Pith. sign in

REVIEW 6 major objections 7 minor 9 references

StreamME: Simplify 3D Gaussian Avatar within Live Stream

T0 review · 6 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read StreamME claims that a 3D head avatar can be reconstructed and visualized on the fly from a live monocular webcam stream in about five minutes, with no offline phase and no MLP deformation network, by binding simplified 3D Gaussian point…

desk verdict The on-the-fly training idea is real and worth peer review, but the headline 5-minute live-stream claim is only backed by qualitative evidence while all quantitative tables are offline. read the letter →

arxiv 2507.17029 v1 pith:AMVE6JXQ submitted 2025-07-22 cs.GR cs.AIcs.CV

Luchuan Song , Yang Zhou , Zhan Xu , Yi Zhou , Deepali Aneja , Chenliang Xu This is my paper · ORCID
classification cs.GRcs.AIcs.CV MSC 68U05
keywords on-the-flytraining3DGaussianSplattingheadavatarreconstructionpointcloudsimplificationmotion-awareanchorslivestreamingmonocularfacialanimation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

StreamME tries to prove that a usable, animatable 3D head avatar no longer needs a separate offline training phase: it can be built on the fly while a person is live streaming from one webcam, reaching recognizable detail in about 10 seconds and photo-realistic quality in roughly 5 minutes. The payoff is that no pre-cached personal model has to be stored or transmitted, only the 3D Gaussian parameters travel, which the paper estimates as about 70% compression versus sending video, and under-trained facial regions can be re-recorded immediately. The method drops the MLP deformation networks used by prior deformable 3DGS avatars and lets a real-time face tracker's mesh carry all facial motion, with Gaussians attached to that mesh and a learnable binary index deciding which points become motion anchors to keep and duplicate. If the claim holds, live-streaming telepresence, VR avatars, and online conferencing gain a privacy-preserving, low-bandwidth path from a single camera to a re-animatable face.

What carries the argument

The central object is the learnable binary index $\text{Idx}\in\{0,1\}^N$ attached to each 3D Gaussian, computed as the thresholded maximum of the inner product between the accumulated density gradient and the motion gradient, where the motion gradient is the position difference of each point relative to the normalized canonical point cloud of the first frame. This index does three jobs: it marks motion-relevant points for cloning, prunes motion-irrelevant points such as those on the forehead, and acts as a mask on opacity and scale whose binary gradients are propagated by a straight-through estimator. Together with a point-to-surface distance regularization that pulls bound points back onto the facial surface, the index is what lets the method strip out MLPs and still adapt quickly to expressions. An optional warm-up phase learns auxiliary texture and lighting parameters ($\text{Tex}$, $\text{SH}$) for roughly 20 seconds to displace coarse template vertices around regions such as hair, and these attributes are frozen before on-the-fly training begins.

What would settle it

Feed the identical webcam video into StreamME twice, once with the real-time tracker's per-frame meshes and once with offline, carefully fitted meshes of the same frames, and compare 5-minute on-the-fly PSNR on held-out expressions; if the real-time-tracker version falls well short of the offline-mesh version, the speed claim comes from the tracker, not the simplification strategy. A second check is to drive the finished avatar with an out-of-distribution expression absent from the stream, where the paper's own limitations predict visible artifacts, which would confirm that 'photo-realistic within 5 minutes' is scoped to the stream's recorded distribution.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the slow part of head-avatar reconstruction is not the rendering representation but the learned deformation field: prior methods spend minutes of optimization fitting MLP networks that map expression coefficients to point offsets. By attaching 3D Gaussian points directly to the deformed head mesh supplied by a real-time tracker, StreamME makes deformation free and reduces the learning problem to appearance, namely color, opacity, scale, rotation, plus a motion-aware selection of which Gaussians matter. The selection is carried by a learnable binary index $\text{Idx}\in\{0,1\}^N$ computed from the density gradient and the motion gradient of each point relative to the canonical first-frame point cloud: points passing a threshold are duplicated as motion anchors, the rest are pruned, and the same index masks opacity and scale during training so that over-small and redundant points disappear. The reported result is an on-the-fly pipeline reaching about 29.7 dB PSNR on self-reenactment with a 2.52 MB model, 139 FPS inference, and progressively faster training as the point cloud sparsifies.

