Pith. sign in

REVIEW 4 major objections 6 minor 52 references

SemFaceEdit: Semantic Face Editing on Generative Radiance Manifolds

T0 review · 4 major / 6 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 SemFaceEdit claims that a 3D generative face model can be made locally editable by giving each facial semantic its own appearance latent code and a shared geometry latent code, with a Semantic Volume Masking layer that sorts every sampled…

desk verdict SemFaceEdit adds a real capability to 3D GAN editing but leaves the load-bearing semantic-assignment assumption untested and skips the closest baseline. read the letter →

arxiv 2506.22833 v1 pith:PVRIENQN submitted 2025-06-28 cs.CV

Shashikant Verma , Shanmuganathan Raman This is my paper
classification cs.CV
keywords Neuralradiancefields3D-awareGANssemanticfaceeditinggenerativemanifoldslatentspacedisentanglementimageinversionappearancetransfergeometry
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

SemFaceEdit claims that a 3D generative model of faces can make geometry and appearance independently editable per facial semantic by giving each semantic its own appearance latent code while sharing one geometry latent code. The method sits on generative radiance manifolds, which confine point sampling to learned two-dimensional surfaces, and adds a Semantic Volume Masking layer that labels each sampled point with a semantic class so that later colour prediction sees only points of that class. If the claim holds, swapping a hairstyle, a smile, or a nose between portraits reduces to exchanging the matching latent code plus a short inversion, with other regions left intact. The paper reports that edits stay view-consistent and that its inversion reaches 0.85 mIoU within 50 iterations, compared with 5,000 iterations for a prior NeRF editing baseline.

What carries the argument

The central mechanism is the Semantic Volume Masking layer, defined by the assignment $S_k(r) = \arg\max_k \sum_{j=k}^K T(x_j)\alpha(x_j)s(x_j,d)$ for points on the $k$-th of $K$ radiance manifolds along a ray, which partitions all sampled points and their appearance descriptors into $n$ semantic collections. Each collection is then processed by a separate mapping network inside the Appearance Module, so that the $i$-th appearance latent $z_i$ conditions only the $i$-th semantic's descriptors; the shared Geometry Module's latent $z$ controls occupancy and semantic layout. This segregation is what turns a global radiance field into locally editable regions.

What would settle it

Render a view of a real image inverted with SemFaceEdit, then compare the semantic label assigned to each point by Eq. 2 against a ground-truth 3D semantic volume obtained from a fitted morphable face model or a synthetic rig; if masking mislabels points at region boundaries under view changes, the per-semantic latent code will also alter neighbouring regions, which can be measured as a per-region SSIM drop when only one $z_i$ is changed.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that semantic control over a radiance field can be obtained without retraining and without global appearance changes. SemFaceEdit trains a Geometry Module to produce, for each point on a radiance manifold, an occupancy value, a semantic radiance vector, and a high-dimensional appearance descriptor, and an Appearance Module that predicts RGB colour after the Semantic Volume Masking layer has partitioned points by Eq. 2's $\arg\max$ over accumulated transmittance-weighted semantic radiance. Because each semantic class is conditioned by its own mapping network, the $i$-th appearance latent $z_i$ controls only the colour of the $i$-th region, and the geometry latent $z$ controls occupancy and semantic layout. The paper demonstrates that this allows transferring the appearance or geometry of one semantic from a source to a target image by swapping the corresponding latent code and solving a mask-and-image inversion objective, while other regions are preserved.

Load-bearing premise

The method assumes that the per-point assignment of 3D locations to facial regions is correct everywhere, even though it is learned only from flat 2D masks with no depth information, so that swapping one region's latent code never alters another region.

