Pith. sign in

REVIEW 3 major objections 6 minor 39 references

A single-image head avatar can run in real time with the full driving pipeline inside the model by splitting compute and facial geometry into specialized parts.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 19:54 UTC pith:KRNYWPZH

load-bearing objection Solid systems paper: full-pipeline timing plus a three-branch Gaussian split that ablations actually support; expression fidelity vs external trackers is the real caveat, not a collapse. the 3 major comments →

arxiv 2607.28032 v1 pith:KRNYWPZH submitted 2026-07-30 cs.CV

Split and Drive: Dual-Axis Disentanglement for Real-Time Gaussian Head Avatars

MD Wahiduzzaman Khan , Mingshan Jia , Xiaolin Zhang , En Yu , Kaska Musial-Gabrys This is my paper
classification cs.CV
keywords face reenactment3D Gaussian Splattinghead avatarssingle-image animationFLAMEreal-time renderingdisentanglement
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Single-image photoreal head avatars usually look fast only because facial tracking is left out of the timing, and because one shared 3D representation has to serve every part of the face. This paper argues that both problems come from a failure to specialize, and that dual-axis disentanglement fixes them. On the compute axis, identity shape is estimated once and a lightweight motion encoder is trained end-to-end through the renderer so per-frame driving lives inside the model. On the feature axis, the head is split into three non-overlapping Gaussian branches for full-head geometry, fine facial appearance, and mouth interior. The result, SpiD, matches or beats strong baselines on reenactment quality while reporting the fastest end-to-end speed among the compared methods on one GPU, including driving.

Core claim

Genuine real-time single-image Gaussian head avatars become possible when driving is internalized and trained for rendering quality, and when the avatar is decomposed into three geometrically specialized Gaussian branches rather than one entangled representation; with that design, SpiD is both competitive on quality metrics and the fastest among compared methods when the complete driving pipeline is included.

What carries the argument

Dual-axis disentanglement: compute-axis internalization of per-frame motion (Phase A caches identity shape; Phase B predicts motion through a renderer-supervised encoder) plus feature-axis split into Static, Dynamic, and Mouth Interior Gaussian branches with exclusive spatial coverage.

Load-bearing premise

A lightweight motion encoder trained mainly for how the picture looks can replace external face trackers at inference without systematically hurting expression accuracy or making the speed comparison unfair.

What would settle it

On the same single-GPU setup, measure full end-to-end latency and cross-identity expression metrics when SpiD’s internal motion encoder is swapped for a strong external tracker (or when that encoder is ablated): if quality drops sharply or a baseline with tracking included becomes both faster and more accurate, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Fair head-avatar benchmarks should report inference speed with tracking or driving included, not excluded as preprocessing.
  • Specialized Gaussian branches for mouth interior and mesh-bound periphery can raise open-mouth and peripheral fidelity without a heavier unified network.
  • End-to-end photometric training of driving parameters can make real-time performance an architectural property rather than a reporting choice.
  • A fast no-refiner variant (SpiD*) and a quality variant can share the same split design while spanning different speed–fidelity trade-offs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If internal driving becomes standard, published FPS numbers across avatar papers will compress and reorder once trackers stop being free.
  • The same split—static appearance plane, mesh-bound dynamics, canonical interior sheet—may transfer to audio-driven or full-body Gaussian avatars where topology gaps also appear.
  • Baked source lighting and single-crop shape, which the paper flags as limits, are the next natural axes to disentangle if the dual-axis idea is pushed further.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. SpiD is a single-image 3D Gaussian head-avatar system that argues two architectural disentanglements are jointly necessary for genuine real-time performance and higher fidelity. Along a compute axis, identity shape is estimated once and cached while a lightweight motion encoder (Phase B, Eqs. 4–9) predicts per-frame FLAME-style drivers end-to-end through the differentiable rasterizer, so inference speed includes the full driving pipeline. Along a feature axis, the head is split into three non-overlapping Gaussian branches—Static (DINOv2-conditioned deformable image plane), Dynamic (densified FLAME-bound Gaussians), and Mouth Interior (canonical oral sheet with a per-identity depth cap)—then optionally refined by StyleUNet. On VFHQ cross- and self-reenactment, SpiD reports leading or competitive reconstruction metrics and 43 FPS (154 FPS without the refiner) on one A100 with driving included; ablations in Table 3 attribute identity, appearance, and coverage roles to the Dynamic/Static/interpolation/plane components.

