The Angel is in the Priors: Improving GAN based Image and Sequence Inpainting with Better Noise and Structural Priors

Arnav Kumar Jain; Avisek Lahiri; Prabir Kumar Biswas

REVIEW 4 major objections 4 minor 31 references

The Angel is in the Priors: Improving GAN based Image and Sequence Inpainting with Better Noise and Structural Priors

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Unsupervised GAN inpainting can run as a single feed-forward pass by learning to predict the noise prior, giving roughly 1500x speedup and stronger temporal consistency.

desk verdict The speedup story is real and the sequence part is genuinely new, but the quality gain over Yeh et al. is confounded by changing the generator and loss at the same time as the inference procedure. read the letter →

arxiv 1908.05861 v1 pith:OTHGOFUV submitted 2019-08-16 cs.CV

Avisek Lahiri , Arnav Kumar Jain , Prabir Kumar Biswas This is my paper · ORCID

classification cs.CV

keywords GANinpaintingunsupervisedimagecompletionnoisepriorpredictionstructuralpriorsfaciallandmarkssequencerecurrentgroupedtemporalconsistency

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

The paper aims to show that the slow, iterative search for a latent noise vector in unsupervised GAN inpainting can be replaced by a learned predictor. A small network is trained, with the generator and discriminator of a pretrained GAN frozen, to map a masked image directly to a matching noise prior. At test time this yields the inpainted image in one feed-forward pass, about 1500 times faster than iterative latent optimization, with usually better reconstruction fidelity. The same predictor is extended to video with a recurrent grouped-prior network and a subsequence consistency loss, giving what the authors describe as the first unsupervised GAN-based sequence inpainting with improved temporal consistency.

What carries the argument

The central object is a learned noise-prior predictor $P_{\theta_z}$, a CNN that maps a masked image to the $d$-dimensional latent vector $z$ expected by a frozen pretrained GAN generator. It is trained with the composite loss $L_{\mathrm{com}}^z = \lambda_1 L_c + \lambda_2 L_r + \lambda_3 L_g$, where $L_c$ is the contextual (unmasked-region) fidelity, $L_r$ is the realism score from the frozen discriminator, and $L_g$ is a gradient-difference loss; this converts the iterative argmin of Eq. (2) into the direct evaluation of Eq. (7). For sequences, a CNN descriptor network feeds an LSTM that emits grouped priors for windows of $W$ frames, optimized by the grouped loss $L_{\mathrm{gr}}^z$ plus a subsequence consistency term $L_{\mathrm{ss}}^z$ that penalizes pairwise latent differences within a window. Structural priors enter as conditional variables $c$, here detected facial landmarks, so the generator learns to respect pose and expression while filling the mask.

What would settle it

Train the noise-prior predictor on random central masks only, then evaluate on freehand and checkerboard masks; if the reconstruction PSNR drops to the iterative baseline's level, the learned prior has not generalized and the 'one feed-forward pass' claim is conditional on matching mask statistics.

Watch

Extended reading notes

Core claim

The central claim is that priors are the deciding factor in GAN inpainting: a data-driven parametric network $P_{\theta_z}$ can predict the latent code a frozen generator needs, so the per-image iterative optimization used by the unsupervised baseline becomes unnecessary. The predictor is trained only on unmasked pixels, using a contextual loss, a realism loss from the frozen discriminator, and a gradient-difference loss; no pixel under the mask is ever supervised. Adding a structural prior, exemplified by facial landmarks used as a conditioning variable $c$, biases completion toward samples that preserve the pose and expression of the reference, and also improves the generator's samples. For sequences, an LSTM-CNN model predicts grouped noise priors over windows of frames, and a subsequence consistency loss penalizes appearance drift between adjacent reconstructions. Across SVHN, Stanford Cars, CelebA, CelebA-HQ, synthetic sequences, and VidTIMIT, the paper reports consistent gains over the unsupervised baseline in reconstruction quality and temporal consistency, with inference speedups around 1500x and performance comparable or better than hybrid supervised+adversarial methods.

Load-bearing premise

The method assumes the generative model used to synthesize images already covers the kind of images being inpainted; the authors concede this fails for broad natural scenes like ImageNet, so the result is bounded by how good the generator is.

