Pith. sign in

REVIEW 3 major objections 5 minor 76 references

IELDG: Suppressing Domain-Specific Noise with Inverse Evolution Layers for Domain Generalized Semantic Segmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Inverse evolution layers — fixed Laplacian-kernel filters that amplify defects — clean both diffusion-generated images and segmentation predictions, lifting cross-domain mIoU by up to 1.6 points.

desk verdict A credible incremental recipe for cleaning diffusion-generated data and decoder features, but the headline gains are inflated by choosing the main hyperparameter on the target test sets. read the letter →

arxiv 2508.19604 v1 pith:W7LOBDDW submitted 2025-08-27 cs.CV cs.AI

Qizhe Fan , Chaoyu Liu , Zhonghua Qiao , Xiaoqin Shen This is my paper
classification cs.CVcs.AI
keywords domaingeneralizedsemanticsegmentationinverseevolutionlayers(IELs)Laplaciankernelpriorsdiffusion-baseddataaugmentationmulti-scalefrequencyfusionsynthetic-to-realgeneralizationartifactsuppressionMask2Formerdecoder
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

Domain-generalized semantic segmentation trains on a labeled synthetic source, such as game-rendered street scenes, and must work on unseen real cities. A popular remedy is to add diffusion-generated images to the training set, but those images carry structural and semantic errors that can hurt rather than help. The paper's central proposal is to insert inverse evolution layers (IELs) — fixed Laplacian-kernel filters that deliberately amplify defects — into both sides of the pipeline: into the diffusion generation loop to make synthetic images cleaner, and into the decoder of a transformer-based segmentation head to amplify prediction errors so the network learns to correct them. With this single mechanism plus a Fourier-domain multi-scale fusion module, the full IELDG framework improves average mIoU on the GTA-trained setting by up to 1.6 points over the Rein baseline and 1.2 over SoMA, with only a small increase in trainable parameters. The contribution is a practical, near parameter-free plug-in that attacks error accumulation in both the data and the model.

What carries the argument

The load-bearing object is the inverse evolution layer (IEL): a fixed, parameter-free ensemble of Laplacian kernels that detects and amplifies 'negative properties' — spatial discontinuities and semantic inconsistencies — in whatever feature map it is inserted after, producing stronger gradient feedback that pushes the network to suppress those defects. It is placed at two sites: inside the diffusion model's generation loop (IELDM) and after each of four feature scales in the segmentation decoder (IELFormer). The secondary mechanism is the multi-scale frequency fusion (MFF) module, which applies a 2D FFT, separates amplitude from phase, fuses each with learnable weights, and returns via inve

What would settle it

Replace the Laplacian kernels inside the IELs with random fixed 3x3 kernels while keeping the same feedback pathway, and retrain both IELDM and IELFormer; if the reported mIoU gains on GTA-to-Cityscapes/BDD100K/Mapillary do not largely disappear, then the specific Laplacian-based defect amplification, rather than the extra feedback loop, is not what drives the improvement.

Watch

Extended reading notes

Core claim

The core discovery, on the paper's own terms, is that the same defect-amplification operator can serve two different roles in domain-generalized semantic segmentation. In the generative role (IELDM), the Laplacian-based IELs are applied to intermediate features of a mask-conditioned text-to-image diffusion model; they make subtle structural and semantic defects more salient and feed that amplified signal back into generation, so the model learns to avoid them. In the segmentation role (IELFormer), IELs are inserted after four feature levels of a Mask2Former-style pixel decoder; they highlight uncertain or erroneous regions in the predictions, effectively giving the network a targeted self-co

Load-bearing premise

The framework rests on the assumption that amplifying low-level spatial defects with Laplacian kernels inside a diffusion model's intermediate features and inside the segmentation decoder actually steers both networks to correct those defects in the intended direction; the paper adopts this mechanism and its theory from the earlier IEL paper instead of re-analyzing the transfer to these settings.

Editorial extensions

If this is right

  • Using IELDM alone to augment the source set adds +0.93 mIoU for Rein while introducing no trainable parameters, since the IEL kernels are fixed and the diffusion model is frozen.
  • Using the full IELFormer decoder (MFF + IELs) adds +0.70 mIoU for Rein beyond the baseline, with MFF alone contributing +0.35 and IELs alone +0.49; the complete IELDG pipeline reaches +1.63 mIoU for Rein and +1.18 for SoMA on the GTA-trained setting.
  • The gains transfer across training configurations: similar improvements appear when training on GTA+SYNTHIA and on GTA+SYNTHIA+UrbanSyn, not only in the single-source case.
  • The optimal operating point is shallow: five IELs outperform ten or twenty in both generated-image quality (CMMD) and downstream segmentation mIoU, because excessive amplification over-enhances minor defects.