Load-bearing premise

The whole 5-minute promise rests on the assumption that the real-time single-image face tracker already provides a metrically reliable, correctly deformed mesh at every frame, so facial motion costs nothing; if the tracker drifts, jitters, or fails under extreme poses or occlusion, the attached Gaussians inherit that error and the reported quality rests on a component the paper does not train.

Editorial extensions

If this is right

  • A live-streaming avatar call can transmit 3D Gaussian parameters instead of images, roughly 70% bandwidth compression in the paper's estimate, so the receiver can rebuild the face without ever storing the sender's raw video.
  • Training gets faster as it runs: because simplification prunes points progressively, per-iteration cost falls over time, from about 1.6e5 points down to roughly 9.8e3 at convergence, so a longer stream keeps improving quality without slowing down.
  • Comparable or better quality is reached much earlier than baselines: the paper reports about 27.2 dB PSNR at 2 minutes, a level FlashAvatar and GaussianBlendshape reach only near 30 minutes.
  • The 5-minute model transfers directly to downstream tasks, including cross-identity animation, text-prompt toonification, and environment relighting with minimal fine-tuning, so one short stream can produce a reusable avatar asset.
  • Because no personal face model is cached on external machines and under-trained areas can be re-recorded immediately, the on-the-fly setup changes where and how face data is stored in VR and conferencing systems.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 'no pre-cached data' framing quietly depends on a pre-trained face tracker; a fair extension would measure output quality against tracker accuracy as a variable, since the anchor-gradient formula is the channel through which tracking errors enter the model.
  • The simplification machinery, binary motion-aware masks with straight-through gradients plus point-to-surface pull, is not face-specific; any reliably tracked deformable object such as hands, clothed bodies, or animals could inherit the same on-the-fly recipe if a tracker with comparable per-frame metric reliability exists.
  • The 70% compression figure suggests a protocol question the paper leaves open: whether incremental model updates can stay below video bitrate under real network jitter, since the model must be transmitted and re-rendered at the receiver.
  • Because the paper's own limitations tie quality to the training stream's pose and expression distribution, the practical claim is '5 minutes for a conversational avatar,' not '5 minutes for arbitrary reenactment'; an honest consumer test would stream a neutral monologue and then drive the avatar with exaggerated expressions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

6 major / 7 minor

Summary. The manuscript proposes StreamME, a 3D Gaussian Splatting-based head avatar reconstruction method intended to operate in an on-the-fly training regime, reconstructing a head avatar from a live monocular webcam stream in about five minutes without pre-cached personal data. The method attaches Gaussian points to a tracked mesh, selects motion-aware anchor points via a learnable binary index, and prunes and simplifies points to reduce computational cost. The authors report offline quantitative comparisons against AvatarMAV, FlashAvatar, and GaussianBlendshape, together with qualitative demonstrations of live reconstruction, animation, toonification, and relighting.

Significance. The efficiency target is genuinely useful: if validated, StreamME would be one of the first methods to make avatar reconstruction available during live streaming, with a compact 2.52 MB model and high inference FPS. The design choice of replacing MLP-based deformation with mesh-bound Gaussians is plausible, and the ablations give some qualitative evidence that the warm-up, anchor selection, and simplification modules each contribute. However, the paper's strongest claims are not yet supported by its experiments: the live five-minute regime is not measured quantitatively, the real-time tracker is unspecified, and the central Idx mechanism is defined inconsistently. No code is made available in the manuscript, so reproducibility currently rests on the textual description.