Editorial extensions

If this is right

Unsupervised GAN inpainting becomes a single feed-forward network evaluation at test time, replacing roughly 1500 gradient-descent iterations per image with one forward pass and a small overhead for structural-prior extraction.
Because the masked pixels are never used in training, the method stays fully unsupervised; better future generators should slot into the same pipeline without retraining the inpainter.
The grouped LSTM prior and subsequence consistency loss improve temporal coherence of reconstructed videos, as measured by higher synthetic-sequence consistency and lower MOVIE scores than the iterative baseline.
Conditioning on structural priors gives the generator a degree of control over pose and expression, so completion respects the reference identity and expression rather than sampling an arbitrary plausible face.
Using progressive GAN training lets the same method operate at 256x256 resolution, where the iterative latent-optimization baseline degrades sharply.

Reading between the lines

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

If the learned-prior trick transfers, other inverse problems that currently rely on GAN latent optimization—such as super-resolution, denoising, or style transfer under masks—could adopt the same single-pass formulation whenever a suitable pretrained generator exists.
The speed and quality gains likely depend on the mask distribution seen during training; a predictor trained on random central masks may not generalize to freehand or checkerboard masks without fine-tuning, so the reported speedup should be read as tied to the experiment's mask protocol.
The failure boundary the authors concede for ImageNet-like scenes suggests the method could serve as a diagnostic for GAN coverage: the gap between single-pass and iterative reconstructions measures how well the generator's latent space covers the data.
Measured end-to-end, including structural-prior extraction and final mask composition, the 1500x speedup may shrink because those constant overheads are excluded from the headline ratio.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper's real contribution is amortizing the iterative latent search of Yeh et al. into a learned feed-forward predictor, plus an LSTM-based grouped prior for unsupervised sequence inpainting. The sequence part is the newest: as far as I can tell this is the first unsupervised GAN method that inpaints videos with temporal consistency, and the subsequence consistency loss is a simple but sensible addition. The structural prior (facial landmarks) is also a nice touch, and the synthetic-sequence consistency metric is a reasonable way to evaluate temporal stability. I believe the authors here; the speedup is real in the sense that 1500 optimization iterations are replaced by one network pass, even if the actual wall-clock gain is not measured and the "1500x" is an iteration-count comparison.

The soft spot is the quality claim. The main tables show large PSNR gains over Yeh et al., but the comparison is not controlled. The proposed model uses a different generator (progressive GAN at 256x256 versus whatever Yeh used), adds a gradient-difference loss, and then swaps iterative optimization for a learned predictor. You cannot attribute the 15.7-to-20.1 dB jump on CelebA-HQ to the learned prior alone. The stress-test note has this right. A fixed-G, fixed-loss ablation would be needed to support the "simultaneous improvement in reconstruction quality" claim. The paper also has the usual reproducibility gaps: no code, no hyperparameters, no error bars, and at least one or two equations are written sloppily (Eq. 3 and Eq. 5 look malformed). Table 1 has an internal inconsistency: M5 improves over M4 on RC but regresses on RF and RCh, which undercuts the claim that the subsequence consistency loss helps across masks. These are addressable.

The authors do concede in Section 6 that current GANs cannot handle natural images like ImageNet, so the method is bounded by GAN coverage. That is honest. The paper shows clear thinking and the literature coverage looks adequate, though the "first sequence inpainting" claim deserves a prior-art check before publication.

Who is this for? People working on GAN-based inpainting, especially anyone who wants to avoid iterative test-time optimization. It deserves a serious referee: the core idea is useful, the sequence extension is novel, and the flaws are fixable with better evaluation rather than being fatal. I would send it out, but with a strong request for controlled ablations and reproducibility details.

Referee Report

4 major / 4 minor

Summary. The paper proposes to accelerate unsupervised GAN-based inpainting by replacing the iterative latent-code optimization of Yeh et al. [29] with a learned feed-forward network P_theta_z that predicts a noise prior z from a masked image. Training of P_theta_z uses a contextual loss, a realism loss from a frozen discriminator, and a gradient-difference loss, with optional structural priors (facial landmarks) and an LSTM-based grouped prior for sequences. The authors claim a roughly 1500x speedup over [29] with improved reconstruction quality, and present the first unsupervised GAN-based sequence inpainting method, adding a subsequence consistency loss for temporal smoothness. Experiments on SVHN, Cars, CelebA, CelebA-HQ, synthetic sequences, and VidTIMIT compare against [29] and hybrid baselines using PSNR, a proposed temporal-consistency metric, and the MOVIE video-quality metric.