Reading between the lines

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

  • Untested extension: the same Laplacian feedback loop could be inserted into other generative models, such as GANs or video generators, and into other dense predictors; the paper's mechanism is not inherently tied to diffusion or semantic segmentation.
  • Untested extension: since depth 5 is globally optimal, an adaptive policy that tunes IEL depth or gain per feature level, or per training stage, could outperform the fixed configuration; the paper reports no such adaptive control.
  • Interpretable check: the MFF module's learned amplitude/phase weights should, under the authors' stated story, favor phase from the high-resolution map and amplitude from the low-resolution map; inspecting them would test the frequency-fusion rationale directly.
  • Focused prediction: because gains concentrate in boundary- and small-instance classes such as sidewalk, rider, motorbike, and traffic light, a boundary- or uncertainty-based evaluation would reveal whether IELs mainly sharpen edges rather than provide general noise suppression; this goes beyond the paper's reported analyses.
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 / 5 minor

Summary. The paper proposes IELDG, a domain-generalized semantic segmentation framework that combines two uses of "inverse evolution layers" (IELs): (i) IELDM, a diffusion-based data augmentation pipeline in which IELs amplify structural defects in generated images so the generator learns to suppress them, and (ii) IELFormer, a Mask2Former-based decoder in which IELs are inserted at four feature scales to highlight prediction artifacts, together with a multi-scale frequency fusion (MFF) module that fuses amplitude and phase components across resolutions. Experiments are reported on GTA/SYNTHIA/UrbanSyn to Cityscapes/BDD100K/Mapillary, claiming consistent improvements of up to +1.6 mIoU over Rein and +1.2 mIoU over SoMA, with ablations attributing the gains to IELDM, MFF, and IELs.

Significance. The topic is timely and the empirical setting is standard for DGSS. If the claimed gains were valid, the paper would offer a practical, lightweight plug-in for two competitive baselines. The paper also gives credit for reporting averaged results over 3 seeds and for using a recent distributional metric (CMMD) for generation quality. However, the central empirical claim is undermined by a test-set leakage in the selection of the IEL depth, and by the absence of any variance reporting. The core mechanism is imported from a self-cited paper [29] with no new derivation, so the novelty rests almost entirely on the empirical case, which is not currently supported.

major comments (3)
  1. [Sec. 4.3, Tab. 4] The IEL depth hyperparameter is selected using the target test benchmarks themselves. Tab. 4 reports average mIoU on GTA -> {City, BDD, Map} for depths 20/10/5/0, and depth 5 is then fixed for all subsequent IELDG comparisons (Tabs. 1, 2, 5). These three datasets are exactly the unseen target domains on which the final method is evaluated, so the selection procedure is a form of test-set leakage. With four depths tried, the best target-domain depth can outperform the average by chance; the paper reports no seed-wise variance. The selected depth's advantage in Tab. 4 is only +0.49/+0.52 mIoU, while the headline gains in Tab. 1 are +1.6/+1.2, so the final claim of 'consistent enhancement' is not supported as an unbiased estimate. A validation-based selection or a complete report of all depths is required.
  2. [Sec. 4.2, Tables 1-2] The text says results are averaged over 3 random seeds, but no standard deviations, confidence intervals, or per-seed values are provided. Many reported differences are small (e.g., Rein+IELDG 60.51 vs Rein 60.16 on GTA->BDD, Tab. 1; +0.35 mIoU). Without any measure of variability, the claim that IELDG 'consistently enhances' both baselines across all settings cannot be statistically assessed. Error bars or significance tests are required for the central empirical claim.
  3. [Sec. 3.1] The core mechanism is imported from ref. [29], and the paper explicitly defers there for 'rigorous theoretical proofs.' No derivation or analysis is given for why Laplacian-kernel negative-property amplification, originally proposed for image segmentation, transfers to (i) intermediate features of Stable Diffusion/ControlNet during generation (Sec. 3.2) and (ii) four pixel-decoder feature levels of Mask2Former (Sec. 3.3). This is a novel application, so the paper should provide at least a mechanism analysis or an ablation that isolates the IEL effect without relying on the target-selected depth. As it stands, the +0.49/+0.52 attribution to IELs in Tab. 5 is entangled with the leakage issue.
minor comments (5)
  1. [Sec. 3.3, Fig. 4] 'IELF ormer' and 'IFFF ormer' typos appear in the architecture description and figure; please fix.
  2. [Tab. 3] The arrow/delta notation (e.g., '↓0.453') is not defined. Clarify which baseline the deltas are computed against.
  3. [Sec. 4.1] 'We adopt only basic data augmentation techniques' is vague; specify the augmentation list or explicitly cite which Rein protocol is followed.
  4. [Tab. 1] The 'Params*' footnote is unclear: CLOUDS is listed as 0.0M, presumably because the backbone is frozen. State this explicitly.
  5. [Figs. 6-7] The visual comparisons in Fig. 6 and Fig. 7 appear to overlap substantially; please reconcile or consolidate to avoid redundancy.

