Pith. sign in

REVIEW 5 major objections 5 minor 35 references

SVIA: A Street View Image Anonymization Framework for Self-Driving Applications

T0 review · 5 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Street view anonymization can hide location and identity by inpainting people, vehicles, signs, roads, and buildings with diffusion-generated content and then harmonizing the result.

desk verdict A useful engineering pipeline whose central privacy claim doesn't survive its own numbers: ACR is 15x chance and worse than simple masking. read the letter →

arxiv 2501.09393 v1 pith:IYS6TBP2 submitted 2025-01-16 cs.CV

Dongyu Liu , Xuhong Wang , Cen Chen , Yanhao Wang , Shengyue Yao , Yilun Lin This is my paper · ORCID
classification cs.CV
keywords streetviewimageanonymizationlatentdiffusionmodelsemanticsegmentationinpaintingprivacyprotectionself-drivingcityre-identificationharmonization
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 proposes SVIA, a three-stage framework for anonymizing street view images collected for self-driving. Its central claim is that replacing privacy-sensitive content—people, vehicles, traffic signs, roads, and buildings—with realistic synthetic content gives a better trade-off between image quality and privacy protection than blurring, pixelation, masking, or GAN-based face anonymization. The authors argue this matters because face-only de-identification still leaves cues that reveal location, trajectories, and identity in urban imagery, and because highly degraded images harm the perception models that self-driving systems depend on. On Cityscapes and Mapillary Vistas, SVIA is reported to lower city re-identification and person similarity while keeping generation quality metrics competitive.

What carries the argument

The pipeline is built from three components. A semantic segmenter based on a DCNv3 architecture produces pixel-level masks for each semantic category. A latent diffusion model (a generative model that denoises a compressed latent representation conditioned on text and image prompts) inpaints each masked region with plausible synthetic content, with Laplacian noise added to the masked area first to suppress residual identity cues. A harmonizer, implemented as a Stable Diffusion image-to-image model, then blends the separately inpainted regions so that hard lines vanish and remaining sensitive details are removed. The selection of which categories to inpaint is itself load-bearing: Grad-CAM on the city classifier identifies roads and buildings as the main features for city re-identification, so the framework anonymizes location-revealing structure, not only faces and bodies.

What would settle it

Train a separate city classifier on Cityscapes images using only regions that SVIA does not inpaint (sky, vegetation, poles, fences, terrain) and measure its top-1 accuracy on SVIA-anonymized images; if it stays far above 29.56%, the claimed location privacy does not hold against that adversary.

Watch

Extended reading notes

Core claim

The paper's central claim is that a street view image can be effectively anonymized by segmenting it into semantic regions, inpainting the five categories that leak identity or location—person, vehicle, traffic sign, road, and building—with a latent diffusion model, and then passing the stitched result through a harmonizer to erase seams and residual sensitive details. The authors train a city classifier on Cityscapes (99.2% accuracy) and use Grad-CAM to show that roads and buildings are the strongest cues for identifying the city in which a photo was taken, which justifies extending anonymization beyond people and faces. In experiments, SVIA reduces top-1 city re-identification accuracy from 99.25% on original Cityscapes images to 29.56%, while achieving FID of 37.74, and on Mapillary Vistas it produces the lowest person-similarity score among the compared methods. The paper concludes that this combination preserves the validity of images for downstream self-driving tasks better than the baselines.

Load-bearing premise

The privacy guarantee stands or falls on treating the city classifier and person-similarity model used for evaluation as realistic stand-ins for a real adversary; if an attacker can use cues outside the five inpainted categories to recognize the location or a person, the claimed protection weakens.

Editorial extensions

If this is right

  • SVIA cuts top-1 city re-identification from 99.25% to 29.56% on Cityscapes while holding FID at 37.74, which the paper presents as the best quality-versus-privacy trade-off among the methods compared.
  • Because roads and buildings are inpainted, anonymized images no longer expose the specific street or city where the photo was taken, closing a location-leakage channel that face-only methods leave open.
  • The modular design means a better segmenter, inpainter, or harmonizer can be substituted without rebuilding the framework, and the harmonizer alone is shown to improve privacy metrics on both datasets.
  • Anonymized street views remain usable for self-driving data pipelines, with lower FID and KID than model-free baselines and a smaller effect on semantic segmentation than the MAT inpainting baseline.

