comparison of self-reenactment results. Our method refines each baseline by restoring facial details such as the eyes, wrinkles, and sharper hair, whereas SDEdit-based refinement alters the source identity. The upper rows show outputs from avatars trained or optimized with fewer input images or views (one image for CAP4D, one video for RGBAvatar, and eight videos for GaussianAvatars), while the lower rows show outputs from avatars trained with more images or videos. Each block uses a different identity and input severity, so comparisons are made within each reconstruction method. The more severely degraded GaussianAvatars example is improved but remains only partially corrected.
Figure 7. Comparison of cross-reenactment results rendered at novel views. Our method refines each face by restoring facial details while effectively removing Gaussian artifacts such as floating blobs. In contrast, SDEdit-based refinement introduces identity shifts and oversmoothing skin texture.
Method
1 view
4 views
8 views
CLIP ↑
L1 ↓
LPIPS ↓
PSNR ↑
CLIP ↑
L1 ↓
LPIPS ↓
PSNR ↑
CLIP ↑
L1 ↓
LPIPS ↓
PSNR ↑
CAP4D
0.9369
0.0584
0.2133
17.5651
0.9393
0.0537
0.1931
18.1435
0.9462
0.0518
0.1847
18.4773
+SDEdit
0.9078
0.0581
0.2728
17.7737
0.9101
0.0535
0.2562
18.3674
0.9151
0.0518
0.2529
18.6645
+Ours
0.9449
0.0580
0.2147
17.5978
0.9488
0.0534
0.1962
18.1927
0.9531
0.0513
0.1884
18.5329
Table 1. Quantitative comparison on CAP4D across different numbers of input views. Best results per column are in red and second best are in yellow.
Method
1 video
4 videos
8 videos
CLIP ↑
L1 ↓
LPIPS ↓
PSNR ↑
CLIP ↑
L1 ↓
LPIPS ↓
PSNR ↑
CLIP ↑
L1 ↓
LPIPS ↓
PSNR ↑
RGBAvatar
0.8423
0.0832
0.3619
16.0469
0.8877
0.0552
0.2524
19.4950
0.9210
0.0502
0.2100
20.5128
+SDEdit
0.8554
0.0827
0.3601
16.5024
0.8639
0.0570
0.2878
19.9181
0.8712
0.0555
0.2752
20.5109
+Ours
0.8897
0.0677
0.3096
17.1533
0.9273
0.0364
0.2121
21.5199
0.9379
0.0332
0.1894
22.9937
Table 2. Quantitative comparison on RGBAvatar across different numbers of input videos . Best results per column are in red and second best are in yellow.
Method
8 videos
10 videos
12 videos
CLIP ↑
L1 ↓
LPIPS ↓
PSNR ↑
CLIP ↑
L1 ↓
LPIPS ↓
PSNR ↑
CLIP ↑
L1 ↓
LPIPS ↓
PSNR ↑
GaussianAvatars
0.7819
0.0852
0.3299
16.2855
0.8407
0.0644
0.2661
18.5994
0.8759
0.0738
0.2303
18.1037
+SDEdit
0.8173
0.0806
0.2894
16.9214
0.8767
0.0649
0.2377
18.9219
0.9025
0.0782
0.2202
18.0437
+Ours
0.8300
0.0739
0.2526
17.3766
0.8977
0.0593
0.2050
19.3206
0.9214
0.0724
0.1875
18.2789
Table 3. Quantitative comparison on GaussianAvatars across different numbers of input videos . Best results per column are in red and second best are in yellow.
Method
1 view
4 views
8 views
Base
0.5510
0.5642
0.5651
+SDEdit
0.3178
0.3432
0.3380
+Ours
0.5169
0.5326
0.5419
Table 4. Quantitative comparison of CSIM across different numbers of input views. Best results per column are in red and second best are in yellow.
Method
1 view
4 views
8 views
CLIP ↑
L1 ↓
LPIPS ↓
PSNR ↑
CLIP ↑
L1 ↓
LPIPS ↓
PSNR ↑
CLIP ↑
L1 ↓
LPIPS ↓
PSNR ↑
w/o Ir
0.9163
0.0595
0.2683
17.5822
0.9222
0.0547
0.2492
18.1634
0.9262
0.0527
0.2448
18.4787
w/o Synthetic
0.9351
0.0583
0.2292
17.6128
0.9407
0.0538
0.2106
18.1925
0.9436
0.0518
0.2033
18.5236
w/o Lexp
0.9373
0.0578
0.2297
17.6271
0.9433
0.0533
0.2107
18.2135
0.9465
0.0512
0.2038
18.5370
w/o MoE
0.9420
0.0580
0.2148
17.5757
0.9456
0.0534
0.1962
18.1689
0.9503
0.0514
0.1884
18.5053
w/ two Experts
0.9433
0.0581
0.2154
17.5824
0.9470
0.0535
0.1967
18.1802
0.9521
0.0514
0.1888
18.5201
Table 5. Ablation study results across different numbers of input views on self-reenactment. Best results per column are in red and second best are in yellow.