Significance. If the results hold under fair end-to-end accounting, the paper makes a useful systems contribution to single-image Gaussian head avatars: it treats tracker latency as part of the model boundary rather than an excluded preprocess, and it shows that region-specialized Gaussian branches (especially explicit oral coverage) improve fidelity where monolithic 3DGS heads are weak. The latency breakdown (Fig. 3b), dual variants (SpiD / SpiD*), component ablation with failure modes matching stated branch roles (Table 3, Fig. 5), and consistent reporting of complete-pipeline FPS are concrete strengths. The dual-axis framing is more rhetorical than theoretical, but the engineering package—internalized driving plus specialized branches—is timely and actionable for real-time telepresence-style settings.

major comments (3)
  1. [§4.1, Eqs. 4–9; Tables 1–2] §4.1 and Tables 1–2: The central speed-and-quality claim treats Phase-B parameters (Eqs. 4–9) as an adequate drop-in for external FLAME trackers. On self-reenactment (Table 2), however, SpiD’s AKD/AED (5.00/4.83) trail GAGAvatar (4.33/4.03) and GPAvatar (3.90/3.39)—the tracker-based baselines—while SpiD leads on PSNR/SSIM/LPIPS. This pattern is consistent with rendering-friendly Θ_d that are not geometrically equivalent to external tracks, especially given heavy L_aux vertex weight (100) mixed with photometric gradients through the rasterizer (§5.1, Eq. 26). Please add a direct comparison: (i) replace Phase B at test time with the same external tracker used by baselines and report metric deltas; and/or (ii) report geometric error of Phase-B ψ/θ/pose vs. tracker GT on held-out frames. Without this, Fig. 1’s quality–speed quadrant and the “complete pipeline included” ranking partly compare
  2. [§5.2–5.5, Table 1, Fig. 1] §5.2–5.5 and Table 1: Several official-baseline rows look failed rather than competitive (LAM PSNR 2.92 / L1 0.599; CVTHead PSNR 3.27 / L1 0.566), which inflates “first in 7 of 10 metrics” and softens the SOTA claim. Please verify evaluation protocols (resolution, color space, crop, FLAME alignment, whether methods receive the same driving signal), exclude or mark clearly broken runs, and restate wins on the validated subset. Separately, for speed fairness, state explicitly which baselines’ published FPS exclude tracker/crop time and, where possible, re-measure those pipelines with tracking included on the same A100—or report SpiD’s Gaussian-only time alongside full Phase-B time so readers can align accounting conventions.
  3. [§4.2, Eq. 21; §5.6 Table 3] §4.2 Mouth Interior Branch and §5.6 Table 3: The paper repeatedly credits the oral branch for mouth fidelity (abstract, §5.4, Fig. 4), and the depth-cap construction (Eq. 21, Fig. 3a) is a load-bearing design choice against depth drift. Table 3 ablates Dynamic, Static, FLAME interpolation, and the deformable plane, but not the Mouth Interior Branch nor the depth cap. Add a w/o-Mouth and w/o-depth-cap ablation (metrics + open-mouth qualitative) so the third branch is evidenced at the same standard as the other two; otherwise the feature-axis claim is only partially supported.
minor comments (6)
  1. [Title, Fig. 1] Title and running text inconsistently space “A vatars” / “Head A vatars” (title block) and “SpiDproduces” (Fig. 1 caption); clean typography throughout.
  2. [§3–4.1, Eqs. 2 and 9] Eq. (9) writes M(β, ψ, θjaw, θeye) while Eq. (2) defines M(β, ψ, θ); clarify how eye/jaw pose and eyelid blendshape deltas are folded into the FLAME call versus applied afterward.
  3. [§5.3–5.4] ArcF(src)/ArcF(drv) are defined in §5.3 but cross-identity Table 1 omits them; either add identity metrics to Table 1 or explain why only self-reenactment reports ArcF.
  4. [Table 3] Table 3 reports CSIM while Tables 1–2 use ArcF; unify identity-similarity naming and the exact network used.
  5. [§2] Related work cites Instant Expressive and several 2025–2026 arXiv works; ensure citation keys and venue status are accurate at camera-ready and that missing tracker-latency discussion in prior Gaussian avatar papers is attributed with specific citations rather than blanket wording.
  6. [§6] Conclusion correctly flags single-crop shape and baked illumination limits; a short quantitative stress test (profile/occluded sources, cross-lighting drivers) would make those limits falsifiable rather than only qualitative.