Reading between the lines

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

  • The privacy evaluation uses the same city classifier that selected the five anonymized categories, so an independent test against a classifier trained on other datasets or on un-inpainted cues (skyline, vegetation, road texture, building architecture) would be the natural audit of the claimed protection.
  • Task-specific measures such as object-detection accuracy or segmentation mIoU on anonymized images would sharpen the utility claim, since FID and KID measure global realism rather than whether a downstream self-driving model still behaves correctly.
  • The per-category inpainting design could extend to video by enforcing temporal consistency across frames, though the paper states SVIA currently handles still images and takes about two minutes per image.
  • A stronger adversary could also exploit subtle perceptual cues that survive inpainting, such as weather patterns or plant species, so the protection level should be read as bounded by the semantic categories chosen rather than as an absolute guarantee.
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.

Referee Report

5 major / 5 minor

Summary. The paper proposes SVIA, a three-stage street-view image anonymization pipeline consisting of a semantic segmenter, a latent-diffusion-based inpainter, and a harmonizer. The pipeline replaces five privacy-sensitive semantic categories (person, vehicle, traffic sign, road, building) with synthetic content, and the authors evaluate it on Cityscapes and Mapillary Vistas using FID, KID, LPIPS, PerSim, and ACR. The central claim, stated in the Abstract and Section IV.C, is that SVIA achieves a much better trade-off between generation quality and privacy protection than existing methods while preserving validity for downstream self-driving tasks. The code is released under an MIT license.

Significance. If the central trade-off claim held, SVIA would be a useful step beyond face-only anonymization, offering a concrete pipeline for reducing identity and location leakage in street-view data collected for self-driving. The pipeline is clearly described, and the release of code is a reproducibility strength. However, the evidence in the paper does not yet establish the claimed trade-off: privacy is measured against a single city classifier that also guided the design of the anonymization categories, the reported ACR is far above chance and worse than GrayMask, and downstream-task validity is asserted rather than measured with actual task metrics. The contribution is therefore promising but not fully supported in its current form.

major comments (5)
  1. [Section III.A and Table II] The privacy evaluation is circular with respect to the threat model. The city classifier trained on Cityscapes is used both to select the five anonymization categories via Grad-CAM (Section III.A) and to measure anonymization success through ACR (Table II). This measures whether SVIA fools that particular classifier, not whether location information is generally removed. Moreover, the ACR numbers are far from chance-level: with 50 cities, random guessing gives 2% top-1 accuracy, while SVIA leaves ACR@1 at 29.5552%, and GrayMask achieves a lower 18.8446%. The paper should report a chance baseline, evaluate against a held-out or adaptive adversary, and test whether cues outside the five chosen categories (e.g., skyline, vegetation, road texture) still reveal the city.
  2. [Abstract, Section IV.C, and Tables I-II] The claim that SVIA achieves a 'much better trade-off' between quality and privacy is stronger than the reported rankings support. In Table I, SVIA is second-best on FID, KID, and LPIPS on Cityscapes and second-best on FID and KID on Mapillary Vistas; on the privacy metric ACR in Table II, SVIA ranks third, behind GrayMask and MAT+Harmonizer. The paper does not define a scalar utility over the five metrics that would justify the 'much better' comparison. The prose in Section IV.C ('SVIA best balances') should be calibrated to the actual multi-objective picture, or a principled aggregation and statistical comparison should be provided.
  3. [Section IV.B and Section IV.C] The downstream-task validity claim is asserted but never measured. Section IV.B states that image generation quality should be measured by the degree to which the accuracy of downstream tasks is affected by anonymization, yet the experiments report only FID and KID. Figure 6 shows qualitative segmentation outputs, but no mIoU, detection accuracy, or any quantitative self-driving task metric is given. The contribution bullet in the Introduction and the Conclusion ('less significant impact on downstream tasks than baselines') therefore require additional experiments with downstream perception models trained or evaluated on the anonymized images.
  4. [Table III] The harmonizer ablation shows that the harmonizer degrades generation quality on Mapillary Vistas: FID increases from 55.0217 to 65.0094 and KID from 0.0053 to 0.0105. The authors acknowledge a 'slight' degradation, but the conclusion still presents the three-component SVIA as achieving the best quality-privacy trade-off. The paper should either justify the harmonizer as dataset-dependent, quantify whether the privacy gains (higher LPIPS, lower PerSim) outweigh the quality loss, or discuss a variant that omits the harmonizer on datasets where it hurts FID/KID.
  5. [Section IV.B] The privacy metrics are not fully aligned with the paper's stated privacy objectives. LPIPS measures perceptual dissimilarity and is not an established privacy metric; a large LPIPS change can result from stylistic alteration without necessarily removing identifying information. PerSim measures person re-identification only, so it does not cover the claimed anonymization of vehicles, roads, buildings, or trajectories. The paper should add direct evaluations of whether vehicle re-identification, trajectory inference, or location inference beyond the single Cityscapes classifier are mitigated.
minor comments (5)
  1. [Algorithm 3] Algorithm 3 uses the variable 'eimg' in the denoising step (line 8 of the loop) but never defines it; it should be defined, for example, as 'eimg ← Eimg(x)' after the image encoder is applied, or the line should be corrected to use the intended conditioning input.
  2. [Section IV.F] The runtime statement that SVIA takes 7 days on Cityscapes and 30 days on Mapillary Vistas is surprising and should be clarified; the paper should state whether these are total wall-clock times for the test sets, report per-image time with hardware details, and discuss how this relates to the claimed deployability in self-driving data pipelines.
  3. [Figure 4] Figure 4 is described as illustrating the trade-off between image quality (KID) and privacy protection (LPIPS), but the paper claims five metrics; the caption should state which metrics are shown and whether the omitted metrics are consistent with the displayed Pareto-style comparison.
  4. [Section V] The limitations paragraph acknowledges the inability to process video, slow generation speed, and possible architecture simplification; these are reasonable, but the conclusion's statement that SVIA 'has the potential to be deployed' should be tempered in light of the runtime numbers and the unmeasured downstream-task impact.
  5. [Table I] DeepPrivacy's nearly zero LPIPS and near-one PerSim indicate that it essentially leaves the input unchanged on street-view images; the text already notes this, but the table would benefit from a footnote clarifying that DeepPrivacy should be read as a non-anonymizing reference point on this data.

