Pith. sign in

REVIEW 3 major objections 6 minor 45 references

GuidPaint: Class-Guided Image Inpainting with Diffusion Models

T0 review · 3 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 GuidPaint claims that adding classifier guidance to the denoising loop of training-free diffusion inpainting improves semantic consistency and visual plausibility over RePaint and CoPaint-TT, reporting lower LPIPS and higher SSIM on…

desk verdict Useful training-free inpainting recipe, but the evaluation hands GuidPaint ground-truth class labels the baselines don't get, so the headline gains don't isolate the method. read the letter →

arxiv 2507.21627 v1 pith:5A5SKOQA submitted 2025-07-29 cs.CV

Qimin Wang , Xinda Liu , Guohua Geng This is my paper
classification cs.CV
keywords ImageInpaintingDiffusionModelsClassifierGuidanceTraining-FreeSamplingStrategyZero-ShotLearningHybridNon-UniformSkip
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

GuidPaint is an image-inpainting method that adds classifier guidance to a training-free diffusion inpaint pipeline. The paper claims that by steering the intermediate denoising steps toward a target class label, the masked region becomes semantically consistent with the rest of the image, and that combining stochastic sampling with a deterministic refinement pass fixes the visually implausible layouts that context-aware baselines sometimes produce. On ImageNet and CelebA-HQ it reports lower LPIPS (perceptual distance) and higher SSIM/PSNR than RePaint and CoPaint-TT on large masks, while a skip-sampling variant cuts inference time relative to both baselines. The contribution is that controllability is achieved without retraining or architecture changes: only a pretrained classifier and the usual diffusion checkpoint are needed. If correct, this makes class-guided inpainting practical as a zero-training plugin for existing diffusion models.

What carries the argument

Two gradient-based corrections run on the same noisy latent $x_t$: the inpainting loss $L_{\mathrm{inp}}$ from CoPaint, which pushes the reconstruction $\hat{x}_0$ to match the known pixels under the mask, and a classifier-guidance loss $L_{\mathrm{guid}}=-\log p_\phi(y\mid \hat{x}_0)$, which pushes the reconstructed image toward the target class. The guidance scale multiplies the classifier gradient by the learnable variance $\Sigma$ of Improved DDPM. Around this core, stochastic sampling forms an intermediate candidate set by compositing the known region with the generated region and re-noising to the current timestep (stopping at $t_{\mathrm{stop}}^{\mathrm{comp}}=130$), and a non-uniform skip schedule allocates more steps to the texture-refinement stage. The hybrid loop is what lets stochastic exploration and deterministic refinement coexist inside a training-free method.

What would settle it

Give the two comparison methods the same class label that GuidPaint receives (for instance, by adding the same classifier-guidance term to their sampling loops) and re-tune every method's hyperparameters on the full validation set rather than the first five images; if the reported LPIPS/SSIM gaps shrink, disappear, or reverse, the claimed improvement over context-aware inpainting is not established.

Watch

Extended reading notes

Core claim

GuidPaint's central claim is that classifier guidance—gradient steps taken on the noisy image estimate to maximize a pretrained classifier's probability for a chosen class—can be inserted into the denoising loop of context-aware inpainting and yields better semantic consistency and visual plausibility than purely context-driven methods. The method applies this guidance either globally to the whole image or locally inside each masked region, and it reruns the known-region consistency constraint of CoPaint on the guided latent. It then exposes a 'Stochastic Intermediate Generation Space' during sampling: stochastic runs produce several plausible intermediate states, the user picks one, and deterministic DDIM refinement turns it into the final image. Quantitative tables report LPIPS 0.340 vs 0.427 and SSIM 0.532 vs 0.460 against CoPaint-TT on ImageNet Expand masks, with comparable gains on CelebA-HQ; the ablation attributes most of the quality gain to the combination of classifier guidance and stochastic sampling.

Load-bearing premise

The comparison assumes that giving GuidPaint the correct object label (read off from the un-masked part of the image) while giving the comparison methods no label at all is a fair test, and that settings chosen on the first five validation images hold up on the hundred-image test set.