Editorial extensions

If this is right

  • Portrait attribute transfer becomes a latent-code swap: to give image A the hair of image B, exchange the hair appearance latent code and run a short inversion, leaving face and background latent codes fixed.
  • Semantic mask editing (e.g., expanding hair or shrinking mouth) changes geometry only in the edited region, with other regions preserved.
  • Each appearance latent $z_i$ controls only colour and texture of its semantic region, so hair colour, face tone, and garment colour can be varied independently and composed freely.
  • Because both semantic and RGB radiance fields live in the same volume, every edit remains multi-view consistent.
  • Inversion to $W$ space takes on the order of $10^2$ iterations in the paper's experiments, compared with $5\times10^3$ for FENeRF, making these edits practical on a single image.

Reading between the lines

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

  • If the masking is as clean as claimed, the same per-semantic latent architecture could carry over to full-body avatars or object categories, where the semantic classes are swapped for body parts or object components.
  • The hard $\arg\max$ in Eq. 2 is a natural place to expect boundary artifacts; a soft or probabilistic assignment could improve hair fidelity, which the paper itself lists as a limitation.
  • One could quantify disentanglement by measuring per-region SSIM and LPIPS before and after changing a single $z_i$; the paper shows qualitative results but does not report such a per-region fidelity metric.
  • The approach suggests that text-conditioned editing could be attached to each semantic latent manifold, since each region already has an isolated low-dimensional code.
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

4 major / 6 minor

Summary. The paper introduces SemFaceEdit, a 3D-aware GAN for semantic face editing built on generative radiance manifolds. The generator is decomposed into a Geometry Module that predicts occupancy, semantic radiance, and appearance descriptors, and an Appearance Module that takes appearance descriptors segregated by a Semantic Volume Masking layer and conditions them on per-semantic latent codes before predicting RGB radiance. Training is two-stage and adversarial on CelebAMask-HQ. The paper claims precise localized editing of particular facial semantics while preserving other regions, and reports qualitative results for appearance transfer, geometry transfer, and semantic-mask editing, along with FID/KID for generated images and an mIoU for inversion. The central promise is per-semantic latent control over both appearance and geometry.

Significance. If the claims hold, SemFaceEdit would offer a practical capability: semantic-specific appearance and geometry control in a 3D-aware radiance-field GAN without retraining, built on the efficient manifold sampling of GRAM. The architecture is modular and internally plausible, the two-stage training recipe is concrete, and the qualitative results, especially the inversion convergence compared with FE-NeRF, are promising. The paper also avoids the need for 3D scans during training. However, the central claim of precise locality is not yet supported by quantitative evidence: generation quality is measured, but editing quality and disentanglement are not. Moreover, one of the core equations appears to define semantic labeling in a way that is inconsistent with the stated preservation guarantee. The contribution is therefore significant in scope but currently under-validated.