Circularity Check

1 steps flagged · score 4.0 of 10

Same city classifier drives both anonymization-category selection and the ACR privacy metric, making the location-privacy claim partially self-referential.

  1. self definitional [Section III-A and Section IV-B (ACR metric); Table II]
    "To verify this, we train a city classifier on the Cityscapes dataset with an accuracy of 99.2%. Then, we utilize the Grad-CAM [25] method to identify the regions in an image that serve as distinguishing criteria for the re-identification of trajectories. ... Consequently, we choose the following five semantic categories in street view images to inpaint: person, vehicle, traffic sign, road, and building. ... We use the city classifier in Section III-A and Fig. 1 to classify the output images and calculate the classification accuracy as ACR."

    The anonymization target categories are selected from Grad-CAM heatmaps of the city classifier, and the location-privacy metric ACR is the accuracy of that same classifier on anonymized images. The evaluation therefore checks whether the cues that this specific classifier flagged were removed; it does not test removal of location information in general. The privacy claim is partly self-referential: the classifier defines both what is sensitive and what counts as success. The result is not forced (ACR@1 remains 29.56% vs 2% random chance, and GrayMask achieves 18.84%), and FID/KID/LPIPS/PerSim are external, so circularity is partial rather than total.

full rationale

The five-category anonymization pipeline is not fitted to the privacy metrics, and the image-quality benchmarks (FID, KID, LPIPS, PerSim) are external, so the core framework retains independent content. However, the location-privacy half of the central trade-off claim is evaluated with the same city classifier that motivated the category selection, creating a closed loop that prevents ACR from independently supporting 'location anonymization.' In addition, the paper asserts downstream-task validity but does not report a quantitative downstream-task benchmark (Section IV-B and V), and it never compares ACR to the 2% chance level; these are missing-support issues rather than circularity. Overall score 4 reflects partial circularity in the privacy evaluation, not in the generation pipeline.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim depends on several hand-chosen numbers and two domain assumptions about what counts as a privacy threat. No new scientific entities are introduced.