Editorial extensions

If this is right

  • Any existing context-aware diffusion inpainting pipeline can be upgraded with class control by adding a pretrained classifier and gradient steps, without retraining.
  • Users can choose what object class appears in the hole (e.g. 'goldfish' instead of 'goldfinch'), which is direct class-guided editing rather than unlabelled completion.
  • The interactive workflow—browsing stochastic intermediate results, choosing one, then deterministically refining—gives human-in-the-loop control over the final inpaint.
  • The non-uniform skip schedule reduces inference time (138s vs 330s for CoPaint-TT and 183s for RePaint in the reported ImageNet setting) while retaining quality.
  • On large masks (Expand, Half) the method reports better perceptual and structural similarity than both baselines, and on Square masks it is comparable or slightly better.

Reading between the lines

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

  • A natural next experiment is to give RePaint and CoPaint-TT the same class label as GuidPaint (e.g., by adding the same classifier-guidance term to their sampling loops); if the gap persists, the advantage is in the hybrid sampling rather than in having extra information.
  • Because the classifier defines the vocabulary, the method's control is bounded by classifier categories; replacing the classifier loss with an open-vocabulary text-conditioned loss is the obvious extension the paper does not pursue.
  • The user-selection step suggests a perceptual user study: do human-selected intermediate states lead to higher downstream fidelity than automatically selecting the first candidate, and does the added interaction time outweigh the quality gain?
  • Hyperparameters (guidance scale $s=1.0$, $I_{\mathrm{inp}}=I_{\mathrm{guid}}=2$, composition-stop threshold) were tuned on five images; testing the same settings across diverse datasets and mask geometries would show whether the guide is robust or needs per-dataset tuning.
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

3 major / 6 minor

Summary. The paper proposes GuidPaint, a training-free class-guided image inpainting method that incorporates classifier guidance into the intermediate denoising steps of a diffusion model. It combines global and local classifier guidance with stochastic and deterministic sampling stages, and introduces a non-uniform skip-sampling schedule for efficiency. The method is evaluated on ImageNet and CelebA-HQ against RePaint and CoPaint-TT, with reported improvements on ImageNet for large masks. The paper also demonstrates interactive selection of intermediate generations and reports ablation results. The central empirical claim is that GuidPaint achieves clear quantitative and qualitative improvements over existing context-aware inpainting methods while remaining training-free.

Significance. If the reported improvements were established under a fair protocol, the contribution would be meaningful: a simple, training-free way to inject semantic class control into inpainting, with an interactive sampling strategy that is easy to adopt. The paper also makes its code available and presents the method transparently. However, the quantitative evaluation is currently not convincing because the comparison is not equivalent: GuidPaint receives a class label derived from the input image via a pretrained classifier, whereas the baselines receive no class information. The CelebA-HQ results, where classifier guidance is disabled, show no consistent gain over CoPaint-TT, which is consistent with the concern that the ImageNet gains largely reflect the added conditioning signal rather than the proposed algorithm. The paper's value is therefore conditional on a re-evaluation that controls for this asymmetry.