major comments (4)
  1. [§3.1, Eq. (2)] Equation (2) computes the semantic label S_k(r) as the argmax of the accumulated semantic radiance from the k-th manifold point to the end of the ray, rather than of the local semantic radiance s(x_k) at that point. Because a later surface with a high semantic response can dominate the suffix sum, an earlier point can be relabeled as belonging to a different semantic class. The Semantic Volume Masking layer would then place that earlier point and its appearance descriptor F into the wrong semantic collection, so editing the latent code of the dominant class could change radiance at points that actually belong to another class. This directly threatens the advertised preservation of non-edited regions. In addition, the hard argmax in Eq. (2) has zero gradient, while Contribution 2 calls this layer "differentiable"; the text does not describe a straight-through estimator or a soft relaxation. Please correct the labeling rule or justify the suffix-sum rule explicitly, and verify per-point label fidelity.
  2. [Contributions and §4.2] The contributions claim semantic-specific control over geometry by manipulating latent spaces specific to each semantic attribute, but the Geometry Module in Eq. (1) takes a single global geometric latent z and outputs a global occupancy field σ; there is no per-semantic geometry latent code. Geometry editing in §4.2 is performed by optimizing a global offset δw+ against a composite semantic mask, not by swapping a semantic-specific geometry code. Consequently, the claim of per-semantic geometry latent control is not supported by the architecture as described, and the preservation of non-edited geometry relies entirely on the optimization not disturbing other regions. Please either introduce per-semantic geometry latents or reframe the claim, and provide a quantitative test of preservation of non-edited geometry during geometry transfer.
  3. [§4.3, Table 1] All reported quantitative metrics, FID and KID, evaluate unconditional generation quality, not editing. The paper's central claim is precise local editing with preservation of other semantic regions, and the evidence for this is exclusively qualitative in Figures 4-8. I request a quantitative editing protocol: for each semantic latent swap, measure semantic mIoU and image-similarity or identity preservation over the regions that should remain unchanged, and measure attribute classification accuracy or mask agreement over the edited region. A comparison with FE-NeRF and IDE-3D under the same protocol would be particularly valuable. Without such measurements, the claimed superiority in "radiance field disentanglement" is not established.
  4. [§4.3, Fig. 9 and Appendix 2.1] The reported 0.85 mIoU for inversion measures agreement between the generator's rendered semantic mask and the target mask that is itself the objective of the inversion optimization in Eq. (10). This is an optimization-fit metric, not a measure of the correctness of the per-point semantic assignments that the Semantic Volume Masking layer relies on. Since the paper's own Limitations section concedes that "any discrepancies in the semantic geometry propagates to final facial appearance," a direct check is needed: compare the Eq. (2) classification of each sampled point with the local semantic radiance s(x_k), or with a denser reference labeling, and report agreement as a function of depth and semantic class. Without this, the locality premise of the entire editing pipeline remains untested.
minor comments (6)
  1. [§3.1, Eq. (2)] The transmittance product writes T(x_j)=∏_{i<j;i≥k}(1−σ(x_j)), but the factor should be σ(x_i); as written the product does not depend correctly on the intermediate points.
  2. [§4.1] Training for 120K iterations on 30K images is 4 epochs, not 3; the second stage of 30K iterations is 1 epoch. Please correct the epoch count or the iteration count.
  3. [§4.2 and Fig. 7 caption] The text states the method converges to 0.85 mIoU within 50 iterations, while the Fig. 7 caption says "rapid convergence within 500 iterations"; these numbers should be reconciled.
  4. [Appendix, Eq. (3)] Under the stated softplus convention f(x)=log(1+exp(x)), the generator terms use f(D_s(...)) while the discriminator terms use f(−D(...)); this is not the standard non-saturating GAN formulation. Please clarify the sign convention or the intended loss direction.
  5. [§3.1] The assertion that learning semantic radiance from masks alone "is bound to fail" without the appearance-descriptor pathway is not supported by an ablation. Either provide the isolating ablation or soften the claim.
  6. [§4.3 and Related Work] IDE-3D [38] is discussed in Related Work as an editing method for tri-planes but is not compared in the experiments; a comparison or a brief explanation of its omission would strengthen the evaluation.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity: the central editing mechanism is an empirical architecture, and the only self-referential element is the reported inversion mIoU, which measures fit to the optimization target rather than external semantic accuracy.

  1. fitted input called prediction [Section 4.3 (Quantitative comparison) and Appendix Eq. (10)]
    "To assess the 3D inversion capability of SemFaceEdit, we perform inversions on 1K images from the CelebAMask-HQ Dataset, calculating the mean Intersection over Union (mIoU) for all n semantics. Across all inverted images, our approach converges to 0.85 mIoU within 50 iterations. ... L(δw+, δw+ i ) = λsLs(S′, Sed) + λimLim(I ′ ⊙ (1 − r), I ⊙ (1 − r)) + λvgg Lvgg (I ′ ⊙ (1 − r), I ⊙ (1 − r)) (10)"

    The 0.85 mIoU is computed between the generator's rendered semantic mask and the same mask used as the target in the inversion loss (Ls in Eq. 9/10). It therefore certifies that the optimization matched its own objective, not that the per-point semantic labels in Eq. 2 are correct. This is a self-referential performance number, but it is not load-bearing for the central editing claim.