free parameters (4)
  • Laplacian noise scale = 0.25
    Added to masked pixels before inpainting (Algorithm 1 line 3); chosen by hand, not optimized or justified.
  • Semantic category set = person, vehicle, traffic sign, road, building
    Selection of which regions count as sensitive; established via Grad-CAM on Cityscapes, then fixed for all experiments.
  • Text prompts for inpainter and harmonizer = not published in paper
    prompti and prompth appear in Algorithm 1 but their exact strings are not given, so outputs depend on unpublished choices.
  • Denoising schedule and steps for Stable Diffusion components = defaults (unspecified)
    Algorithms 2 and 3 rely on alpha, sigma and step count d; the paper does not report the values used.
assumptions (5)
  • standard math The latent diffusion denoising equations (Algorithms 2 and 3) are correct as specified in the cited literature.
    The paper relies on Rombach et al. (2022) without re-deriving the noise schedule.
  • domain assumption Stable Diffusion v2 and InternImage pretrained weights behave as described by their original papers.
    The framework reuses published models without retraining verification in this paper.
  • domain assumption The city classifier trained on Cityscapes is an adequate proxy for an adversary trying to infer location.
    Section III-A uses Grad-CAM from this classifier to justify inpainting roads/buildings, and Table II uses the same classifier to measure privacy.
  • ad hoc to paper The five anonymized categories (person, vehicle, traffic sign, road, building) cover all privacy-relevant information in street scenes.
    This is inferred from one Grad-CAM analysis on Cityscapes and assumed for Mapillary Vistas and other scenes.
  • domain assumption Semantic segmentation masks are accurate enough that inpainting only the masked regions removes the sensitive content.
    Any segmentation error leaves or destroys content; the paper provides no robustness analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SVIA: A Street View Image Anonymization Framework for Self-Driving Applications." pith.science (2026). https://pith.science/paper/IYS6TBP2

@misc{pith2026250109393,
  author       = {Pith},
  title        = {Pith review of: SVIA: A Street View Image Anonymization Framework for Self-Driving Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IYS6TBP2}},
  note         = {Machine review of arXiv:2501.09393}
}
read the original abstract

In recent years, there has been an increasing interest in image anonymization, particularly focusing on the de-identification of faces and individuals. However, for self-driving applications, merely de-identifying faces and individuals might not provide sufficient privacy protection since street views like vehicles and buildings can still disclose locations, trajectories, and other sensitive information. Therefore, it remains crucial to extend anonymization techniques to street view images to fully preserve the privacy of users, pedestrians, and vehicles. In this paper, we propose a Street View Image Anonymization (SVIA) framework for self-driving applications. The SVIA framework consists of three integral components: a semantic segmenter to segment an input image into functional regions, an inpainter to generate alternatives to privacy-sensitive regions, and a harmonizer to seamlessly stitch modified regions to guarantee visual coherence. Compared to existing methods, SVIA achieves a much better trade-off between image generation quality and privacy protection, as evidenced by experimental results for five common metrics on two widely used public datasets.

Figures

Figures reproduced from arXiv: 2501.09393 by the authors.

Figure 1
Figure 1. Illustration of Grad-CAM results for city classification, where the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of the SVIA pipeline. Algorithm 2 Inpainting Model Ω(ϕ, x, prompt) Input: Masked noisy image x˜ ∈ [0, 1]3×h×w; mask ϕ ∈ {0, 1} n×h×w; text prompt prompt; text encoder Et : String → Rs ; image encoder Eimg : [0, 1]3×h×w → Rs and decoder Dimg : Rs → [0, 1]3×h×w, denoising encoder Es(t) : R → Rs , UNet noise sampler S(x, . . . ) : Rs × · · · → Rs ; number of denoising steps d; latent size s; noise strength α1,… view at source ↗
Figure 3
Figure 3. Detailed Architectures of (A) the segmenter, (B) the inpainter, and (C) the harmonizer. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Illustration of the trade-off between image quality (KID) and privacy [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visual comparison of four original images and their corresponding images generated by SVIA w/o Harmonizer, SVIA, and MAT+Harmonizer. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Illustration of the results for semantic segmentation on the images shown in Fig. 5. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 27 canonical work pages

  1. [9]

    LDFA: Latent diffusion face anonymization for self-driving applications,

    M. Klemp, K. R ¨osch, R. Wagner, J. Quehl, and M. Lauer, “LDFA: Latent diffusion face anonymization for self-driving applications,” in CVPRW, 2023, pp. 3198–3204

    work page 2023

  2. [1]