major comments (6)
  1. [§4.1.1, Table 1, Figure 1] The central claim of on-the-fly reconstruction is validated only qualitatively. Section 4.1.1 states that the six-subject quantitative experiments are conducted in the 'offline training setup' on public datasets, while the on-the-fly results are self-captured webcam videos shown in Figure 1. Table 1's PSNR=29.7 dB and Table 2's time-course numbers therefore do not directly establish the live-stream claim. In addition, Figure 1's caption reports PSNR=26 dB within 5 minutes, which is inconsistent with the offline 29.7 dB. The authors should report live PSNR/LPIPS values or explicitly state that the offline numbers are used as a proxy, and they should reconcile the two reported PSNR values.
  2. [§2.2, §4.1.4] The real-time face estimation module is never named or evaluated. The on-the-fly pipeline attaches Gaussian points to deformed meshes from this module (Figure 2a), so its accuracy and frame rate directly bound the quality of reconstruction. Without reporting the tracker's identity and its accuracy and latency, the comparison with baselines that use precomputed MICA tracks is not on equal footing, and the claim of operating without pre-cached data is left ambiguous about pre-trained models. Please specify the tracker and measure its tracking error in the live setup.
  3. [§3.3–3.4, Eq. (4)–(6)] The role of Idx is inconsistent. Equation (4) defines Idx by a deterministic comparison of accumulated optimizer and motion gradients to a threshold, whereas Eq. (6) treats Idx as a continuous learnable parameter binarized with a straight-through estimator. The text does not state whether Eq. (4) is used only for initialization, whether Eq. (5) uses the binarized or soft value, or how often Eq. (4) is recomputed during training. Because Idx drives both anchor duplication and point pruning, this ambiguity makes the core simplification mechanism under-specified and difficult to reproduce.
  4. [§4.2–4.3, Table 1] Quantitative comparisons lack variance information and statistical testing. Table 1 reports single PSNR/MSE/LPIPS values without standard deviations or number of runs, and the 30-participant user study is reported only as aggregate means without spread or a test of significance. Given the observed margins (e.g., StreamME 29.7 versus FlashAvatar 27.8 PSNR), the reader cannot tell whether the differences are meaningful. Also, the exclusion of INSTA and other related methods in §4.2 is asserted rather than demonstrated; the claim of superiority over the field needs either a broader comparison or a principled inclusion criterion.
  5. [§4.1.4, Figure 8] The warm-up duration is inconsistent and its dependency on pose diversity is unmeasured. Section 4.1.4 allocates 'approximately 30 seconds' for the warm-up, while Figure 8 states it 'takes about 10 seconds and several appearances from different perspectives.' A typical frontal webcam stream may not provide the required viewpoint variation, and this is a load-bearing assumption for the on-the-fly claim. Please measure warm-up duration and pose coverage in the live setup, or show robustness to restricted head motion.
  6. [Abstract, §1, §2.2] The statement 'without any pre-cached data' is overstated if interpreted literally, since the pipeline uses RobustVideoMatting and a pre-trained face tracker. If the intended meaning is 'no pre-cached personal identity data,' the manuscript should say so explicitly; otherwise the privacy and bandwidth claims based on this phrase are not established.
minor comments (7)
  1. [Throughout] There are numerous typos, including 'tpoic' in Section 1, 'reconstitution' in Section 2.1, 'warp-up' in Section 4.6.2, 'phrase' for 'phase' in Sections 3.2 and 4.6.1, and 'ccomparison' in the Figure 14 caption; a careful proofread is needed.
  2. [Table 2] The column header '30in' should read '30min', and the scientific-notation iteration counts (e.g., '1.4 e5') should be explained in the caption or text.
  3. [Figure 2 caption] The notation for Idx, o*, and Tex/SH is introduced in the figure but not fully explained in the main text; in particular, the distinction between frozen Tex/SH and optimized Tex/SH should be stated in Section 3.2.
  4. [Eq. (7)] The summation 'over A in Idx' is unclear because Idx is defined as a vector of binary values rather than a set of anchor indices; please define the indexing explicitly.
  5. [§4.2] The sentence about excluding INSTA because of 'on-the-fly reconstruction setups' is confusing, since INSTA is not presented as an on-the-fly method in the related work; the inclusion and exclusion criteria should be restated.
  6. [§2.2] The MICA citation appears as '[MIC 2022]' with no full reference entry; please provide the complete citation.
  7. [Reproducibility] The paper does not state whether source code will be released; for reproducibility, please add a code availability statement.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: StreamME's reported results are held-out empirical measurements, not predictions forced by fitted parameters or self-citation.