Circularity Check

2 steps flagged · score 6.0 of 10

Test-set selection of IEL depth underpins the headline gains, while the core IEL mechanism is deferred to a self-cited prior paper.

  1. fitted input called prediction [Sec. 4.3, Tab. 4; carried into Tabs. 1, 2, 5]
    "Tab. 4 presents the average mIoU results on the GTA →{City, BDD, Map } benchmark under varying IEL depths within the IELFormer framework. ... The optimal performance is attained at an IEL depth of 5, where the mIoU improves by +0.49 for Rein and +0.52 for SoMA."

    The IEL depth is a hyperparameter selected by maximizing the average mIoU on exactly the three target test domains (City, BDD, Map) that later define the headline 'unseen target' evaluation in Tab. 1. Because Tab. 1's Rein+IELDG and SoMA+IELDG use this optimum depth, the reported +1.6/+1.2 gains are not independent predictions; they are the result of choosing the configuration on the test benchmark. With only four depth values (20/10/5/0) and no seed-wise error bars, selecting the best target-domain value biases the comparison upward even under pure noise. This is test-set leakage presented as a validated improvement.

  2. self citation load bearing [Sec. 3.1; relied on in Sec. 3.2 and Sec. 3.3]
    "For comprehensive technical details regarding IELs’ implementation and rigorous theoretical proofs of their effectiveness, we direct readers to the foundational reference [29], which provides exhaustive mathematical analysis and experimental validations."

    The core mechanism behind both IELDM and IELFormer is imported without derivation from [29], a prior paper by overlapping authors (Liu and Qiao are authors of both works). The present paper contributes no proof or analysis that the Laplacian-kernel 'negative property amplification' transfers to Stable Diffusion/ControlNet intermediate features or to Mask2Former's pixel-decoder levels; it explicitly defers to [29] for the theory. Thus the central premise that IELs suppress artifacts in these new settings rests on a self-citation rather than on evidence developed here. The later ablations (Tabs. 4-5) do test the module, but they are performed on the target test domains, so they do not provide clean independent support.

full rationale

The paper's mathematical constructions are not definitionally circular: IELDM and MFF are described with concrete operations (FFT/IFFT, Laplacian kernels, learnable weighted sums), and the final system is benchmarked against many external methods. However, the central evidence for the claim of 'consistent enhancement' is weakened by selection-on-test-set procedures. Sec. 4.3 chooses the IEL depth by maximizing average mIoU on the exact three target domains that are later used as unseen test domains in Tab. 1 and Tab. 2; the component ablation in Sec. 4.4 similarly selects components on those same targets. This makes the reported gains partly a product of post-hoc selection rather than independent predictions. In addition, the paper's core mechanism (IELs) is imported from a self-cited prior work by overlapping authors, with the theoretical justification explicitly deferred to that citation and no transfer analysis given for the new settings. Consequently, the empirical claim is not fully independent of its own fitted choices. The paper is not equivalent to its inputs by definition, so a score of 6 (partial circularity) is appropriate rather than 8 or 10.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The framework rests on a self-cited prior mechanism (IELs) that is not derived here, plus a hyperparameter (IEL depth) selected using the evaluation benchmarks. These are the main unverified inputs beyond standard model training components.