Significance. If the central claims hold, the paper makes a useful contribution: it converts an iterative, slow unsupervised inpainting procedure into a one-shot feed-forward inference, and it extends unsupervised GAN inpainting to sequences. The idea of learning a noise-prior predictor is simple and potentially impactful, and the paper is honest about the current limitation that fully unsupervised GAN inpainting does not scale to ImageNet/Places2 (Section 6). The synthetic-sequence temporal-consistency evaluation and the use of the MOVIE metric are sensible evaluation choices. However, the empirical evidence as presented is not yet convincing because the main comparison against [29] changes multiple factors at once, the reported gains for the subsequence consistency loss are internally inconsistent, the temporal-consistency metric is defined incorrectly in the text, and crucial implementation details are missing. The core concept is defensible, but the load-bearing experimental support needs substantial revision.

major comments (4)

[Section 4.1, Eq. (6), Table 1] The central claim of simultaneous quality improvement over [29] is not supported by a controlled comparison. The proposed M1 differs from the Yeh et al. baseline in at least three coupled ways: (1) a different pretrained generator, which Section 5.1.2 argues is more realistic and Section 5.1.3 trains with progressive growing at 256x256; (2) the additional gradient-difference loss in Eq. (5), which is absent from the baseline's objective in Eq. (2); and (3) the learned feed-forward predictor itself. Consequently, the large PSNR gains such as 15.7 dB to 20.1 dB on CelebA-HQ in Table 1 cannot be attributed solely to replacing iterative optimization with P_theta_z. No experiment holds G, D, and the loss function fixed while varying only the inference procedure. This is the key ablation needed to support the paper's main contribution.
[Table 1, Section 5.2.2] The claimed benefit of the subsequence consistency loss is contradicted by the reported numbers. In Table 1 on CelebA@128, M5 (with consistency loss) improves over M4 (without it) on RC (27.6 vs. 26.3) but regresses on RF (28.0 vs. 29.8) and RCh (26.9 vs. 29.4). The paper in Section 5.2.2 asserts that L_ss improves temporal consistency, but the data show mixed results, and M5 is only reported for one dataset and mask set. Without error bars or a clear explanation for the regression, this claim is not established.
[Eq. (12), Section 5.2.1, Table 1 caption] The temporal consistency metric is defined inconsistently. The text states 'Ideally, we want eta_temp = 0,' but Eq. (12) defines eta_temp as the mean pairwise PSNR, for which higher values mean more consistent (identical frames give infinite PSNR). The Table 1 caption says 'Higher values of consistency are better,' which contradicts the 'eta_temp = 0' statement. This needs to be corrected and the metric re-stated unambiguously, since it is a primary evaluation tool for the sequence contributions.
[Section 4, Equations (3)-(6), Section 5, Tables 1-2] The experimental sections lack the details needed to reproduce or verify the central speed and quality claims. The exact architectures of P_theta_z, the LSTM-CNN module, and the pretrained GANs are not specified; the loss weights lambda_1 through lambda_4, the group size W, and the noise dimension d are not given; the gradient-difference term in Eq. (5) contains an incomplete expression 'G()' with no argument; and no wall-clock timings or error bars are reported for the 1500x speedup claim or for any PSNR/MOVIE numbers. The paper would be substantially strengthened by releasing code, specifying all hyperparameters, and reporting standard deviations over multiple runs.

minor comments (4)

[Abstract and Section 1] There are several typos and grammatical slips, including 'accelarate' (Abstract), 'simultaneousness' (Section 1), and 'Standford' for Stanford (used throughout, e.g., Section 5.1). These should be corrected.
[Eq. (8)] The conditional GAN objective in Eq. (8) writes the second expectation as E_z[1 - D(G(z))], but when conditioning on c, the discriminator should also receive c in the fake term, i.e., E_{z,c}[1 - D(G(z), c)] or an equivalent formulation. Please clarify.
[Section 5.1.2] The human evaluation is described as 'the classiﬁcation accuracy is 5.8% higher for DIP(p = 10^-3)' but it is unclear what 'DIP' refers to here and what statistical test yields the reported p-values. Please specify the hypothesis and the test used.
[Figure 3 and Section 4.3] There is a stray word 'ames' in the paragraph before Figure 3, and the figure caption states 'Unlock sign means parameters to update' but the figure itself is not visible in the text. Please ensure the figure and caption are consistent.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the learned noise-prior predictor is an amortized version of the iterative objective, and all reported gains are checked against external baselines and held-out test data.