full rationale

The central reconstruction-quality and efficiency claims are empirical, not derivational. Section 4.1.1 states that quantitative experiments use a held-out test split ('the test dataset includes frames with novel expressions and poses (about 20%)'), and Section 4.3.1 defines the metrics (PSNR, LPIPS, MSE) on self-reenactment. Table 2 reports PSNR at fixed training times on this held-out protocol, so the 29.7 dB result is a measured generalization number rather than a value implied by the model's own definitions. The training losses are standard L1 and SSIM against ground-truth frames (Section 4.1.3), so visual fidelity is not defined in terms of any fitted parameter. The Idx motion-anchor rule (Eq. 4) and the simplification mask (Eq. 5) are algorithmic heuristics whose benefit is established by ablations (Figures 9-12), not by circular reasoning; the claim that simplification preserves quality is verified with error maps and held-out comparisons. Self-citations exist (e.g., TextToon [Song et al. 2024a] in Section 5.2 as the toonification recipe, and prior Song et al. works in related-work lists), but they are not used to justify the 5-minute reconstruction claim or to exclude alternative methods; they are implementation or background citations and are not load-bearing. The strongest concerns in the paper are evidentiary rather than circular: the live 'approximately 5 minutes' claim is supported only qualitatively (Figure 1), the real-time face tracker and RobustVideoMatting are pre-trained external dependencies that complicate the 'no pre-cached data' framing, and the warm-up duration is inconsistently reported as 'approximately 30 seconds' in Section 4.1.4 versus 'about 10 seconds' in Figure 8. These are scope, consistency, and support issues; none of them makes a prediction equivalent to its inputs by construction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The paper builds on 3DGS and FLAME tracking from prior literature. Its central efficiency claim depends on several hand-tuned thresholds and durations (epsilon, warm-up length, prune interval) and two geometric assumptions about point placement that are asserted rather than derived. No new physical entities are introduced; the Idx binary mask is a learned auxiliary variable.

free parameters (5)
  • Motion anchor threshold epsilon (Eq. 4) = 0.01
    Controls the clone/prune decision for Gaussian points; directly sets how sparse the point cloud becomes and therefore training speed.
  • Binarization threshold epsilon' (Eq. 6) = 0.01
    Sets the mask threshold in the straight-through estimator; affects which points are simplified away.
  • Warm-up duration = 30 seconds
    Chosen by hand; Section 4.1.4 warns that longer warm-up may cause collapse. Directly enters the 5-minute total reconstruction claim.
  • Pruning/cloning interval = every 1500 iterations
    Determines how often the point cloud is simplified; affects the training speed profile and the final point count.
  • Loss weights = L1=1, SSIM=0.1, reg=0.01, dark channel=10
    Tuned loss weights for the reconstruction and warm-up phases; needed to reproduce the reported quality numbers.
assumptions (5)
  • domain assumption 3D Gaussian Splatting forward model and differentiable rasterizer are correct and applicable to dynamic head rendering.
    The entire method relies on the 3DGS formulation from Kerbl et al. 2023, introduced in Section 3.1.
  • domain assumption Real-time face tracker (MICA/FLAME) provides a sufficiently accurate, metrically consistent mesh at frame rate.
    Section 2.2 and Figure 2 use a tracked deformed mesh as the geometric backbone; no error model for tracking is given.
  • ad hoc to paper Single-view monocular video over 5 minutes is sufficient to determine identity, expression, and appearance.
    The on-the-fly efficiency claim presumes the webcam stream provides enough pose and expression diversity; this is asserted, not demonstrated.
  • ad hoc to paper Small-size Gaussians and points inside the surface contribute negligibly to image quality.
    Section 3.4 states these as foundational assumptions for simplification; they drive the pruning that produces the speed gain.
  • standard math Straight-through estimator provides usable gradients for the binary Idx mask.
    Section 3.4, Eq. 6, cites Bengio et al. 2013; accepted practice for binary latent variables, though not proven for this objective.
invented entities (2)
  • Idx binary learnable mask (anchor index)
    purpose: Labels each 3D Gaussian as motion-relevant (1) or prunable (0); drives duplication and simplification.
    New latent variable introduced in Section 3.3; it is learned during training and has no observable handle outside the training loop.
  • Anchor points with bound point clusters
    purpose: Sparse set of 3D Gaussians that survive pruning and to which other points are bound; used for motion-aware duplication and surface-distance regularization (Eq. 7).
    Algorithmic construct from Sections 3.3 and 3.4; no external validation that the anchors correspond to semantic facial features.