full rationale

The paper's core pipeline is an empirical architecture: Geometry and Appearance modules are trained adversarially, semantic masking is defined by Eq. 2, and editing is performed by swapping latent codes and optimizing an inversion offset. No equation reduces to its own input, and no fitted parameter is later presented as an external prediction. The only self-referential element is the headline inversion mIoU, which measures agreement with the mask used as the inversion objective; that is a standard reconstruction metric rather than a circular derivation. The Limitations section openly concedes semantic-geometry discrepancies, which is a correctness risk, not circularity. No load-bearing self-citations appear: references to GRAM, FENeRF, and NeRFFaceEditing are external works. Score 2 reflects the minor self-referential metric.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper pulls the GRAM manifold predictor, FiLM SIREN backbone, volumetric rendering equations, and the pivotal-tuning inversion procedure from prior work; those are listed as axioms and domain assumptions. The central claim rests on one ad hoc assumption: that per-point hard semantic assignment (Eq. 2) is accurate enough despite training on projected 2D masks with no depth, an assumption the authors themselves flag as fragile in the Limitations. There are no invented physical entities: the Semantic Volume Masking layer is a network component, not a new particle, force, or conserved quantity. No central-claim-defining scalar is fitted to data; the numbers listed as free parameters are hand-set hyperparameters and a hand-chosen semantic grouping.

free parameters (3)
  • Semantic clubbing (19 to 4 classes) = n = 4: hair, face, garment, background
    The 19 CelebAMask-HQ labels are hand-grouped into 4 semantics; all per-semantic latent codes and the claimed editing granularity are defined by this choice (Section 4.1).
  • Loss weights (stage 1, stage 2, inversion) = lam_im=5, lam_s=1, lam_p=10, lam_l=1; then lam_im=1, lam_p=10; inversion lam_s=10, lam_im=1, lam_vgg=1
    Hand-set in Eqs. 3-11; they influence the reported FID and mIoU numbers but are not fitted constants on which the central claim reduces.
  • Network depth (FiLM layers) = 8 layers in both modules
    Selected from an ablation over depths 4, 6, 8 (Table 3) by FID; a hand choice governing quality versus compute.
assumptions (5)
  • standard math Volumetric rendering equations (transmittance, opacity accumulation) apply to points sampled on 2D manifolds
    Invoked in Eq. 2 and final RGB integration via [30, 47] and the GRAM pipeline [9]; used without derivation.
  • domain assumption Adversarial training on 2D images yields a multi-view-consistent 3D radiance field
    The standard 3D GAN assumption inherited from pi-GAN/GRAM [7, 9]; the paper verifies consistency only through qualitative multi-view figures (Figure 12).
  • domain assumption The GRAM manifold predictor supplies sufficient iso-surfaces (K=24) for semantic and RGB radiance learning
    Taken wholesale from [9] (Section 3.1, Manifold Predictor); no ablation of K is presented.
  • ad hoc to paper Semantic radiance can be learned from 2D masks if appearance descriptors F provide gradient flow into the geometry module
    Section 3.1 asserts mask-only learning 'is bound to fail' and that the F pathway fixes it, but the isolating ablation is not in the text.
  • ad hoc to paper The hard arg-max assignment in Eq. 2 labels points correctly; semantic-geometry errors are small enough not to corrupt appearance
    The Semantic Volume Masking layer depends on this; the paper's Limitations concede 'any discrepancies in the semantic geometry propagates to final facial appearance.'

how reviews work

0 comments
Cite this review

Pith. "Pith review of SemFaceEdit: Semantic Face Editing on Generative Radiance Manifolds." pith.science (2026). https://pith.science/paper/PVRIENQN

@misc{pith2026250622833,
  author       = {Pith},
  title        = {Pith review of: SemFaceEdit: Semantic Face Editing on Generative Radiance Manifolds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PVRIENQN}},
  note         = {Machine review of arXiv:2506.22833}
}
read the original abstract