    Advances and open problems in federated learning,

    P. Kairouz et al., “Advances and open problems in federated learning,” Foundations and Trends® in Machine Learning , vol. 14, no. 1–2, pp. 1–210, 2021

    2021

  3. [2]

    Calibrating noise to sensitivity in private data analysis,

    C. Dwork, F. McSherry, K. Nissim, and A. D. Smith, “Calibrating noise to sensitivity in private data analysis,” in TCC, 2006, pp. 265– 284

    work page 2006

  4. [3]

    Anonymization techniques for privacy preserv- ing data publishing: A comprehensive survey,

    A. Majeed and S. Lee, “Anonymization techniques for privacy preserv- ing data publishing: A comprehensive survey,” IEEE Access , vol. 9, pp. 8512–8545, 2021

    work page 2021

  5. [4]

    Flash memory in the emerging age of autonomy,

    S. Heinrich and L. Motors, “Flash memory in the emerging age of autonomy,” in FMS, 2017, pp. 1–10

    work page 2017

  6. [5]

    Garfinkel et al

    S. Garfinkel et al. , De-identification of Personal Information . US Department of Commerce, National Institute of Standards and Tech- nology, 2015

    work page 2015

  7. [6]

    AnonymousNet: Natural face de-identification with measurable privacy,

    T. Li and L. Lin, “AnonymousNet: Natural face de-identification with measurable privacy,” in CVPRW, 2019, pp. 56–65

    work page 2019

  8. [7]

    DeepPrivacy: A generative adversarial network for face anonymization,

    H. Hukkel ˚as, R. Mester, and F. Lindseth, “DeepPrivacy: A generative adversarial network for face anonymization,” in ISVC, 2019, pp. 565– 578

    work page 2019

Show all 35 references
  1. [8]

    CIAGAN: Conditional identity anonymization generative adversarial networks,

    M. Maximov, I. Elezi, and L. Leal-Taix ´e, “CIAGAN: Conditional identity anonymization generative adversarial networks,” in CVPR, 2020, pp. 5447–5456

    2020

  2. [10]

    DeepPrivacy2: Towards realistic full- body anonymization,

    H. Hukkel ˚as and F. Lindseth, “DeepPrivacy2: Towards realistic full- body anonymization,” in WACV, 2023, pp. 1329–1338

    2023

  3. [11]

    Generative adversarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” Advances In Neural Information Processing Systems , vol. 27, pp. 2672–2680, 2014

    2014

  4. [12]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in Neural Information Processing Systems, vol. 33, pp. 6840–6851, 2020

    2020

  5. [13]

    Diffusion models beat GANs on im- age synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat GANs on im- age synthesis,” Advances in Neural Information Processing Systems , vol. 34, pp. 8780–8794, 2021

    2021

  6. [14]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in CVPR, 2022, pp. 10 684–10 695

    2022

  7. [15]

    PrivacyNet: Semi-adversarial networks for multi-attribute face privacy,

    V . Mirjalili, S. Raschka, and A. Ross, “PrivacyNet: Semi-adversarial networks for multi-attribute face privacy,”IEEE Transactions on Image Processing, vol. 29, pp. 9400–9412, 2020

    2020

  8. [16]

    Analyzing and improving the image quality of StyleGAN,

    T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, and T. Aila, “Analyzing and improving the image quality of StyleGAN,” in CVPR, 2020, pp. 8110–8119

    2020

  9. [17]

    Live face de-identification in video,

    O. Gafni, L. Wolf, and Y . Taigman, “Live face de-identification in video,” in ICCV, 2019, pp. 9377–9386

    2019

  10. [18]

    Ad- versarial autoencoders,

    A. Makhzani, J. Shlens, N. Jaitly, I. Goodfellow, and B. Frey, “Ad- versarial autoencoders,” CoRR, vol. abs/1511.05644, 2015

    2015 arXiv

  11. [19]

    CFA-Net: Controllable face anonymization network with identity representation manipulation,

    T. Ma, D. Li, W. Wang, and J. Dong, “CFA-Net: Controllable face anonymization network with identity representation manipulation,” CoRR, vol. abs/2105.11137, 2021

    2021 arXiv

  12. [20]

    U-Net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional networks for biomedical image segmentation,” in MICCAI, 2015, pp. 234–241

    2015

  13. [21]