major comments (3)
  1. [§4.1, Fig. 4, Table 1] The quantitative comparison is not equivalent. GuidPaint obtains the target class label by running a pretrained classifier on the ground-truth image (or the masked ground truth) and applies classifier guidance throughout denoising, while RePaint and CoPaint-TT are evaluated without any class label or guidance signal. Therefore the reported ImageNet gains (e.g., Expand LPIPS 0.340 vs 0.427; SSIM 0.532 vs 0.460) do not isolate the value of GuidPaint's hybrid sampling or its particular guidance formulation; they also reflect the effect of giving the generative model the correct ImageNet class. This concern is supported by the CelebA-HQ rows, where GuidPaint* uses no classifier guidance and is not consistently better than CoPaint-TT (e.g., Expand LPIPS 0.367 vs 0.362; Square LPIPS 0.096 vs 0.090). To establish the claimed improvement, the authors should compare against class-conditioned versions of the baselines, e.g., by applying the same classifier-gradient guidance to RePaint/CoPaint-TT, or by ablating GuidPaint with the class label withheld. Without such a control, the central claim of 'clear improvements' is not established.
  2. [§4.1, Table 1] Hyperparameters are selected on only the first 5 validation images, and the reported test results are averaged over 100 images without error bars, confidence intervals, or significance tests. Given that several differences are small (e.g., Square-mask SSIM 0.805 vs 0.802; CelebA-HQ Half LPIPS 0.201 vs 0.203), the reader cannot tell whether the reported gains are robust or within run-to-run variation. The authors should provide per-image results or at least standard errors, and ideally report how the choice of validation images affects the conclusions.
  3. [§4.2, Table 2] The efficiency claim is not clearly supported by the table. Full GuidPaint takes 381s on ImageNet, which is slower than CoPaint-TT (330s), while only GuidPaint-Skip is faster (138s). The abstract and Section 4.2 state that non-uniform skip sampling accelerates inference, which is true for the skip variant, but the full method does not. Please clarify the scope of the efficiency claim and report the quality/time trade-off for both variants.
minor comments (6)
  1. [Throughout] There are several typos and grammatical issues, e.g., 'plausiblility' in §3, 'unresonable' in §4.2, 'Metics' in §4.1, and 'these methods struggles' in §1. A careful proofread is needed.
  2. [Fig. 1 caption] The caption uses 'w/o y' and 'y=18' but does not define y or the class-index convention in the caption; please define these in the text or figure.
  3. [§3.2] The paper mentions predicting Top-k class labels but does not specify how the target class y is selected among the top-k predictions, nor how this interacts with the cross-entropy loss in Eq. (8). Please clarify.
  4. [§3.2, Eq. (8)] The notation Σ in Eq. (8) is used for the learned variance, but the explanation in Eq. (9) is brief and the relationship between υ and the predicted variance could be stated more explicitly for reproducibility.
  5. [§2, §4.1] RAD is mentioned in the related work as a context-aware diffusion inpainting method but is not included in the quantitative comparison. Since the paper claims improvement over 'existing context-aware inpainting methods,' a comparison with RAD or a justification for its omission would strengthen the evaluation.
  6. [§5] The stated limitations (class labels only, single-class multi-instance generation) are appropriate and should ideally be mentioned in the abstract or introduction to set expectations for the method's scope.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: GuidPaint combines externally cited components (classifier guidance from [8], inpainting constraint from CoPaint [42]); no self-citations and no derived quantity reduces to its own input by construction.

full rationale

The derivation chain is self-contained and non-circular. The class-conditional guidance in Eq. 8 (xguid_t = x_t - sSigma grad Lguid with Lguid = -log p_y(hat x0)) is exactly the classifier guidance introduced by Dhariwal and Nichol (reference [8]), an external, machine-checked result. The inpainting constraint in Eq. 6 (Linp = ||xgt_0⊙M - hat x0⊙M||^2 + lambda_reg||xt - mu_t||^2) is adapted from CoPaint (reference [42]), also external. The stochastic sampling of Eq. 10 (xcomp_0 = xgt_0⊙M + hat x0⊙(1-M), then forward-noised) is a new composition rule, but it uses the ground truth only in known regions, as all inpainting methods do, and does not define the output in terms of itself. No fitted parameter is renamed as a prediction; hyperparameters (Iguid, Iinp, tcomp_stop, guidance scale) are hand-chosen or tuned on a small validation set, which is standard practice and not circularity. There are no self-citations in the reference list, so no self-citation chain is load-bearing. The quantitative evaluation gives GuidPaint ground-truth class labels while RePaint and CoPaint-TT receive no class information, which may confound the reported gains, but this is an experimental fairness issue, not a circular derivation. The method's equations do not assume the result they are meant to produce, and the improvements are empirical rather than forced by construction.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The central method uses existing diffusion models, a pretrained classifier, and well-known inpainting constraints. The free parameters are mostly hyperparameters for gradient steps and thresholds, some tuned on a 5-image validation set. No new physical or conceptual entities are introduced. The key assumptions are the validity of classifier gradients on noisy intermediates and the distributional compatibility of the compositing step.