Despite multiple view consistency offered by 3D-aware GAN techniques, the resulting images often lack the capacity for localized editing. In response, generative radiance manifolds emerge as an efficient approach for constrained point sampling within volumes, effectively reducing computational demands and enabling the learning of fine details. This work introduces SemFaceEdit, a novel method that streamlines the appearance and geometric editing process by generating semantic fields on generative radiance manifolds. Utilizing latent codes, our method effectively disentangles the geometry and appearance associated with different facial semantics within the generated image. In contrast to existing methods that can change the appearance of the entire radiance field, our method enables the precise editing of particular facial semantics while preserving the integrity of other regions. Our network comprises two key modules: the Geometry module, which generates semantic radiance and occupancy fields, and the Appearance module, which is responsible for predicting RGB radiance. We jointly train both modules in adversarial settings to learn semantic-aware geometry and appearance descriptors. The appearance descriptors are then conditioned on their respective semantic latent codes by the Appearance Module, facilitating disentanglement and enhanced control. Our experiments highlight SemFaceEdit's superior performance in semantic field-based editing, particularly in achieving improved radiance field disentanglement.

Figures

Figures reproduced from arXiv: 2506.22833 by the authors.

Figure 1
Figure 1. An overview of our proposed framework. We sample points in volume by de￾termining intersections of casted rays with isosurfaces predicted by Manifold Predictor [9]. Subsequently, the Geometry Module conditions these points using a latent vector sampled from a Gaussian distribution, resulting in diverse predictions for occupancy (σ), semantic radiance, and Appearance Descriptor F. To segregate points and their appear… view at source ↗
Figure 2
Figure 2. The network architecture of Geometry Module and Appearance Module [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Renderings of Semantic-radiance and RGB-radiance on image space generated by our approach by random latent code z ∈ R d and zi ∈ R d . mation from the geometry module to control the appearance of each semantic point on these manifolds in the volume. 3D-aware Neural Face Image Synthesis. The combination of Generative Adversarial Networks (GANs), as originally proposed [13], and neural implicit radiance fields (NeRF) … view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Latent space disentanglement and interpolations. In (a), linear interpolation is performed on the geometric latent space z while keeping all appearance latent zi fixed. (b-d) demonstrate changes in appearance latent variables corresponding to hair, face, and background…
Figure 5
Figure 5. Figure 5: Appearance Latent Codes (zi) influence on Geometric Shapes. The image (i,j) in grid (a),(b), and (c) are generated with same zi’s, highlighting resemblance in ap￾pearance across hair, background, and facial regions with (i,j)th image in each grid. 4 Experiments and Res…
Figure 6
Figure 6. Figure 6: (a-f) displays two smaller images, with the top image representing the source attribute and the bottom image representing the target attribute of appear￾ance/geometry. Appearance Transfer: In (a), we present three different views of the source image showcasing the tran…
Figure 7
Figure 7. Figure 7: Comparison of pivotal inversion results between our proposed approach (b,e) and FE-NerF [39] (a,d). Our method achieves rapid convergence within 500 iterations, contrasting with FE-NerF’s 5K iterations. Additionally, we compare with Diffusion￾Rig [11] (c,f), which firs…
Figure 8
Figure 8. Figure 8: Qualitative comparison of SemFaceEdit with Co-Diff [16], Diff-Rig [11], SofGAN [8], GRAM [9], and FE-NerF [39]. ⊠ represents that the respective approach don’t support or yield the required result. Input mask in [16] and image in [11] are taken from CelebAMask-HQ Datas…
Figure 9
Figure 9. Figure 9: During the optimization process for 3D inversion on 1000 CelebAMask-HQ Dataset images, we track the Mean Intersection over Union (mIoU). Embedded images qualitatively illustrate the inversion progression of an image in a challenging pose. Furthermore, we track the Mean…
Figure 10
Figure 10. Figure 10: Effect of manipulation in the semantic mask by expanding the hair region semantics and the effect of shrinking at the mouth region compared with Co-Diff [16] [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Artifacts due to discrepancy in semantic radiance (a) and Hair quality ob￾tained by different methods (c), (d) compared with Ours (b). position. In contrast Co-Diff [16] employs a diffusion-based model to generate image in 2D space. 3.2 Limitations The generation of R…
Figure 12
Figure 12. Figure 12: Renderings of Semantic-radiance and RGB-radiance on image space generated by our approach by random latent code z ∈ R d and zi ∈ R d . using images sourced from the web or repositories, the facial images must first be aligned as previously discussed. Subsequently, a s…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 32 canonical work pages