free parameters (2)
  • IEL depth (number of stacked IELs) = 5
    Selected via ablation on generation quality (Tab. 3, CMMD and mIoU) and on the actual DGSS target benchmarks (Tab. 4, GTA to City/BDD/Map). This is a test-set-dependent hyperparameter choice that inflates reported generalization.
  • Number of synthetic images used for augmentation = 5,000
    Fixed in all experiments; no sensitivity analysis or derivation, but it affects the training distribution and results.
assumptions (4)
  • domain assumption IELs as described in reference [29] effectively amplify negative properties via Laplacian kernels and improve network learning.
    The paper does not restate or derive IELs; Sec. 3.1 defers to [29] for 'rigorous theoretical proofs'. The current work relies on this external, self-authored result.
  • domain assumption Feeding amplified structural defects back through the diffusion model and the segmentation decoder produces corrective gradients that reduce artifacts rather than distorting the output.
    This is the core causal mechanism behind IELDM and IELFormer (Sec. 3.2, 3.3), stated qualitatively with no gradient or stability analysis.
  • ad hoc to paper Amplitude-phase decomposition with learnable weights improves cross-scale semantic consistency in the decoder.
    The MFF module (Sec. 3.3, Eqs. 2-7) is introduced to address a 'lack of multi-scale aggregation' in Mask2Former; its benefit is supported only by a small ablation without error bars.
  • domain assumption The reported average mIoU differences are not explained by test-set hyperparameter tuning or random seed variation.
    No held-out validation is used; Tab. 4 evaluates candidate IEL depths directly on the target test domains, and no standard deviations are reported (Tabs. 1, 5).

how reviews work

0 comments
Cite this review

Pith. "Pith review of IELDG: Suppressing Domain-Specific Noise with Inverse Evolution Layers for Domain Generalized Semantic Segmentation." pith.science (2026). https://pith.science/paper/W7LOBDDW

@misc{pith2026250819604,
  author       = {Pith},
  title        = {Pith review of: IELDG: Suppressing Domain-Specific Noise with Inverse Evolution Layers for Domain Generalized Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W7LOBDDW}},
  note         = {Machine review of arXiv:2508.19604}
}
read the original abstract

Domain Generalized Semantic Segmentation (DGSS) focuses on training a model using labeled data from a source domain, with the goal of achieving robust generalization to unseen target domains during inference. A common approach to improve generalization is to augment the source domain with synthetic data generated by diffusion models (DMs). However, the generated images often contain structural or semantic defects due to training imperfections. Training segmentation models with such flawed data can lead to performance degradation and error accumulation. To address this issue, we propose to integrate inverse evolution layers (IELs) into the generative process. IELs are designed to highlight spatial discontinuities and semantic inconsistencies using Laplacian-based priors, enabling more effective filtering of undesirable generative patterns. Based on this mechanism, we introduce IELDM, an enhanced diffusion-based data augmentation framework that can produce higher-quality images. Furthermore, we observe that the defect-suppression capability of IELs can also benefit the segmentation network by suppressing artifact propagation. Based on this insight, we embed IELs into the decoder of the DGSS model and propose IELFormer to strengthen generalization capability in cross-domain scenarios. To further strengthen the model's semantic consistency across scales, IELFormer incorporates a multi-scale frequency fusion (MFF) module, which performs frequency-domain analysis to achieve structured integration of multi-resolution features, thereby improving cross-scale coherence. Extensive experiments on benchmark datasets demonstrate that our approach achieves superior generalization performance compared to existing methods.

Figures

Figures reproduced from arXiv: 2508.19604 by the authors.

Figure 1
Figure 1. (A) Comparison of synthetic images generated by original DM and our IELDM. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Core Mechanism of IELs. IELs operate as intelligent negative property amplifiers through their Laplacian kernel￾based architecture, which systematically detects and exacerbates undesirable characteristics in neural network outputs. When suboptimal features (e.g., noise artifacts or geometric irregularities) are fed into IELs, the architecture performs spectral decomposition to quantify these deficiencies, subsequent… view at source ↗
Figure 3
Figure 3. Overview of the IELDM framework for enhanced image generation. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: (A) Overall pipeline of the proposed DGSS framework integrated with IELFormer. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of segmentation results on GTA → Cityscapes. Our proposed IELDG framework demonstrates superior visual accuracy and coherence compared to the two state-of-the-art baselines, Rein and SoMA, effectively preserving fine details and reducing artifact…
Figure 6
Figure 6. Figure 6: Visual comparison of generated images under different IEL depths. Without IELs, the outputs suffer from severe artifacts and misclassification. As the IEL depth decreases, both structural fidelity and semantic accuracy improve, with the depth of 5 yielding the most rea…
Figure 7
Figure 7. Figure 7: Fidelity and realism comparison of generated images with and without IELs. Incorporating IELs leads to improved fidelity and enhanced realism. in visually implausible artifacts such as blurry fences, merged humans, and stairs that lack distinct steps or proper structur…
Figure 8
Figure 8. Figure 8: Visual comparison of multi-scale feature maps with (red boxes) and without (blue boxes) the proposed MFF module. The inclusion of MFF enhances semantic coherence and structural consistency across layers, confirming its effectiveness in facilitating cross-scale feature …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