Method
1 view
4 views
8 views
w/o Ir
0.4508
0.4716
0.4769
w/o Synthetic
0.4971
0.5244
0.5242
w/o Lexp
0.4645
0.4930
0.4927
w/o MoE
0.5177
0.5287
0.5355
w/ two Experts
0.5162
0.5301
0.5363
Ours
0.5169
0.5326
0.5419
Table 6. Ablation study results across different numbers of input views on cross-reenactment measured by CSIM. Best results per column are in red and second best are in yellow.
Pipeline
Base
+EDGAR
CAP4D
1.759±0.210
1.696±0.199
RGBAvatar
7.424±1.411
5.574±1.512
GaussianAvatars
4.929±1.111
5.938±1.932
Average
4.704±0.803
4.403±1.105
Table 7. Foreground tWE ( ×10−3 ; lower is better). Values are mean ± identity-level standard deviation. The lower value in each row is red.
Figure 8. Matched-subject comparison at the same evaluation frame. Native inputs are four images for CAP4D, four videos for RGBAvatar, and ten videos for GaussianAvatars. Identity and evaluation frame are shared, while input regimes and initial reconstruction quality differ across pipelines.
Pipeline
One-time cost (min)
Output FPS
Base opt.
Avatar refine.
Base
EDGAR
CAP4D
484.0±80.5
5.04±0.11
50.3±0.3
49.4±0.6
RGBAvatar
105.6±4.4
7.96±0.98
813.8±36.4
826.3±16.9
GaussianAvatars
238.7±8.8
16.60±5.98
153.4±7.4
156.1±0.7
Table 8. Computational cost and output throughput on an RTX 3090. Refinement time includes loading, pseudo-target generation, avatar optimization, and checkpoint writing. Values are mean ± sample standard deviation.
Avatar family
Inputs
Metric
Base
EDGAR
CAP4D
1
Ed↓
0.0028±0.0004
0.0025±0.0003
C↑
0.9135±0.0181
0.9217±0.0159
4
Ed↓
0.0028±0.0004
0.0025±0.0004
C↑
0.9145±0.0166
0.9222±0.0153
8
Ed↓
0.0028±0.0004
0.0025±0.0004
C↑
0.9141±0.0169
0.9218±0.0157
Table 9. Cross-view geometric consistency results. Values are reported as mean ± standard deviation. Better values are highlighted in red.
Foreground L1↓
Foreground LPIPS ↓
Pipeline
Condition
Base
EDGAR
Base
EDGAR
CAP4D
Large pose
0.07500
0.07601
0.22353
0.21704
Open mouth
0.06906
0.06660
0.21057
0.21975
Both
0.07475
0.08084
0.23054
0.23507
RGBAvatar
Large pose
0.08894
0.06121
0.17845
0.14399
Open mouth
0.06410
0.05372
0.12051
0.10368
Table 10. Pose-and-expression evaluation at representative native input settings (CAP4D: four images; RGBAvatar: four videos; GaussianAvatars: ten videos). Values are equally weighted identity means. Large pose denotes ∣yaw∣≥30∘ , and open mouth denotes a normalized lip gap ≥0.10 . Better values are highlighted in red.
Figure 9. Large-yaw with optional open-mouth examples for RGBAvatar with four input videos. Rows show examples with absolute yaw angles of 37.4∘ , 37.7∘ , and 36.3∘ , respectively. Examples are selected near each identity’s median ground-truth yaw and mouth opening within the combined subset.
Figure 10. Ablation study results for Ir and Synthetic data. Some of the artifacts in the hair, ears, and floating blobs exhibited by the input Gaussian avatar are mitigated by the ablated variants. Nonetheless, the variant without Ir produces altered details and overly smooth textures due to the absence of reference information, while the variant without synthetic data retains artifacts such as distortions around the ears and floating Gaussian blobs. In contrast, the full model better preserves identity while effectively removing artifacts and improving overall visual quality. Visible artifacts in the ear and floating Gaussian blobs are highlighted in red boxes.