Circularity Check

0 steps flagged

No derivation-chain circularity: empirical systems method trained and scored on external reenactment benchmarks.

full rationale

SpiD is an engineering/ML systems paper whose load-bearing claims are architectural (internalized Phase-A/B driving; three non-overlapping Gaussian branches) and empirical (VFHQ cross-/self-reenactment tables; A100 FPS with the full pipeline). Training mixes photometric L1/perceptual losses with geometric anchors L_aux against tracked FLAME parameters and a landmark loss; those anchors supervise the motion encoder during learning but are not re-labeled as held-out scientific predictions, and inference truly drops the external tracker. Reported metrics (AKD, AED, APD, PSNR, SSIM, LPIPS, ArcFace, MUSIQ, TopIQ, FPS) are standard external protocols against third-party baselines, not algebraic restatements of the training objective or fitted inputs. There is no self-definitional loop, no uniqueness theorem imported from overlapping authors, no ansatz smuggled in via self-citation, and no renaming of a known closed-form result. Mild dependence on FLAME pseudo-GT at train time is ordinary supervised learning, not circularity under the stated criteria. Score 0; steps empty.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

Load-bearing structure is standard 3DGS + FLAME avatar practice plus architectural choices and many hand-set hyperparameters. No new physical entities; ‘invented’ items are engineering modules. Claims rest on domain assumptions that FLAME topology (plus a mouth sheet) covers animatable head appearance and that photometric end-to-end training yields usable drivers.

free parameters (5)
  • Loss weights λ1..λ4 and L_aux internals = λ1–4=1.0; aux=100,10,5,10
    Set to 1.0 each for main terms; aux internals 100/10/5/10 for verts/rot/trans/eye (§5.1). Directly shape the trade-off between appearance and geometric anchoring that the compute-axis claim depends on.
  • Static plane grid size n and δ_max = n=296, δ_max=0.05
    n=296 and max deformation 0.05 control Static Branch capacity and how far the deformable plane may move (§4.2, §5.1).
  • Mouth interior grid K×K and depth-cap margin ε = K=32; ε small anatomical margin
    32×32 canonical sheet and anatomical cap formula with margin ε bound oral Gaussians (§4.2); chosen design knobs, not derived.
  • Training schedule (lr, steps, batch) = 1e-4, bs=8, 500k steps
    Adam 1e-4, batch 8, 500k steps on A100 (§5.1); standard fitted optimization regime affecting final metrics.
  • Gaussian color feature dimension (32) and render resolution = 32-dim features; 512×512
    32-channel features and 512×512 images are design choices coupling branches to StyleUNet fusion (§3, §4.3, §5.1).
axioms (5)
  • domain assumption Differentiable 3D Gaussian splatting alpha-compositing is a sufficient real-time renderer for photoreal head avatars.
    Invoked throughout §3–4 via Kerbl et al. formulation (Eq. 1); quality and speed claims presuppose this representation.
  • domain assumption FLAME’s low-dimensional shape/expression/pose space plus eyelid blendshape deltas adequately scaffold driven head geometry for animation.
    Eqs. 2 and 9; Dynamic and Mouth branches bind to FLAME vertices and lip centroids.
  • domain assumption Frozen DINOv2 features provide identity-discriminative appearance cues suitable for Gaussian attribute decoding without finetuning.
    Used in Static/Dynamic/Mouth decoders (§4.2); backbone frozen (§5.1).
  • ad hoc to paper Photometric and perceptual losses on paired same-video frames, with geometric anchors, supervise motion parameters that transfer to cross-identity driving at test time.
    Core compute-axis training story in §4.1 and §4.3; not guaranteed by prior tracker literature.
  • ad hoc to paper A soft FLAME coverage map plus three branch coordinate frames yields spatially non-overlapping, complementary Gaussian support.
    Feature-axis design claim in §4.2; enforced by construction of coverage map and branch domains.
invented entities (3)
  • Dual-axis disentanglement framing (compute axis + feature axis) no independent evidence
    purpose: Organize the paper’s two contributions—internalized driving and multi-branch Gaussians—as one system thesis.
    Rhetorical/architectural construct; not an independently measured natural object. Useful taxonomy, but evidence is the system’s empirical scores.
  • Three specialized Gaussian branches (Static deformable plane, densified Dynamic mesh, Mouth Interior canonical sheet) no independent evidence
    purpose: Assign geometrically distinct facial domains to separate Gaussian populations and decoders.
    New module split relative to monolithic 3DGS avatars cited; validated only via ablations inside this paper.
  • Per-identity mouth depth cap from bowl depth and lip thickness no independent evidence
    purpose: Prevent oral Gaussian depth drift and flickering on open mouths.
    Heuristic anatomical clamp (Eq. 21, Fig. 3a); no external clinical validation.