full rationale

The paper's central step is training P_theta_z (Eq. 6) to minimize essentially the same masked-region loss that Yeh et al. optimize per test image (Eq. 2). This is a standard amortized-inference construction, not a tautology: the learned predictor is not defined in terms of the reported outputs, and its quality is measured against an external iterative baseline on held-out images. The claimed 1500x speedup is arithmetic (1500 test-time iterations versus one forward pass), not a derived prediction. The subsequence consistency loss (Eq. 10) penalizes distances in latent space, while the temporal-consistency metric (Eq. 12) is pairwise PSNR in image space; although related, these are not the same quantity, so the observed consistency improvement is an empirical outcome rather than a forced identity. No load-bearing self-citation appears, and the comparisons with hybrid methods and with [29] are external. The main validity concern is that the quality comparison to [29] changes generator, loss, and inference procedure simultaneously, but that is an experimental confound, not a circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The ledger is dominated by undisclosed experimental choices and domain assumptions. There are no invented physical entities. The free parameters are the loss weights, the LSTM window size, and the latent dimension, none of which are reported. The central assumptions are the reliability of the frozen GAN as a supervision signal, the availability of landmarks under masks, the validity of L2 distance in latent space as a perceptual proxy, and the transfer from synthetic to real video. These assumptions are load-bearing for the sequence claims in particular.

free parameters (4)

lambda_1, lambda_2, lambda_3 (loss weights in Eq. 6) = not reported
Relative weights of contextual, realism, and gradient difference losses are hand-chosen and undisclosed; they determine the balance between visible-pixel fidelity and manifold realism in every experiment.
lambda_4 (subsequence consistency weight, Eq. 11) = not reported
Controls how strongly the latent codes of nearby frames are pulled together; no sensitivity study is provided, so the M5 gains cannot be separated from the choice of this weight.
W (LSTM group size) = not reported
The number of frames grouped by the recurrent predictor is a free design choice; the paper never states W, even though the subsequence loss is defined over a W-frame neighborhood.
noise dimension d = not reported
The dimensionality of the GAN latent space is not specified; it bounds the capacity of the noise-prior predictor and affects the LSTM consistency loss.

assumptions (4)

domain assumption The frozen pretrained GAN generator and discriminator supply a differentiable objective landscape whose optima correspond to plausible inpaintings.
Section 4.1 freezes G and D and trains P_theta_z using Lc, Lr, and Lg; if the GAN manifold has gaps or the realism loss is uninformative, the learned predictor inherits the failure, which is why Section 6 limits the method to narrow domains.
domain assumption Facial landmarks can be extracted from masked or corrupted images and serve as valid structural priors.
Section 4.2 conditions the generator on landmarks from Kazemi et al.; masks that occlude landmarks are not analyzed, so the availability of the prior at test time is assumed.
domain assumption Euclidean distance in latent space is a faithful proxy for perceptual or temporal dissimilarity of generated frames.
Section 4.4 defines the subsequence consistency loss as pairwise L2 distance between z vectors; this requires an approximately isometric latent space, which the paper does not establish.
domain assumption Synthetic sequences made by masking one image with different masks capture the temporal consistency needed in real video.
Section 5.2.1 defines eta_temp on synthetic sequences and states real-video alignment is not feasible; the transfer to real video is supported only by the MOVIE metric on VidTIMIT.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Angel is in the Priors: Improving GAN based Image and Sequence Inpainting with Better Noise and Structural Priors." pith.science (2026). https://pith.science/paper/OTHGOFUV

@misc{pith2026190805861,
  author       = {Pith},
  title        = {Pith review of: The Angel is in the Priors: Improving GAN based Image and Sequence Inpainting with Better Noise and Structural Priors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OTHGOFUV}},
  note         = {Machine review of arXiv:1908.05861}
}