  1. [1]

    ACM Transactions on Graphics (ToG)40(3), 1–21 (2021)

    Abdal, R., Zhu, P., Mitra, N.J., Wonka, P.: Styleflow: Attribute-conditioned ex- ploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG)40(3), 1–21 (2021)

    2021

  2. [2]

    arXiv preprint arXiv:2303.13071 (2023)

    An, S., Xu, H., Shi, Y., Song, G., Ogras, U., Luo, L.: Panohead: Geometry-aware 3d full-head synthesis in 360. arXiv preprint arXiv:2303.13071 (2023)

    arXiv 2023

  3. [3]

    In: 2023 IEEE 17th International Conference on Auto- matic Face and Gesture Recognition (FG)

    Athar, S., Shu, Z., Samaras, D.: Flame-in-nerf: Neural control of radiance fields for free view face animation. In: 2023 IEEE 17th International Conference on Auto- matic Face and Gesture Recognition (FG). pp. 1–8. IEEE (2023)

    work page 2023

  4. [4]

    arXiv preprint arXiv:1801.01401 (2018)

    Bińkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. arXiv preprint arXiv:1801.01401 (2018)

    arXiv 2018

  5. [5]

    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)

    work page 2017

  6. [6]

    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)

    work page 2022

  7. [7]

    In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition

    Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5799–5809 (2021)

    2021

  8. [8]

    ACM Transactions on Graphics41(1) (2022) SemFaceEdit: Semantic Face Editing on Generative Radiance Manifolds 15

    Chen, A., Liu, R., Xie, L., Chen, Z., Su, H., Yu, J.: Sofgan: A portrait image generator with dynamic styling. ACM Transactions on Graphics41(1) (2022) SemFaceEdit: Semantic Face Editing on Generative Radiance Manifolds 15

    work page 2022

Show all 52 references
  1. [9]

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

    Deng, Y., Yang, J., Xiang, J., Tong, X.: Gram: Generative radiance manifolds for 3d-aware image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10673–10683 (2022)

    2022

  2. [10]

    In: IEEE Computer Vision and Pattern Recognition Workshops (2019)

    Deng, Y., Yang, J., Xu, S., Chen, D., Jia, Y., Tong, X.: Accurate 3d face reconstruc- tion with weakly-supervised learning: From single image to image set. In: IEEE Computer Vision and Pattern Recognition Workshops (2019)

    2019

  3. [11]

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

    Ding, Z., Zhang, X., Xia, Z., Jebe, L., Tu, Z., Zhang, X.: Diffusionrig: Learning personalized priors for facial appearance editing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12736–12746 (2023)

    2023

  4. [12]

    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

  5. [13]

    Advances in neural infor- mation processing systems27 (2014)

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural infor- mation processing systems27 (2014)

    2014

  6. [14]

    arXiv preprint arXiv:2110.08985 (2021)

    Gu, J., Liu, L., Wang, P., Theobalt, C.: Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985 (2021)

    2021 arXiv

  7. [15]

    Advances in neural information processing systems30 (2017)

    Heusel,M.,Ramsauer,H.,Unterthiner,T.,Nessler,B.,Hochreiter,S.:Ganstrained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems30 (2017)

    2017

  8. [16]

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

    Huang, Z., Chan, K.C., Jiang, Y., Liu, Z.: Collaborative diffusion for multi-modal face generation and editing. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 6080–6090 (2023)

    2023

  9. [17]

    In: SIGGRAPH Asia 2022 Conference Papers

    Jiang, K., Chen, S.Y., Liu, F.L., Fu, H., Gao, L.: Nerffaceediting: Disentangled face editing in neural radiance fields. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1–9 (2022)