pith-pipeline@v1.2.0-daily-grok45 · 17517 in / 4223 out tokens · 86602 ms · 2026-07-31T19:54:11.249223+00:00 · methodology

0 comments
read the original abstract

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.

Figures

Figures reproduced from arXiv: 2607.28032 by En Yu, Kaska Musial-Gabrys, MD Wahiduzzaman Khan, Mingshan Jia, Xiaolin Zhang.

Figure 1
Figure 1. Figure 1: SpiD produces high-fidelity avatar animation from a single image across di￾verse identities and expressions (left). Among all compared methods, SpiD achieves the strongest quality-speed trade-off, placing in the real-time high-fidelity quadrant with the complete driving pipeline included (right). SpiD* denotes our speed variant without neural refinement. Abstract. Creating photorealistic animatable head av… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SpiD. Given a single source image and a driving frame, SpiD operates along two disentanglement axes. Along the compute axis, Phase A estimates identity shape once per subject via a frozen shape encoder, while Phase B predicts per-frame motion parameters Θd via a lightweight motion encoder. Along the fea￾ture axis, three specialized Gaussian branches process complementary spatial domains: the St… view at source ↗
Figure 3
Figure 3. Figure 3: (a) Per-identity depth bound for the Mouth Interior Branch. Side view of the source FLAME geometry showing inner lip vertices (circles), outer lip vertices (diamonds), and canonical mouth Gaussians (dots). Two anatomical quantities define the admissible depth interval: dbowl, the distance from the canonical interior sheet to the inner lip opening, and dlip, the front-to-back thickness of the upper lip. The… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison against state-of-the-art methods on cross-identity reen￾actment. SpiD produces more faithful expression transfer and sharper appearance re￾construction across diverse identities. plane with a fixed grid degrades perceptual quality (LPIPS 0.178 → 0.206), con￾firming that content-aware plane deformation improves local appearance cap￾ture. 6 Conclusion We presented SpiD, a real-time sin… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative ablation results. Removing the Static Branch washes out fine fa￾cial detail. Removing the Dynamic Branch causes identity collapse. Removing FLAME mesh interpolation produces visible mesh artifacts on the face surface. Replacing the Deformable Plane with a fixed grid reduces local appearance fidelity. The Mouth Inte￾rior Branch contribution is most evident in open-mouth cross-reenactment scenari… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 5 linked inside Pith

  1. [1]

    In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp

    Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 157–164 (2023)