76 extracted references · 71 canonical work pages

  1. [29]

    C. Liu, Z. Qiao, C. Li, C.-B. Sch¨ onlieb, Inverse evolution layers: Physics-informed regularizers for image segmentation, SIAM Journal on Mathematics of Data Science 7 (1) (2025) 55–85

    work page 2025

  2. [1]

    L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, A. L. Yuille, Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs, IEEE Trans. Pattern Anal. Mach. Intell. 40 (4) (2017) 834–848

    2017

  3. [2]

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, P. Luo, Segformer: Simple and efficient design for semantic segmentation with transformers, in: NeurIPS, 2021

    work page 2021

  4. [3]

    Cheng, I

    B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, R. Girdhar, Masked-attention mask transformer for universal image segmentation, in: CVPR, 2022, pp. 1290–1299

    work page 2022

  5. [4]

    Quinonero-Candela, M

    J. Quinonero-Candela, M. Sugiyama, A. Schwaighofer, N. D. Lawrence, Dataset shift in machine learning, The MIT Press, 2008. 16

    work page 2008

  6. [5]

    Y. Ai, Y. Qi, B. Wang, Y. Cheng, X. Wang, R. T. Tan, Domain-adaptive 2d human pose estimation via dual teachers in extremely low-light conditions, in: ECCV, 2024, pp. 221–239

    work page 2024

  7. [6]

    T. Chen, B. Lin, Y. Jin, W. Yan, W. Ye, Y. Yuan, R. T. Tan, Dual-rain: Video rain removal using assertive and gentle teachers, in: ECCV, 2024, pp. 127–143

    work page 2024

  8. [7]

    W. Yan, R. T. Tan, B. Zeng, S. Liu, Deep homography mixture for single image rolling shutter correction, in: ICCV, 2023, pp. 9868–9877

    work page 2023

Show all 76 references
  1. [8]

    Cordts, M

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, B. Schiele, The cityscapes dataset for semantic urban scene understanding, in: CVPR, 2016, pp. 3213–3223

    2016

  2. [9]

    F. Yu, H. Chen, X. Wang, W. Xian, Y. Chen, F. Liu, V. Madhavan, T. Darrell, Bdd100k: A diverse driving dataset for heterogeneous multitask learning, in: CVPR, 2020, pp. 2636–2645

    2020

  3. [10]

    Neuhold, T

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

    2017

  4. [11]

    Tsai, W.-C

    Y.-H. Tsai, W.-C. Hung, S. Schulter, K. Sohn, M.-H. Yang, M. Chandraker, Learning to adapt structured output space for semantic segmentation, in: CVPR, 2018, pp. 7472–7481

    2018

  5. [12]

    Tranheden, V

    W. Tranheden, V. Olsson, J. Pinto, L. Svensson, Dacs: Domain adaptation via cross-domain mixed sampling, in: W ACV, 2021, pp. 1379–1389

    2021

  6. [13]

    Zhang, B

    P. Zhang, B. Zhang, T. Zhang, D. Chen, Y. Wang, F. Wen, Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation, in: CVPR, 2021, pp. 12414–12424

    2021

  7. [14]

    Hoyer, D

    L. Hoyer, D. Dai, L. Van Gool, Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation, in: CVPR, 2022, pp. 9924–9935

    2022

  8. [15]

    Hoyer, D

    L. Hoyer, D. Dai, L. Van Gool, Hrda: Context-aware high-resolution domain-adaptive semantic segmentation, in: ECCV, 2022, pp. 372–391

    2022

  9. [16]

    Hoyer, D

    L. Hoyer, D. Dai, H. Wang, L. Van Gool, Mic: Masked image consistency for context-enhanced domain adaptation, in: CVPR, 2023, pp. 11721–11732

    2023

  10. [17]

    B. Pak, B. Woo, S. Kim, D.-h. Kim, H. Kim, Textual query-driven mask transformer for domain generalized segmentation, in: ECCV, 2024, pp. 37–54