read the original abstract

Contemporary deep learning based inpainting algorithms are mainly based on a hybrid dual stage training policy of supervised reconstruction loss followed by an unsupervised adversarial critic loss. However, there is a dearth of literature for a fully unsupervised GAN based inpainting framework. The primary aversion towards the latter genre is due to its prohibitively slow iterative optimization requirement during inference to find a matching noise prior for a masked image. In this paper, we show that priors matter in GAN: we learn a data driven parametric network to predict a matching prior for a given image. This converts an iterative paradigm to a single feed forward inference pipeline with a massive 1500X speedup and simultaneous improvement in reconstruction quality. We show that an additional structural prior imposed on GAN model results in higher fidelity outputs. To extend our model for sequence inpainting, we propose a recurrent net based grouped noise prior learning. To our knowledge, this is the first demonstration of an unsupervised GAN based sequence inpainting. A further improvement in sequence inpainting is achieved with an additional subsequence consistency loss. These contributions improve the spatio-temporal characteristics of reconstructed sequences. Extensive experiments conducted on SVHN, Standford Cars, CelebA and CelebA-HQ image datasets, synthetic sequences and ViDTIMIT video datasets reveal that we consistently improve upon previous unsupervised baseline and also achieve comparable performances(sometimes also better) to hybrid benchmarks.

Figures

Figures reproduced from arXiv: 1908.05861 by the authors.

**Figure 1.** Learning and inferencing(inpainting) with our learned noise prior model. Step 1: Learn a GAN model. Step 2: Freeze GAN modules and learn to infer noise prior based on masked input image. Step 3: During inference, given a masked image, predict a matching noise vector and use pretrained GAN generator(G) to yield final output. The dashed arrows show flow of error gradients during training phase. completion possibility … view at source ↗

**Figure 2.** Illustration of multi model image completion possibility of GAN based inpainting methods. Given a corrupted image, an unconditioned inpainting algorithms(top row) such as [29, 23, 11, 30] samples from a uniform distribution of viable inpainted images. However, if conditioned by structural priors(bottom row), the sampling distribution is biased towards samples which preserve original facial pose and expression. dis… view at source ↗

**Figure 3.** Grouped noise prior learning with a combined LSTM-CNN framework. Unlock sign means parameters to update. to more consistent reconstructions in terms of appearance. ames Since, RNNs suffer from vanishing gradients problem[3] and are unable to capture long dependencies, we use Long Short Term Memory (LSTM) [10] Networks [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (9 more)

**Figure 6.** Figure 6: Structural priors enables GAN to disentangle facial pose and appearance cues. Left: Faces samples with same z vector but different structural priors. Right: Faces sampled with different z vectors for a given structural prior. Even if some keypoints are missing/occlude…

**Figure 5.** Figure 5: Proposed structural priors in GAN help in generating better samples compared to vanilla GAN. Random samples from proposed framework are structurally more consistent and complete. This is eventually important for better inpainting. So, L gr z helps in learning the tem…

**Figure 8.** Figure 8: Benefit of structural prior augmented GAN based inpainting. In each sub figure, Column 1: Original image, Column 2: Masked image, Colimn 3: Inpainted by a GAN model conditioned on proposed learned noise prior Column 4:Inpainted by a GAN model conditioned on proposed …

**Figure 7.** Figure 7: Comparative visualization of inpainting. In each set, column 1: original image, column 2: masked images, column 3: unsupervised baseline of [29], column 4: Proposed learned noise prior conditioned model(Eq. 6). Proposed reconstructions are usually better, yet our model…

**Figure 9.** Figure 9: Visualizing consistency of inpainting synthetic sequences. A synthetic sequence is created by masking a given image with different corruption patterns. Ideally we want an inpainter to yield exactly same outputs for a given synthetic sequence.; Top: Masked synthetic seq…

**Figure 10.** Figure 10: Benefit of subsequence consistency loss(Eq.10) augmented with grouped prior loss(Eq.11). Left: A synthetic sequence in which same image(sample from CelebA-HQ @ 256×256) is masked differently. Ideally we want a model to inpaint both frames identically. Middle:Inpaintin…

**Figure 11.** Figure 11: Inpainting on VidTIMIT sequences by [29](3rd row) and proposed method: version M5(4th row). See Table1 for definition of M5 [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]

**Figure 12.** Figure 12: Visualization of inpainting by contemporary baselines(GLCIC[11], GIP[30], GFC [19]) trained without supervised reconstruction loss but only with unsupervised adversarial loss. It is evident that training with only unsupervised loss diminishes the efficacy of the metho…

**Figure 13.** Figure 13: Visual comparison of inpainting with hybrid methods of CE[23], GLCIC[11], GFC[19], GIP [30]. Our unsupervised method performs reasonably comparable even though it is totally unsupervised. Note, that at 256×256(CelebAHQ), performance of unsupervised baseline of [29] d…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 28 canonical work pages