    2023

  2. [2]

    In: Proceedings of the IEEE international conference on computer vision

    Bulat, A., Tzimiropoulos, G.: How far are we from solving the 2d & 3d face align- ment problem?(and a dataset of 230,000 3d facial landmarks). In: Proceedings of the IEEE international conference on computer vision. pp. 1021–1030 (2017)

    2017

  3. [3]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16123–16133 (2022)

    2022

  4. [4]

    IEEE Transactions on Image Processing33, 2404–2418 (2024)

    Chen, C., Mo, J., Hou, J., Wu, H., Liao, L., Sun, W., Yan, Q., Lin, W.: Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing33, 2404–2418 (2024)

    2024

  5. [5]

    Advances in Neural Information Processing Systems37, 57642–57670 (2024)

    Chu, X., Harada, T.: Generalizable and animatable gaussian head avatar. Advances in Neural Information Processing Systems37, 57642–57670 (2024)

    2024

  6. [6]

    arXiv preprint arXiv:2401.10215 (2024)

    Chu, X., Li, Y., Zeng, A., Yang, T., Lin, L., Liu, Y., Harada, T.: Gpavatar: Gener- alizable and precise head avatar from image (s). arXiv preprint arXiv:2401.10215 (2024)

    Pith/arXiv arXiv 2024

  7. [7]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Daněček, R., Black, M.J., Bolkart, T.: Emoca: Emotion driven monocular face capture and animation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 20311–20322 (2022)

    2022

  8. [8]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690–4699 (2019)

    2019

  9. [9]

    In: European Conference on Computer Vision

    Deng, Y., Wang, D., Wang, B.: Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. In: European Conference on Computer Vision. pp. 316–333. Springer (2024)

    2024

  10. [10]

    ACM Transactions on Graphics (ToG)40(4), 1–13 (2021)

    Feng, Y., Feng, H., Black, M.J., Bolkart, T.: Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG)40(4), 1–13 (2021)

    2021

  11. [11]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Grassal, P.W., Prinzler, M., Leistner, T., Rother, C., Nießner, M., Thies, J.: Neu- ral head avatars from monocular rgb videos. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18653–18664 (2022)

    2022

  12. [12]

    arXiv preprint arXiv:2504.14373 (2025)

    Guo, C., Su, Z., Wang, L., Wang, J., Li, S., Chang, X., Li, Z., Zhao, Y., Wang, G., Liu, Y., et al.: Sega: Drivable 3d gaussian head avatar from a single image. arXiv preprint arXiv:2504.14373 (2025)

    arXiv 2025

  13. [13]

    arXiv preprint arXiv:2407.03168 (2024)

    Guo, J., Zhang, D., Liu, X., Zhong, Z., Zhang, Y., Wan, P., Zhang, D.: Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168 (2024)

    Pith/arXiv arXiv 2024

  14. [14]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Guo, Y., Chen, K., Liang, S., Liu, Y.J., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5784–5794 (2021)

    2021

  15. [15]

    In: Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers

    He, Y., Gu, X., Ye, X., Xu, C., Zhao, Z., Dong, Y., Yuan, W., Dong, Z., Bo, L.: Lam: large avatar model for one-shot animatable gaussian head. In: Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers. pp. 1–13 (2025)

    2025

  16. [16]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Hong, Y., Peng, B., Xiao, H., Liu, L., Zhang, J.: Headnerf: A real-time nerf-based parametric head model. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 20374–20384 (2022) 16 M. W. Khan et al

    2022

  17. [17]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314–1324 (2019)

    2019

  18. [18]

    ACM Transactions on Graphics (ToG)36(6), 1–14 (2017)

    Hu, L., Saito, S., Wei, L., Nagano, K., Seo, J., Fursund, J., Sadeghi, I., Sun, C., Chen, Y.C., Li, H.: Avatar digitization from a single image for real-time rendering. ACM Transactions on Graphics (ToG)36(6), 1–14 (2017)

    2017

  19. [19]

    arXiv preprint arXiv:2512.16893 (2025)