    Rethinking atrous convolution for semantic image segmentation,

    L. Chen, G. Papandreou, F. Schroff, and H. Adam, “Rethinking atrous convolution for semantic image segmentation,” CoRR, vol. abs/1706.05587, 2017

    2017 arXiv

  14. [22]

    InternImage: Exploring large- scale vision foundation models with deformable convolutions,

    W. Wang, J. Dai, Z. Chen, Z. Huang, Z. Li, X. Zhu, X. Hu, T. Lu, L. Lu, H. Li, X. Wang, and Y . Qiao, “InternImage: Exploring large- scale vision foundation models with deformable convolutions,” in CVPR, 2023, pp. 14 408–14 419

    2023

  15. [23]

    Large scale image completion via co-modulated generative adversarial networks,

    S. Zhao, J. Cui, Y . Sheng, Y . Dong, X. Liang, E. Chang, and Y . Xu, “Large scale image completion via co-modulated generative adversarial networks,” in ICLR, 2021

    2021

  16. [24]

    MAT: Mask- aware transformer for large hole image inpainting,

    W. Li, Z. Lin, K. Zhou, L. Qi, Y . Wang, and J. Jia, “MAT: Mask- aware transformer for large hole image inpainting,” in CVPR, 2022, pp. 10 758–10 768

    2022

  17. [25]

    Grad-CAM: Visual explanations from deep networks via gradient-based localization,

    R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-CAM: Visual explanations from deep networks via gradient-based localization,” in ICCV, 2017, pp. 618–626

    2017

  18. [26]

    The cityscapes dataset for semantic urban scene understanding,

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benen- son, U. Franke, S. Roth, and B. Schiele, “The cityscapes dataset for semantic urban scene understanding,” in CVPR, 2016, pp. 3213–3223

    2016

  19. [27]

    The mapillary vistas dataset for semantic understanding of street scenes,

    G. Neuhold, T. Ollmann, S. Rota Bulo, and P. Kontschieder, “The mapillary vistas dataset for semantic understanding of street scenes,” in ICCV, 2017, pp. 4990–4999

    2017

  20. [28]

    GANs trained by a two time-scale update rule converge to a local Nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochre- iter, “GANs trained by a two time-scale update rule converge to a local Nash equilibrium,” Advances In Neural Information Processing Systems, vol. 30, pp. 6626–6637, 2017

    2017

  21. [29]

    Demystify- ing MMD GANs,

    M. Bi ´nkowski, D. Sutherland, M. Arbel, and A. Gretton, “Demystify- ing MMD GANs,” in ICLR, 2018

    2018

  22. [30]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in CVPR, 2018, pp. 586–595

    2018

  23. [31]

    PASS: Part-aware self-supervised pre-training for person re-identification,

    K. Zhu, H. Guo, T. Yan, Y . Zhu, J. Wang, and M. Tang, “PASS: Part-aware self-supervised pre-training for person re-identification,” in ECCV, 2022, pp. 198–214

    2022

  24. [32]

    Align your latents: High-resolution video synthesis with latent diffusion models,

    A. Blattmann, R. Rombach, H. Ling, T. Dockhorn, S. W. Kim, S. Fidler, and K. Kreis, “Align your latents: High-resolution video synthesis with latent diffusion models,” in CVPR, 2023, pp. 22 563– 22 575

    2023

  25. [33]

    Sora: A review on background, technology, limitations, and opportunities of large vision models,

    Y . Liu, K. Zhang, Y . Li, Z. Yan, C. Gao, R. Chen, Z. Yuan, Y . Huang, H. Sun, J. Gao, L. He, and L. Sun, “Sora: A review on background, technology, limitations, and opportunities of large vision models,” CoRR, vol. abs/2402.17177, 2024

    2024 arXiv

  26. [34]

    Adversarial diffusion distillation,

    A. Sauer, D. Lorenz, A. Blattmann, and R. Rombach, “Adversarial diffusion distillation,” CoRR, vol. abs/2311.17042, 2023

    2023 arXiv

  27. [35]

    SSMG: Spatial-semantic map guided diffusion model for free-form layout-to-image generation,

    C. Jia, M. Luo, Z. Dang, G. Dai, X. Chang, M. Wang, and J. Wang, “SSMG: Spatial-semantic map guided diffusion model for free-form layout-to-image generation,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 3, pp. 2480–2488, 2024

    2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.