[29]

R. A. Yeh, C. Chen, T. Y . Lim, A. G. Schwing, M. Hasegawa- Johnson, and M. N. Do. Semantic image inpainting with deep generative models. In CVPR, pages 5485–5493, 2017. 1, 2, 3, 5, 6, 7, 8, 9

work page 2017
[1]

Ballester, M

C. Ballester, M. Bertalmio, V . Caselles, G. Sapiro, and J. Verdera. Filling-in by joint interpolation of vector ﬁelds and gray levels. IEEE transactions on image processing , 10(8):1200–1211, 2001. 2

2001
[2]

Barnes, E

C. Barnes, E. Shechtman, A. Finkelstein, and D. B. Gold- man. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Transactions on Graphics (ToG), 28(3):24, 2009. 1

work page 2009
[3]

Bengio, P

Y . Bengio, P. Simard, and P. Frasconi. Learning long-term dependencies with gradient descent is difﬁcult. IEEE trans- actions on neural networks, 5(2):157–166, 1994. 4 Figure 13: Visual comparison of inpainting with hybrid methods of CE[23], GLCIC[11], GFC[19], GIP [30]. Our unsupervised method performs reasonably comparable even though it is totally uns...

work page 1994
[4]

Bertalmio, G

M. Bertalmio, G. Sapiro, V . Caselles, and C. Ballester. Image inpainting. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques , pages 417–

work page
[5]

Caballero, C

J. Caballero, C. Ledig, A. Aitken, A. Acosta, J. Totz, Z. Wang, and W. Shi. Real-time video super-resolution with spatio-temporal networks and motion compensation. CVPR,

work page
[6]

A. A. Efros and W. T. Freeman. Image quilting for tex- ture synthesis and transfer. In Proceedings of the 28th an- nual conference on Computer graphics and interactive tech- niques, pages 341–346. ACM, 2001. 2

work page 2001
[7]

A. A. Efros and T. K. Leung. Texture synthesis by non- parametric sampling. In ICCV, page 1033. IEEE, 1999. 2

work page 1999

Show all 31 references

[8]

Goodfellow, J

I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio. Gen- erative adversarial nets. In NIPS, pages 2672–2680, 2014. 1, 3

2014
[9]

Hays and A

J. Hays and A. A. Efros. Scene completion using millions of photographs. In ACM Transactions on Graphics (TOG), volume 26, page 4. ACM, 2007. 1

2007
[10]

Hochreiter and J

S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997. 4

1997
[11]

Iizuka, E

S. Iizuka, E. Simo-Serra, and H. Ishikawa. Globally and locally consistent image completion. ACM Transactions on Graphics (TOG), 36(4):107, 2017. 1, 2, 3, 6, 7, 8, 9

2017
[12]

Karras, T

T. Karras, T. Aila, S. Laine, and J. Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In ICLR, 2018. 5, 6, 9

2018
[13]

Kazemi and J

V . Kazemi and J. Sullivan. One millisecond face alignment with an ensemble of regression trees. In CVPR, pages 1867– 1874, 2014. 4, 5

2014
[14]

D. P. Kingma and M. Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1

2013 arXiv
[15]

K ¨ohler, C

R. K ¨ohler, C. Schuler, B. Sch ¨olkopf, and S. Harmeling. Mask-speciﬁc inpainting with deep neural networks. In Ger- man Conference on Pattern Recognition , pages 523–534. Springer, 2014. 2

2014
[16]

Krause, M

J. Krause, M. Stark, J. Deng, and L. Fei-Fei. 3d object rep- resentations for ﬁne-grained categorization. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 554–561, 2013. 5

2013
[17]

Ledig, L

C. Ledig, L. Theis, F. Husz ´ar, J. Caballero, A. Cunningham, A. Acosta, A. P. Aitken, A. Tejani, J. Totz, Z. Wang, et al. Photo-realistic single image super-resolution using a genera- tive adversarial network. In CVPR, volume 2, page 4, 2017. 7

2017
[18]

C. Li, Y . Ding, B. Yu, M. Xu, and Q. Zhang. Inpaint- ing of continuous frames of old movies based on deep neu- ral network. In 2018 International Conference on Audio, Language and Image Processing (ICALIP), pages 132–137. IEEE, 2018. 6

2018
[19]

Y . Li, S. Liu, J. Yang, and M.-H. Yang. Generative face completion. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, page 3, 2017. 6, 7, 8, 9