how reviews work

0 comments
Cite this review

Pith. "Pith review of StreamME: Simplify 3D Gaussian Avatar within Live Stream." pith.science (2026). https://pith.science/paper/AMVE6JXQ

@misc{pith2026250717029,
  author       = {Pith},
  title        = {Pith review of: StreamME: Simplify 3D Gaussian Avatar within Live Stream},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AMVE6JXQ}},
  note         = {Machine review of arXiv:2507.17029}
}
read the original abstract

We propose StreamME, a method focuses on fast 3D avatar reconstruction. The StreamME synchronously records and reconstructs a head avatar from live video streams without any pre-cached data, enabling seamless integration of the reconstructed appearance into downstream applications. This exceptionally fast training strategy, which we refer to as on-the-fly training, is central to our approach. Our method is built upon 3D Gaussian Splatting (3DGS), eliminating the reliance on MLPs in deformable 3DGS and relying solely on geometry, which significantly improves the adaptation speed to facial expression. To further ensure high efficiency in on-the-fly training, we introduced a simplification strategy based on primary points, which distributes the point clouds more sparsely across the facial surface, optimizing points number while maintaining rendering quality. Leveraging the on-the-fly training capabilities, our method protects the facial privacy and reduces communication bandwidth in VR system or online conference. Additionally, it can be directly applied to downstream application such as animation, toonify, and relighting. Please refer to our project page for more details: https://songluchuan.github.io/StreamME/.

Figures

Figures reproduced from arXiv: 2507.17029 by the authors.