free parameters (7)
  • Iguid (guidance gradient steps) = 2
    Number of gradient descent steps for classifier guidance, set to 2 by default. Tuned on 5 validation images.
  • Iinp (inpainting gradient steps) = 2
    Number of gradient descent steps for inpainting constraint, set to 2 by default. Tuned on 5 validation images.
  • lambda_reg (regularization weight) = 0.01
    Weight for the regularization term in the inpainting loss (Eq. 6). Set by default.
  • eta_t (inpainting learning rate) = 0.02*sqrt(alpha_bar_t)*1.012^(T-t)
    Learning rate schedule for inpainting gradient descent (Eq. 7). The constants 0.02 and 1.012 are hand-chosen.
  • tcomp_stop = 130 (GuidPaint), 124 (GuidPaint-Skip)
    Early-stopping threshold for composition in stochastic sampling. Set based on validation.
  • guidance scale s = 1.0
    Scale for classifier guidance, consistent with Dhariwal and Nichol. Not tuned.
  • skip-step counts = {50,50,25,25,5}
    Number of sampling steps in each of the four stages for GuidPaint-Skip. The stage partition is not algorithmically specified.
assumptions (3)
  • standard math DDPM/DDIM reverse sampling equations hold (Eqs. 1-5).
    The method relies on the standard diffusion model framework; these are well-established equations from the cited literature.
  • domain assumption A pretrained classifier trained on clean images provides meaningful gradients when applied to intermediate denoising estimates xhat0.
    Classifier guidance assumes the classifier's gradient direction improves semantic alignment even though xhat0 is noisy and imperfect (Section 3.2, Eq. 8). This is an empirical assumption not proven in the paper.
  • domain assumption The composited image x_comp0 (known pixels plus generated pixels) follows the same noise distribution as natural images during forward noising.
    The stochastic sampling step (Eq. 10) re-noises the composite; if the generated region is out-of-distribution, the subsequent denoising may drift. The early-stopping threshold is a hand-coded mitigation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GuidPaint: Class-Guided Image Inpainting with Diffusion Models." pith.science (2026). https://pith.science/paper/5A5SKOQA

@misc{pith2026250721627,
  author       = {Pith},
  title        = {Pith review of: GuidPaint: Class-Guided Image Inpainting with Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5A5SKOQA}},
  note         = {Machine review of arXiv:2507.21627}
}
read the original abstract

In recent years, diffusion models have been widely adopted for image inpainting tasks due to their powerful generative capabilities, achieving impressive results. Existing multimodal inpainting methods based on diffusion models often require architectural modifications and retraining, resulting in high computational cost. In contrast, context-aware diffusion inpainting methods leverage the model's inherent priors to adjust intermediate denoising steps, enabling high-quality inpainting without additional training and significantly reducing computation. However, these methods lack fine-grained control over the masked regions, often leading to semantically inconsistent or visually implausible content. To address this issue, we propose GuidPaint, a training-free, class-guided image inpainting framework. By incorporating classifier guidance into the denoising process, GuidPaint enables precise control over intermediate generations within the masked areas, ensuring both semantic consistency and visual realism. Furthermore, it integrates stochastic and deterministic sampling, allowing users to select preferred intermediate results and deterministically refine them. Experimental results demonstrate that GuidPaint achieves clear improvements over existing context-aware inpainting methods in both qualitative and quantitative evaluations.

Figures

Figures reproduced from arXiv: 2507.21627 by the authors.

Figure 1
Figure 1. GuidPaint (Ours) enables class-guided control over masked regions (top), while stochastic sampling generates visually plausible content (bottom), compared to context-aware inpainting methods. Here, class label y denotes conditional diffusion model with classifier guidance (11: “goldfinch”, 1: “goldfish”), and w/o y indicates unconditional diffusion model without classifier guidance. Bottom: y=18 (“magpie”). (VAEs) [… view at source ↗
Figure 2
Figure 2. When T=250, CoPaint’s intermediate results show that flawed initial layouts lead to visually flawed outputs. For comparison, CoPaint-TT’s refined generation is presented. The introduction of Denoising Diffusion Probabilistic Models (DDPM) [11] has elevated Diffusion Models (DMs) [30] to state-of-the-art status in generative paradigms, demonstrating remarkable capability in producing high-fidelity and diverse images.… view at source ↗
Figure 3
Figure 3. GuidPaint overview. Input a ground truth image, masks and class labels. Skip sampling from random Gaussian noise, stochastic sampling generates diverse interme￾diate candidates, a selected candidate is then refined through deterministic sampling to produce the final output. Context-Aware Image Inpainting RePaint [20] introduces DDPM [11] into image inpainting by concatenating noisy versions of known regions with int… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Class label prediction using a pretrained classifier [8]. Lguid = − log pϕ(y|xˆ0) x guid t = xt − sΣ∇xtLguid (8) where s is the guidance scale (default=1.0, consistent with CG [8]), and Σ de￾notes the learnable reverse process variance proposed in Improved DDPM [23], p…
Figure 5
Figure 5. Figure 5: (a) Visualization of intermediate denoising results at different timesteps; (b) Stage partitioning of the denoising process using a linear noise schedule. While DDIM [31] enables skip sampling, uniform skip intervals are subopti￾mal across the entire denoising process.…
Figure 6
Figure 6. Figure 6: Qualitative results of ImageNet (a) and CelebA-HQ (b) [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results of Local Guidance on ImageNet (18: “magpie”, 130: “flamingo”) [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Ablation study examples on ImageNet with Half mask (CG: Classifier Guidance, SS: Stochastic Sampling) [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Ablation study examples on ImageNet with Skip Sampling [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 27 canonical work pages