    2024

  11. [18]

    Benigmim, S

    Y. Benigmim, S. Roy, S. Essid, V. Kalogeiton, S. Lathuili` ere, Collaborating foundation models for domain generalized semantic segmentation, in: CVPR, 2024, pp. 3108–3119

    2024

  12. [19]

    Z. Wei, L. Chen, Y. Jin, X. Ma, T. Liu, P. Ling, B. Wang, H. Chen, J. Zheng, Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation, in: CVPR, 2024, pp. 28619–28630

    2024

  13. [20]

    Q. Bi, J. Yi, H. Zheng, H. Zhan, Y. Huang, W. Ji, Y. Li, Y. Zheng, Learning frequency-adapted vision foundation model for domain generalized semantic segmentation, NeurIPS 37 (2024) 94047–94072

    2024

  14. [21]

    Zhang, R

    X. Zhang, R. T. Tan, Mamba as a bridge: Where vision foundation models meet vision language models for domain- generalized semantic segmentation, in: CVPR, 2025, pp. 14527–14537

    2025

  15. [22]

    S. Yun, S. Chae, D. Lee, Y. Ro, Soma: Singular value decomposed minor components adaptation for domain generalizable representation learning, in: CVPR, 2025, pp. 25602–25612

    2025

  16. [23]

    Q. Bi, J. Yi, H. Huang, H. Zheng, H. Zhan, Y. Huang, Y. Li, X. Wu, Y. Zheng, Nightadapter: Learning a frequency adapter for generalizable night-time scene segmentation, in: CVPR, 2025, pp. 23838–23849

    2025

  17. [24]

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

    2020

  18. [25]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High-resolution image synthesis with latent diffusion models, in: CVPR, 2022, pp. 10684–10695

    2022

  19. [26]

    Zhang, A

    L. Zhang, A. Rao, M. Agrawala, Adding conditional control to text-to-image diffusion models, in: ICCV, 2023, pp. 3836–3847

    2023

  20. [27]

    Y. Jia, L. Hoyer, S. Huang, T. Wang, L. Van Gool, K. Schindler, A. Obukhov, Dginstyle: Domain-generalizable semantic segmentation with image diffusion models and stylized semantic control, in: ECCV, 2024, pp. 91–109

    2024

  21. [28]

    Niemeijer, M

    J. Niemeijer, M. Schwonberg, J.-A. Term¨ ohlen, N. M. Schmidt, T. Fingscheidt, Generalization by adaptation: Diffusion- based domain extension for domain-generalized semantic segmentation, in: W ACV, 2024, pp. 2830–2840

    2024

  22. [30]

    X. Pan, P. Luo, J. Shi, X. Tang, Two at once: Enhancing learning and generalization capacities via ibn-net, in: ECCV, 2018, pp. 464–479

    2018

  23. [31]

    Ahn, G.-Y

    W.-J. Ahn, G.-Y. Yang, H.-D. Choi, M.-T. Lim, Style blind domain generalized semantic segmentation via covariance alignment and semantic consistence contrastive learning, in: CVPR, 2024, pp. 3616–3626

    2024

  24. [32]

    M. Jing, X. Zhen, J. Li, C. G. Snoek, Order-preserving consistency regularization for domain adaptation and generalization, in: ICCV, 2023, pp. 18916–18927

    2023

  25. [33]

    Z. Wu, X. Wu, X. Zhang, L. Ju, S. Wang, Siamdoge: Domain generalizable semantic segmentation using siamese network, in: ECCV, 2022, pp. 603–620

    2022

  26. [34]

    Y. Zhao, Z. Zhong, N. Zhao, N. Sebe, G. H. Lee, Style-hallucinated dual consistency learning for domain generalized semantic segmentation, in: ECCV, 2022, pp. 535–552

    2022

  27. [35]

    S. Choi, S. Jung, H. Yun, J. T. Kim, S. Kim, J. Choo, Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening, in: CVPR, 2021, pp. 11580–11590

    2021

  28. [36]

    D. Peng, Y. Lei, M. Hayat, Y. Guo, W. Li, Semantic-aware domain generalized segmentation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 2594–2605

    2022

  29. [37]

    H. Niu, L. Xie, J. Lin, S. Zhang, Exploring semantic consistency and style diversity for domain generalized semantic segmentation, in: AAAI, Vol. 39, 2025, pp. 6245–6253. 17

    2025

  30. [38]