2017
[20]

Z. Liu, P. Luo, X. Wang, and X. Tang. Deep learning face at- tributes in the wild. InProceedings of the IEEE International Conference on Computer Vision, pages 3730–3738, 2015. 5

2015
[21]

Mathieu, C

M. Mathieu, C. Couprie, and Y . LeCun. Deep multi-scale video prediction beyond mean square error. ICLR, 2016. 3

2016
[22]

D. Nie, R. Trullo, J. Lian, C. Petitjean, S. Ruan, Q. Wang, and D. Shen. Medical image synthesis with context-aware generative adversarial networks. In MICCAI, pages 417–
[23]

Pathak, P

D. Pathak, P. Krahenbuhl, J. Donahue, T. Darrell, and A. A. Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536–2544, 2016. 1, 2, 3, 6, 7, 8, 9

2016
[24]

Sanderson and B

C. Sanderson and B. C. Lovell. Multi-region probabilis- tic histograms for robust and scalable identity inference. In International Conference on Biometrics, pages 199–208. Springer, 2009. 7

2009
[25]

Seshadrinathan and A

K. Seshadrinathan and A. C. Bovik. Motion tuned spatio- temporal quality assessment of natural videos. IEEE trans- actions on image processing, 19(2):335–350, 2010. 7, 8

2010
[26]

Shrivastava, T

A. Shrivastava, T. Pﬁster, O. Tuzel, J. Susskind, W. Wang, and R. Webb. Learning from simulated and unsupervised im- ages through adversarial training. CVPR, pages 2107–2116,
[27]

X. Sun, R. Szeto, and J. J. Corso. A temporally-aware inter- polation network for video frame inpainting. arXiv preprint arXiv:1803.07218, 2018. 6

2018 arXiv
[28]

L. Xu, J. S. Ren, C. Liu, and J. Jia. Deep convolutional neural network for image deconvolution. In Advances in Neural Information Processing Systems, pages 1790–1798, 2014. 2

2014
[30]

J. Yu, Z. Lin, J. Yang, X. Shen, X. Lu, and T. S. Huang. Generative image inpainting with contextual attention. In CVPR, 2018. 2, 3, 6, 7, 8, 9

2018
[424]

ACM Press/Addison-Wesley Publishing Co., 2000. 2

2000

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [29]

R. A. Yeh, C. Chen, T. Y . Lim, A. G. Schwing, M. Hasegawa- Johnson, and M. N. Do. Semantic image inpainting with deep generative models. In CVPR, pages 5485–5493, 2017. 1, 2, 3, 5, 6, 7, 8, 9

work page 2017

[2] [1]

Ballester, M

C. Ballester, M. Bertalmio, V . Caselles, G. Sapiro, and J. Verdera. Filling-in by joint interpolation of vector ﬁelds and gray levels. IEEE transactions on image processing , 10(8):1200–1211, 2001. 2

2001

[3] [2]

Barnes, E

C. Barnes, E. Shechtman, A. Finkelstein, and D. B. Gold- man. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Transactions on Graphics (ToG), 28(3):24, 2009. 1

work page 2009

[4] [3]

Bengio, P

Y . Bengio, P. Simard, and P. Frasconi. Learning long-term dependencies with gradient descent is difﬁcult. IEEE trans- actions on neural networks, 5(2):157–166, 1994. 4 Figure 13: Visual comparison of inpainting with hybrid methods of CE[23], GLCIC[11], GFC[19], GIP [30]. Our unsupervised method performs reasonably comparable even though it is totally uns...

work page 1994

[5] [4]

Bertalmio, G

M. Bertalmio, G. Sapiro, V . Caselles, and C. Ballester. Image inpainting. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques , pages 417–

work page

[6] [5]

Caballero, C

J. Caballero, C. Ledig, A. Aitken, A. Acosta, J. Totz, Z. Wang, and W. Shi. Real-time video super-resolution with spatio-temporal networks and motion compensation. CVPR,

work page

[7] [6]

A. A. Efros and W. T. Freeman. Image quilting for tex- ture synthesis and transfer. In Proceedings of the 28th an- nual conference on Computer graphics and interactive tech- niques, pages 341–346. ACM, 2001. 2

work page 2001

[8] [7]

A. A. Efros and T. K. Leung. Texture synthesis by non- parametric sampling. In ICCV, page 1033. IEEE, 1999. 2

work page 1999