    2022

  10. [18]

    arXiv preprint arXiv:2112.03517 (2021)

    Jo, K., Shim, G., Jung, S., Yang, S., Choo, J.: Cg-nerf: Conditional generative neural radiance fields. arXiv preprint arXiv:2112.03517 (2021)

    2021 arXiv

  11. [19]

    In: Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14

    Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694–711. Springer (2016)

    2016

  12. [20]

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

    Kim, G., Kwon, T., Ye, J.C.: Diffusionclip: Text-guided diffusion models for robust image manipulation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2426–2435 (2022)

    2022

  13. [21]

    arXiv preprint arXiv:1412.6980 (2014)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  14. [22]

    In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)

    Lee, C.H., Liu, Z., Wu, L., Luo, P.: Maskgan: Towards diverse and interactive facial image manipulation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2020)

    2020

  15. [23]

    arXiv preprint arXiv:2109.09378 (2021)

    Leimkühler, T., Drettakis, G.: Freestylegan: Free-view editable portrait rendering with the camera manifold. arXiv preprint arXiv:2109.09378 (2021)

    2021 arXiv

  16. [24]

    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

  17. [25]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2020)

    Liao, Y., Schwarz, K., Mescheder, L., Geiger, A.: Towards unsupervised learning of generative models for 3d controllable image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2020)

    2020

  18. [26]

    Mescheder, L., Geiger, A., Nowozin, S.: Which training methods for gans do ac- tually converge? In: International conference on machine learning. pp. 3481–3490. PMLR (2018) 16 S. Verma et al

    2018

  19. [27]

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

    Michalkiewicz, M., Pontes, J.K., Jack, D., Baktashmotlagh, M., Eriksson, A.: Im- plicit surface representations as layers in neural networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4743–4752 (2019)

    2019

  20. [28]

    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

  21. [29]

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

    Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11453–11464 (2021)

    2021

  22. [30]

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

    Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589–5599 (2021)

    2021

  23. [31]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

    Or-El, R., Luo, X., Shan, M., Shechtman, E., Park, J.J., Kemelmacher-Shlizerman, I.: Stylesdf: High-resolution 3d-consistent image and geometry generation. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 13503–13513 (2022)

    2022

  24. [32]

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

    Park, T., Liu, M.Y., Wang, T.C., Zhu, J.Y.: Semantic image synthesis with spatially-adaptive normalization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2337–2346 (2019)

    2019

  25. [33]

    In: 2009 sixth IEEE interna- tional conference on advanced video and signal based surveillance

    Paysan, P., Knothe, R., Amberg, B., Romdhani, S., Vetter, T.: A 3d face model for pose and illumination invariant face recognition. In: 2009 sixth IEEE interna- tional conference on advanced video and signal based surveillance. pp. 296–301. Ieee (2009)

    2009

  26. [34]

    ACM Transactions on graphics42(1), 1–13 (2022)

    Roich, D., Mokady, R., Bermano, A.H., Cohen-Or, D.: Pivotal tuning for latent- based editing of real images. ACM Transactions on graphics42(1), 1–13 (2022)

    2022

  27. [35]

    arXiv preprint arXiv:1409.1556 (2014)

    Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)

    2014 arXiv

  28. [36]

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

    Sitzmann, V., Thies, J., Heide, F., Nießner, M., Wetzstein, G., Zollhofer, M.: Deepvoxels: Learning persistent 3d feature embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2437– 2446 (2019)

    2019

  29. [37]

    In: Proceedings of the ieee/cvf conference on computer vision and pattern recognition

    Sun, J., Deng, Q., Li, Q., Sun, M., Ren, M., Sun, Z.: Anyface: Free-style text- to-face synthesis and manipulation. In: Proceedings of the ieee/cvf conference on computer vision and pattern recognition. pp. 18687–18696 (2022)

    2022

  30. [38]

    ACM Transactions on Graphics (ToG)41(6), 1–10 (2022)