  1. [1]

    In: The IEEE/CVF Conference on Computer Vision and Pattern Recog- nition 2023

    Anciukevicius, T., Xu, Z., Fisher, M., Henderson, P., Bilen, H., Mitra, N.J., Guer- rero, P.: Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and gen- eration. In: The IEEE/CVF Conference on Computer Vision and Pattern Recog- nition 2023. pp. 12608–12618. Institute of Electrical and Electronics Engineers (2023)

    work page 2023

  2. [2]

    ACM transactions on graphics (TOG) 42(4), 1–11 (2023)

    Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM transactions on graphics (TOG) 42(4), 1–11 (2023)

    work page 2023

  3. [3]

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

    Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18208–18218 (2022)

    2022

  4. [4]

    arXiv preprint arXiv:2209.14687 (2022)

    Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sam- pling for general noisy inverse problems. arXiv preprint arXiv:2209.14687 (2022)

    arXiv 2022

  5. [5]

    In: NeurIPS (2022)

    Chung, H., Sim, B., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. In: NeurIPS (2022)

    work page 2022

  6. [6]

    In: ICLR 2023 (Eleventh International Conference on Learning Representations) (2023)

    Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based se- mantic image editing with mask guidance. In: ICLR 2023 (Eleventh International Conference on Learning Representations) (2023)

    work page 2023

  7. [7]

    In: CVPR (2009)

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)

    2009

  8. [8]

    In: NeurIPS (2021)

    Dhariwal, P., Nichol, A.Q.: Diffusion models beat GANs on image synthesis. In: NeurIPS (2021)

    2021

Show all 45 references
  1. [9]

    Advances in neural in- formation processing systems 27 (2014)

    Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural in- formation processing systems 27 (2014)

    2014

  2. [10]

    science 313(5786), 504–507 (2006) GuidPaint 13

    Hinton, G.E., Salakhutdinov, R.R.: Reducing the dimensionality of data with neu- ral networks. science 313(5786), 504–507 (2006) GuidPaint 13

    2006

  3. [11]

    NeurIPS 33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840–6851 (2020)

    2020

  4. [12]

    arXiv preprint arXiv:2207.12598 (2022)

    Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)

    2022 arXiv

  5. [13]

    Advances in Neural Information Processing Systems 35, 8633– 8646 (2022)

    Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. Advances in Neural Information Processing Systems 35, 8633– 8646 (2022)

    2022

  6. [14]

    Advances in Neural Information Processing Systems 35, 23593–23606 (2022)

    Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. Advances in Neural Information Processing Systems 35, 23593–23606 (2022)

    2022

  7. [15]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Kim, S., Suh, S., Lee, M.: Rad: Region-aware diffusion models for image inpainting. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 2439–2448 (2025)

    2025

  8. [16]

    arXiv e-prints arXiv:1312.6114 (Dec 2013)

    Kingma, D.P., Welling, M.: Auto-Encoding Variational Bayes. arXiv e-prints arXiv:1312.6114 (Dec 2013). https://doi.org/10.48550/arXiv.1312.6114

    2013 doi arXiv

  9. [17]

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

    Liu, H., Wang, Y., Qian, B., Wang, M., Rui, Y.: Structure matters: Tackling the semantic discrepancy in diffusion models for image inpainting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8038–8047 (2024)

    2024

  10. [18]

    In: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16

    Liu, H., Jiang, B., Song, Y., Huang, W., Yang, C.: Rethinking image inpainting via a mutual encoder-decoder with feature equalizations. In: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16. pp. 725–741. Springer (2020)

    2020

  11. [19]

    In: ICCV (2015)

    Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: ICCV (2015)

    2015

  12. [20]

    CVPR (2022)

    Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Gool, L.V.: Repaint: Inpainting using denoising diffusion probabilistic models. CVPR (2022)

    2022

  13. [21]

    In: ICLR (2022)

    Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022)

    2022

  14. [22]

    In: ICML (2022)

    Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: GLIDE: towards photorealistic image generation and edit- ing with text-guided diffusion models. In: ICML (2022)

    2022

  15. [23]

    In: International conference on machine learning

    Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: International conference on machine learning. pp. 8162–8171. PMLR (2021)

    2021

  16. [24]

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

    Peng, J., Liu, D., Xu, S., Li, H.: Generating diverse structure for image inpainting with hierarchical vq-vae. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10775–10784 (2021)

    2021

  17. [25]

    arXiv preprint arXiv:2204.06125 1(2), 3 (2022)

    Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text- conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)

    2022 arXiv

  18. [26]

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

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684–10695 (2022)

    2022

  19. [27]

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation (2015), https://arxiv.org/abs/1505.04597

    2015 arXiv

  20. [28]

    In: ACM SIGGRAPH (2022)

    Saharia, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH (2022)

    2022

  21. [29]

    Advances in neural information processing systems 35, 36479–36494 (2022) 14 Q

    Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text- to-image diffusion models with deep language understanding. Advances in neural information processing systems 35...

    2022

  22. [30]

    In: International conference on machine learning

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsuper- vised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256–2265. pmlr (2015)

    2015

  23. [31]

    arXiv preprint arXiv:2010.02502 (2020)

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)

    2020 arXiv

  24. [32]

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

    Suvorov, R., Logacheva, E., Mashikhin, A., Remizova, A., Ashukha, A., Silvestrov, A., Kong, N., Goka, H., Park, K., Lempitsky, V.: Resolution-robust large mask inpainting with fourier convolutions. In: Proceedings of the IEEE/CVF winter conference on applications of computer v...

    2022

  25. [33]

    In: International conference on machine learning

    Van Den Oord, A., Kalchbrenner, N., Kavukcuoglu, K.: Pixel recurrent neural networks. In: International conference on machine learning. pp. 1747–1756. PMLR (2016)

    2016

  26. [34]

    Advances in neural information pro- cessing systems 30 (2017)

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems 30 (2017)

    2017

  27. [35]

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

    Wan, Z., Zhang, J., Chen, D., Liao, J.: High-fidelity pluralistic image completion with transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4692–4701 (2021)

    2021

  28. [36]

    arXiv preprint arXiv:2212.00490 (2022)

    Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490 (2022)

    2022 arXiv

  29. [37]

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

    Xie, S., Zhang, Z., Lin, Z., Hinz, T., Zhang, K.: Smartbrush: Text and shape guided object inpainting with diffusion model. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 22428–22437 (2023)

    2023

  30. [38]

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

    Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Chen, D., Wen, F.: Paint by example: Exemplar-based image editing with diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18381–18391 (2023)

    2023

  31. [39]

    Entropy 25(10), 1469 (2023)

    Yang, R., Srivastava, P., Mandt, S.: Diffusion probabilistic modeling for video generation. Entropy 25(10), 1469 (2023)

    2023

  32. [40]

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

    Yeh, R.A., Chen, C., Yian Lim, T., Schwing, A.G., Hasegawa-Johnson, M., Do, M.N.: Semantic image inpainting with deep generative models. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5485–5493 (2017)

    2017

  33. [41]

    In: Proceedings of the 29th ACM international conference on multimedia

    Yu, Y., Zhan, F., Wu, R., Pan, J., Cui, K., Lu, S., Ma, F., Xie, X., Miao, C.: Diverse image inpainting with bidirectional and autoregressive transformers. In: Proceedings of the 29th ACM international conference on multimedia. pp. 69–78 (2021)

    2021

  34. [42]

    In: International Conference on Machine Learning (2023)

    Zhang, G., Ji, J., Zhang, Y., Yu, M., Jaakkola, T., Chang, S.: Towards coher- ent image inpainting using denoising diffusion implicit models. In: International Conference on Machine Learning (2023)

    2023

  35. [43]

    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

  36. [44]

    In: ICLR (2021)

    Zhao, S., Cui, J., Sheng, Y., Dong, Y., Liang, X., Chang, E.I., Xu, Y.: Large scale image completion via co-modulated generative adversarial networks. In: ICLR (2021)

    2021

  37. [45]

    In: European Conference on Computer Vision

    Zhao, Y., Lian, Z.: Udifftext: A unified framework for high-quality text synthesis in arbitrary images via character-aware diffusion models. In: European Conference on Computer Vision. pp. 217–233. Springer (2024)

    2024

Pith tools

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