    Jiang, K., Li, X., Park, S., Ramamoorthi, R., De Mello, S., Nagano, K.: Instant expressive gaussian head avatar via 3d-aware expression distillation. arXiv preprint arXiv:2512.16893 (2025)

    Pith/arXiv arXiv 2025

  20. [20]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5148–5157 (2021)

    2021

  21. [21]

    ACM Trans

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G., et al.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph.42(4), 139–1 (2023)

    2023

  22. [22]

    In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition

    Kirschstein, T., Giebenhain, S., Nießner, M.: Flexavatar: Learning complete 3d head avatars with partial supervision. In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition. pp. 18193–18203 (2026)

    2026

  23. [23]

    ACM Transactions on Graphics (TOG)42(4), 1–14 (2023)

    Kirschstein, T., Qian, S., Giebenhain, S., Walter, T., Nießner, M.: Nersemble: Multi-view radiance field reconstruction of human heads. ACM Transactions on Graphics (TOG)42(4), 1–14 (2023)

    2023

  24. [24]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Kirschstein, T., Romero, J., Sevastopolsky, A., Nießner, M., Saito, S.: Avat3r: Large animatable gaussian reconstruction model for high-fidelity 3d head avatars. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12089–12100 (2025)

    2025

  25. [25]

    ACM Trans

    Li, T., Bolkart, T., Black, M.J., Li, H., Romero, J.: Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph.36(6), 194–1 (2017)

    2017

  26. [26]

    In: 2024 International Conference on 3D Vision (3DV)

    Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 2024 International Conference on 3D Vision (3DV). pp. 800–809. IEEE (2024)

    2024

  27. [27]

    Commu- nications of the ACM65(1), 99–106 (2021)

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Commu- nications of the ACM65(1), 99–106 (2021)

    2021

  28. [28]

    arXiv preprint arXiv:2304.07193 (2023)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

    Pith/arXiv arXiv 2023

  29. [29]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Ren, Y., Li, G., Chen, Y., Li, T.H., Liu, S.: Pirenderer: Controllable portrait im- age generation via semantic neural rendering. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13759–13768 (2021)

    2021

  30. [30]

    arXiv preprint arXiv:2404.04104 (2024)

    Retsinas, G., Filntisis, P.P., Danecek, R., Abrevaya, V.F., Roussos, A., Bolkart, T., Maragos, P.: Smirk: 3d facial expressions through analysis-by-neural-synthesis. arXiv preprint arXiv:2404.04104 (2024)

    Pith/arXiv arXiv 2024

  31. [31]

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

    Shi, Y., Li, Z., Wang, Y., Zhu, H., Cao, X., Liu, L.: Dex-portrait: Disentangled and expressive portrait animation via explicit and latent motion representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 40009–40019 (2026)

    2026

  32. [32]

    Advances in neural information processing systems32 (2019)

    Siarohin, A., Lathuilière, S., Tulyakov, S., Ricci, E., Sebe, N.: First order motion model for image animation. Advances in neural information processing systems32 (2019)

    2019

  33. [33]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wang, T.C., Mallya, A., Liu, M.Y.: One-shot free-view neural talking-head syn- thesis for video conferencing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10039–10049 (2021) SpiD: Dual-Axis Disentanglement for Gaussian Head Avatars 17

    2021

  34. [34]

    IEEE transactions on image processing 13(4), 600–612 (2004)

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600–612 (2004)

    2004

  35. [35]

    In: The thrity-seventh asilomar conference on signals, systems & computers, 2003

    Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The thrity-seventh asilomar conference on signals, systems & computers, 2003. vol. 2, pp. 1398–1402. Ieee (2003)

    2003

  36. [36]

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

    Xie, L., Wang, X., Zhang, H., Dong, C., Shan, Y.: Vfhq: A high-quality dataset and benchmark for video face super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 657–666 (2022)

    2022

  37. [37]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 20331– 20341 (2024)

    2024

  38. [38]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586–595 (2018)

    2018

  39. [39]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation rep- resentations in neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5745–5753 (2019)

    2019