Figure 11. Ablation study results on Lexp . Removing Lexp degrades the preservation of facial expression details.
Figure 12. Qualitative results for the MoE ablation. All variants reduce artifacts, but the variants without MoE and with two experts retain sharp white artifacts around the ears. The full model suppresses these artifacts more effectively in this example.
Method
Animation
Preservation
Preference
Ours vs CAP4D
87.33%
91.33%
96.67%
Ours vs CAP4D+SDEdit
94.67%
98.0%
98.0%
Table 11. Selection ratio of our method over the baselines in the user study.
We propose S-Avatar, a novel method for generating photorealistic 3D head avatars from a single image using a diffusion-guided 3D model generation module and strategies for animating 3D Gaussian Splatting (3DGS). While single-image head avatar reconstruction is crucial for lifelike Virtual Reality (VR) applications, existing approaches often struggle to preserve 3D consistency under unseen viewpoints. S-Avatar addresses this limitation through a three-stage pipeline. First, a high-resolution 3DGS is synthesized directly from a single image using a diffusion-based Gaussian splat generation module. Next, the parametric head model FLAME is aligned with the generated 3DGS by optimizing its parameters and spatial transformations. Finally, to adapt the 3DGS to FLAME variations, we construct a binding template that encodes the spatial relationship between the initial splats and FLAME. The dynamic 3D head avatar can then be rendered in real time by deforming the 3DGS with the binding template. By combining diffusion-guided canonical 3DGS generation with FLAME-based control, our method achieves efficient and accurate reconstruction with enhanced 3D consistency. Evaluations on public datasets demonstrate that S-Avatar outperforms state-of-the-art methods in novel-view and expression generation, achieving superior realism and consistency. Consequently, our approach represents a significant advance in accessible avatar creation, applicable to a wide range of VR/AR applications. The project page is available at https://github.com/hailsong/savatar.
Creating photorealistic animatable head avatars from a single image remains a fundamental challenge in digital human synthesis. While recent 3D Gaussian Splatting methods have achieved promising results, they rely on external tracking pipelines whose latency is excluded from inference measurements. Furthermore, they adopt unified representations that entangle geometrically distinct facial regions, limiting both expressiveness and rendering fidelity. We propose SpiD (Split and Drive), a single-image Gaussian head avatar framework built on two disentanglement axes. The compute axis internalizes per-frame driving, eliminating external tracking dependency at inference. The feature axis decomposes the avatar into three specialized Gaussian branches, each modeling a geometrically distinct facial domain. Extensive experiments demonstrate consistently strong performance against state-of-the-art methods while achieving the fastest inference speed among all compared methods on a single GPU with the complete driving pipeline included.
MD Wahiduzzaman Khan, Mingshan Jia, Xiaolin Zhang +2
University of Technology Sydney, Australia · Shandong University of Science and Technology, China
High-quality 4D head avatars from one or a few source portraits are central to telepresence, AR/VR, and digital-human interaction. 3D Gaussian Splatting (3DGS) has emerged as the dominant representation, with two complementary regimes (generalizable feed-forward predictors and per-subject refiners) maturing in parallel. However, existing feed-forward predictors are trained on a single dataset family with a hard-coded source count, inheriting the corresponding domain bias. Per-subject refiners require 300K--600K iterations and rely on adaptive densification that destroys upstream Gaussian layouts, preventing the two regimes from sharing a representation end-to-end. To bridge both regimes we propose SpatialAvatar-0 on a shared FLAME-mesh-bound Gaussian representation: a feed-forward generator with a parameter-free K-source mean-pool and a monocular-temporal to multi-view-spatial two-phase schedule that anchors against identity-prior collapse onto the smaller multi-view set. We further introduce a 10K-iter layout-preserving per-subject refinement loop that freezes the FLAME-binding and Gaussian count and replaces densification with a three-component anti-spike regularization. On VFHQ/HDTF cross-domain zero-shot we surpass the in-domain leader GAGAvatar by +1.5 dB PSNR despite never training on either test domain, and on the SplattingAvatar monocular benchmark we lead every reported metric, surpassing the 300K-iter GeoAvatar by +1.3 dB PSNR at up to 60x shorter per-subject schedule than common SOTA baselines. Website: https://spatialwalk.github.io/SpatialAvatar-0.