Figure 1
Figure 1. The StreamME takes live stream (or monocular) video as input to enable rapid 3D head avatar reconstruction. It achieves impressive speed, capturing the basic facial appearance within 10 seconds (PSNR = 21 dB) and reaching high-quality fidelity (PSNR = 26 dB) within 5 minutes, as shown in (a). Notably, StreamME reconstructs facial features through on-the-fly training, allowing simultaneous recording and modeling with… view at source ↗
Figure 2
Figure 2. The overview the pipeline of StreamME. We list three components at here. i) The 3DGS Properties Warming-Up (Optional): we introduce two auxiliary learnable 3D Gaussian attribute texture and illumination, refining the UV vertex positions to improve facial geometry detail (e.g. here, we show the coarse displacement for the vertices around the hair). This step is optional, and the users may also opt to use the tracked … view at source ↗
Figure 3
Figure 3. We employ the proposed Idx for anchor binding, elimination the motion irrelevant points and simplification the over-small size/opacity points to reduce computational overhead. Specifically, (a) The anchor points are sampled from the mesh and multiple duplicated for detail representation. (b) The motion-aware Idx tends to remove points, as there are none motion gradients around the forehead in canonical space. (c) Th… view at source ↗
Figures from the paper (10 more)
Figure 2
Figure 2. Figure 2: It is worth noting that this explicit geometric deformation [PITH_FULL_IMAGE:figures/full_fig_p004_2.png]
Figure 4
Figure 4. Figure 4: The perceptual evaluation of our method and baselines for self [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: The visualization of efficiency comparison with baseline methods, [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: The visualize of ablation study on Gaussian properties warm-up [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: The visualize of image quality via 3D Gaussian anchor. We demon [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 8
Figure 8. Figure 8: The warm-up phase helps geometry adjustment for 3D Gaussian [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 11
Figure 11. Figure 11: The visualize of the 3D Gaussian simplification ablation study. We [PITH_FULL_IMAGE:figures/full_fig_p009_11.png]
Figure 12
Figure 12. Figure 12: Comparisons in the ablation study on w/ and w/o 3D Gaussian simplification. For w/o simplification, as the number of iterations grows, the time per iteration also increases due to the overhead from redundant points. In contrast, simplified pruning (w/ simplification) …
Figure 14
Figure 14. Figure 14: The visualize of toonification head reconstruction. We apply the [PITH_FULL_IMAGE:figures/full_fig_p010_14.png]
Figure 15
Figure 15. Figure 15: The visualize of limitation of monocular head reconstruction via [PITH_FULL_IMAGE:figures/full_fig_p010_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 6 canonical work pages

  1. [3]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Real-time 3D-aware portrait video relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6221–6231. Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun- Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shoou-I Yu, et al. 2022. Authentic volumetric avatars from a phone scan. ACM T...

    work page 2022

  2. [4]

    Portrait Video Editing Empowered by Multimodal Generative Priors

    Portrait Video Editing Empowered by Multimodal Generative Priors. arXiv preprint arXiv:2409.13591 (2024). Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang

    work page Pith review arXiv 2024

  3. [5]

    ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia) 41, 6 (2022)

    Reconstructing Personalized Semantic Facial NeRF Models From Monocular Video. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia) 41, 6 (2022). https://doi.org/10.1145/3550454.3555501 Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rünz, Lourdes Agapito, and Matthias Nießner. 2024. MonoNPHM: Dynamic Head Reconstruc- tion from Mon...

    arXiv 2022

  4. [6]

    InEuropean conference on computer vision

    KeypointNeRF: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. InEuropean conference on computer vision. Springer, 179–197. Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ra- mamoorthi, and Ren Ng. 2020. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In Compute...

    arXiv 2022

  5. [7]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    3dgs-avatar: Animatable avatars via deformable 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 5020–5030. Haonan Qiu, Zhaoxi Chen, Yuming Jiang, Hang Zhou, Xiangyu Fan, Lei Yang, Wayne Wu, and Ziwei Liu. 2024. Relitalk: Relightable talking portrait generation from a single video. International J...

    work page 2024

  6. [8]

    SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting

    High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10684– 10695. Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. 2024. Splattingavatar: Realistic real-time human avatars with mesh-embedded gaussian splat...

    work page Pith review arXiv 2024

  7. [2022]

    Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh

    Towards Metrical Reconstruction of Human Faces. Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh

    work page

  8. [2023]

    MotionGS: Exploring Explicit Motion Guidance for Deformable 3D Gaussian Splatting

    Pointavatar: Deformable point-based head avatars from videos. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 21057– 21067. Ruijie Zhu, Yanzhe Liang, Hanzhi Chang, Jiacheng Deng, Jiahao Lu, Wenfei Yang, Tianzhu Zhang, and Yongdong Zhang. 2024. MotionGS: Exploring Explicit Motion Guidance for Deformable 3D Gaussian Spla...

    work page Pith review arXiv 2024

Show all 9 references
  1. [2024]

    arXiv preprint arXiv:2404.03613 (2024)

    Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting. arXiv preprint arXiv:2404.03613 (2024). Yoshua Bengio, Nicholas Léonard, and Aaron Courville. 2013. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv ...

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.