    Zhong, Y

    Z. Zhong, Y. Zhao, G. H. Lee, N. Sebe, Adversarial style augmentation for domain generalized urban-scene segmentation, Vol. 35, 2022, pp. 338–350

    2022

  31. [39]

    Chattopadhyay, K

    P. Chattopadhyay, K. Sarangmath, V. Vijaykumar, J. Hoffman, Pasta: Proportional amplitude spectrum training aug- mentation for syn-to-real domain generalization, in: ICCV, 2023, pp. 19288–19300

    2023

  32. [40]

    D. Peng, Y. Lei, L. Liu, P. Zhang, J. Liu, Global and local texture randomization for synthetic-to-real semantic segmen- tation, IEEE Transactions on Image Processing 30 (2021) 6594–6608

    2021

  33. [41]

    Podell, Z

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. M¨ uller, J. Penna, R. Rombach, Sdxl: Improving latent diffusion models for high-resolution image synthesis, in: ICLR, 2024

    2024

  34. [42]

    Fahes, T.-H

    M. Fahes, T.-H. Vu, A. Bursuc, P. P´ erez, R. De Charette, A simple recipe for language-guided domain generalized segmentation, in: CVPR, 2024, pp. 23428–23437

    2024

  35. [43]

    Q. Fan, M. Segu, Y.-W. Tai, F. Yu, C.-K. Tang, B. Schiele, D. Dai, Towards robust object detection invariant to real-world domain shifts, in: ICLR, 2023

    2023

  36. [44]

    S. Lee, H. Seong, S. Lee, E. Kim, Wildnet: Learning domain generalized semantic segmentation from the wild, in: CVPR, 2022, pp. 9936–9946

    2022

  37. [45]

    Vidit, M

    V. Vidit, M. Engilberge, M. Salzmann, Clip the gap: A single domain generalization approach for object detection, in: CVPR, 2023, pp. 3219–3229

    2023

  38. [46]

    Udupa, P

    S. Udupa, P. Gurunath, A. Sikdar, S. Sundaram, Mrfp: Learning generalizable semantic segmentation from sim-2-real with multi-resolution feature perturbation, in: CVPR, 2024, pp. 5904–5914

    2024

  39. [47]

    Brown, B

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al., Language models are few-shot learners, NeurIPS 33 (2020) 1877–1901

    2020

  40. [48]

    Touvron, L

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, et al., Llama 2: Open foundation and fine-tuned chat models, arXiv preprint arXiv:2307.09288 (2023)

    2023 arXiv

  41. [49]

    D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, et al., Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, arXiv preprint arXiv:2501.12948 (2025)

    2025 arXiv

  42. [50]

    Benigmim, S

    Y. Benigmim, S. Roy, S. Essid, V. Kalogeiton, S. Lathuili` ere, One-shot unsupervised domain adaptation with personalized diffusion models, in: CVPR, 2023, pp. 698–708

    2023

  43. [51]

    M. B. Sarıyıldız, K. Alahari, D. Larlus, Y. Kalantidis, Fake it till you make it: Learning transferable representations from synthetic imagenet clones, in: CVPR, 2023, pp. 8011–8021

    2023

  44. [52]

    R. Gong, M. Danelljan, H. Sun, J. D. Mangas, L. Van Gool, Prompting diffusion representations for cross-domain semantic segmentation, arXiv preprint arXiv:2307.02138 (2023)

    2023 arXiv

  45. [53]

    Qiu, Y.-X

    R.-Z. Qiu, Y.-X. Wang, K. Hauser, Aligndiff: aligning diffusion models for general few-shot segmentation, in: ECCV, 2024, pp. 384–400

    2024

  46. [54]

    Namekata, A

    K. Namekata, A. Sabour, S. Fidler, S. W. Kim, Emerdiff: Emerging pixel-level semantic knowledge in diffusion models, in: ICLR, 2024

    2024

  47. [55]

    Couairon, M

    P. Couairon, M. Shukor, J.-E. Haugeard, M. Cord, N. Thome, Diffcut: Catalyzing zero-shot semantic segmentation with diffusion features and recursive normalized cut, NeurIPS 37 (2024) 13548–13578

    2024

  48. [56]

    J. Tian, L. Aggarwal, A. Colaco, Z. Kira, M. Gonzalez-Franco, Diffuse attend and segment: Unsupervised zero-shot segmentation using stable diffusion, in: CVPR, 2024, pp. 3554–3563

    2024

  49. [57]