    Sun, J., Wang, X., Shi, Y., Wang, L., Wang, J., Liu, Y.: Ide-3d: Interactive disen- tangled editing for high-resolution 3d-aware portrait synthesis. ACM Transactions on Graphics (ToG)41(6), 1–10 (2022)

    2022

  31. [39]

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

    Sun, J., Wang, X., Zhang, Y., Li, X., Zhang, Q., Liu, Y., Wang, J.: Fenerf: Face editing in neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7672–7682 (2022)

    2022

  32. [40]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

    Tucker, R., Snavely, N.: Single-view view synthesis with multiplane images. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 551–560 (2020)

    2020

  33. [41]

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

    Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752–5761 (2021)

    2021

  34. [42]

    International journal of computer vision129, 3051–3068 (2021) SemFaceEdit: Semantic Face Editing on Generative Radiance Manifolds 17

    Yu, C., Gao, C., Wang, J., Yu, G., Shen, C., Sang, N.: Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International journal of computer vision129, 3051–3068 (2021) SemFaceEdit: Semantic Face Editing on Generative Radiance Manifolds 17

    2021

  35. [43]

    In: Proceedings of the European conference on computer vision (ECCV)

    Yu, C., Wang, J., Peng, C., Gao, C., Yu, G., Sang, N.: Bisenet: Bilateral segmenta- tion network for real-time semantic segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 325–341 (2018)

    2018

  36. [44]

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

    Zheng, Y., Yang, H., Zhang, T., Bao, J., Chen, D., Huang, Y., Yuan, L., Chen, D., Zeng, M., Wen, F.: General facial representation learning in a visual-linguistic manner. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18697–18709 (2022)

    2022

  37. [45]

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

    Zheng, Y., Abrevaya, V.F., Bühler, M.C., Chen, X., Black, M.J., Hilliges, O.: Im avatar: Implicit morphable head avatars from videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13545– 13555 (2022)

    2022

  38. [46]

    arXiv preprint arXiv:2110.09788 (2021)

    Zhou, P., Xie, L., Ni, B., Tian, Q.: Cips-3d: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. arXiv preprint arXiv:2110.09788 (2021)

    2021 arXiv

  39. [47]

    Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnification: Learn- ingviewsynthesisusingmultiplaneimages.arXivpreprintarXiv:1805.09817(2018)

    2018 arXiv

  40. [48]

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

    Zhu, P., Abdal, R., Qin, Y., Wonka, P.: Sean: Image synthesis with semantic region- adaptive normalization. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 5104–5113 (2020)

    2020

  41. [49]

    In: European Conference on Computer Vision

    Zhuang, Y., Zhu, H., Sun, X., Cao, X.: Mofanerf: Morphable facial neural radiance field. In: European Conference on Computer Vision. pp. 268–285. Springer (2022) 18 S. Verma et al. Appendix 1 Loss Functions As discussed in the main paper, we train our network in two stages in ...

    2022

  42. [50]

    Note that since our approach generates semantic and rgb-radiances in a vol- ume, by design it supports generation of multi-view images by changing camera SemFaceEdit: Semantic Face Editing on Generative Radiance Manifolds 21 Fig. 11. Artifacts due to discrepancy in semantic ra...

  43. [51]

    Each variant is trained for 60K iterations (2 epoch) jointly learning weights for both Geometry Module and Appearance Module (First stage)

    Dataset images with different weight sharing in Appearance Module. Each variant is trained for 60K iterations (2 epoch) jointly learning weights for both Geometry Module and Appearance Module (First stage). Method FID ↓ KID (×103) ↓ No shared weights 30.36 41.62 Fully Shared A...

  44. [52]

    Each variant is trained for 60K iterations (2 epochs), jointly learning weights for both Geometry Module and Appearance Module (First stage)

    Dataset images with different architecture depth of Geometry Module and Appear- ance Module. Each variant is trained for 60K iterations (2 epochs), jointly learning weights for both Geometry Module and Appearance Module (First stage). Depth Geometric Module Appearance Module F...

Pith tools

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