    Y. Liu, T. Yang, S. Huang, Z. Zhang, H. Huang, F. Wei, W. Deng, F. Sun, Q. Zhang, Text diffusion with reinforced conditioning, in: AAAI, Vol. 38, 2024, pp. 14069–14077

    2024

  50. [58]

    Zhang, X

    H. Zhang, X. Liu, J. Zhang, Diffusum: Generation enhanced extractive summarization with diffusion, 2023, pp. 13089– 13100

    2023

  51. [59]

    Lovelace, V

    J. Lovelace, V. Kishore, Y. Chen, K. Q. Weinberger, Diffusion guided language modeling, 2024, pp. 14936–14952

    2024

  52. [60]

    Q. Fang, Y. Zhou, S. Guo, S. Zhang, Y. Feng, Llama-omni2: Llm-based real-time spoken chatbot with autoregressive streaming speech synthesis, 2025, pp. 18617–18629

    2025

  53. [61]

    Zhang, L

    Z. Zhang, L. Yao, B. Wang, D. Jha, G. Durak, E. Keles, A. Medetalibeyoglu, U. Bagci, Diffboost: Enhancing medical image segmentation via text-guided diffusion model, IEEE Transactions on Medical Imaging (2024)

    2024

  54. [62]

    J. Wu, W. Ji, H. Fu, M. Xu, Y. Jin, Y. Xu, Medsegdiff-v2: Diffusion-based medical image segmentation with transformer, in: AAAI, Vol. 38, 2024, pp. 6030–6038

    2024

  55. [63]

    S. R. Richter, V. Vineet, S. Roth, V. Koltun, Playing for data: Ground truth from computer games, in: ECCV, 2016, pp. 102–118

    2016

  56. [64]

    G. Ros, L. Sellart, J. Materzynska, D. Vazquez, A. M. Lopez, The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes, in: CVPR, 2016, pp. 3234–3243

    2016

  57. [65]

    J. L. G´ omez, M. Silva, A. Seoane, A. Borr´ as, M. Noriega, G. Ros, J. A. Iglesias-Guitian, A. M. L´ opez, All for one, and one for all: Urbansyn dataset, the third musketeer of synthetic driving scenes, Neurocomputing 637 (2025) 130038

    2025

  58. [66]

    Contributors, MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark, https://github.com/ open-mmlab/mmsegmentation (2020)

    M. Contributors, MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark, https://github.com/ open-mmlab/mmsegmentation (2020)

    2020

  59. [67]

    Oquab, T

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

    2023

  60. [68]

    Loshchilov, F

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization, arXiv preprint arXiv:1711.05101 (2017)

    2017 arXiv

  61. [69]

    N. Ruiz, Y. Li, V. Jampani, Y. Pritch, M. Rubinstein, K. Aberman, Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation, in: CVPR, 2023, pp. 22500–22510

    2023

  62. [70]

    Jayasumana, S

    S. Jayasumana, S. Ramalingam, A. Veit, D. Glasner, A. Chakrabarti, S. Kumar, Rethinking fid: Towards a better evaluation metric for image generation, in: CVPR, 2024, pp. 9307–9315

    2024

  63. [71]

    H¨ ummer, M

    C. H¨ ummer, M. Schwonberg, L. Zhou, H. Cao, A. Knoll, H. Gottschalk, Vltseg: Simple transfer of clip-based vision- 18 language representations for domain generalized semantic segmentation, in: ACCV, 2024

    2024

  64. [72]

    Liu, C.-Y

    S.-Y. Liu, C.-Y. Wang, H. Yin, P. Molchanov, Y.-C. F. Wang, K.-T. Cheng, M.-H. Chen, Dora: Weight-decomposed low-rank adaptation, in: ICML, 2024

    2024

  65. [73]

    M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, S.-N. Lim, Visual prompt tuning, in: ECCV, 2022, pp. 709–727

    2022

  66. [74]

    S. Chen, C. Ge, Z. Tong, J. Wang, Y. Song, J. Wang, P. Luo, Adaptformer: Adapting vision transformers for scalable visual recognition, NeurIPS 35 (2022) 16664–16678

    2022

  67. [75]

    D. Lian, D. Zhou, J. Feng, X. Wang, Scaling & shifting your features: A new baseline for efficient model tuning, NeurIPS 35 (2022) 109–123

    2022

  68. [76]

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, et al., Lora: Low-rank adaptation of large language models., ICLR 1 (2) (2022) 3. 19

    2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.