Pith. sign in

REVIEW 4 major objections 6 minor 29 references

Mask Consistency Regularization in Object Removal

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Training diffusion inpainting with a mask-consistency loss — forcing the model to predict the same noise for original, dilated, and reshaped masks — substantially reduces mask hallucinations and shape-dependent fillings in object removal, o

desk verdict A plausible regularizer for diffusion inpainting, but the evidence as written is not enough to support the claimed gains — the paper admits experiments are still in progress and never compares against the same model without the consistency loss. read the letter →

arxiv 2509.10259 v1 pith:XJXVST7G submitted 2025-09-12 cs.CV

Hua Yuan , Jin Yuan , Yicheng Jiang , Yao Zhang , Xin Geng , Yong Rui This is my paper
classification cs.CV
keywords objectremovalimageinpaintingdiffusionmodelsconsistencyregularizationmaskhallucinationmask-shapebiasControlNetStableXL
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

This paper tries to establish that a single additional training loss — Mask Consistency Regularization (MCR) — can fix two failure modes in diffusion-based object removal: hallucinating content inside the masked region, and filling the region with an object that mirrors the mask's shape. The loss forces the model to predict the same noise for the original mask as for a dilated version and a reshaped version, making the inpainting process agnostic to mask geometry. On BrushBench and a self-constructed 300-image removal test set with ground-truth after-removal images, the method beats LaMa, BrushNet, and SDXL-I on most metrics, and ablations show both perturbation types contribute. If right, it gives a drop-in training objective for mask-conditioned diffusion inpainting that yields more contextually coherent and shape-agnostic results.

What carries the argument

The central object is the consistency loss L_cons (Eq. 5): the squared distance between the predicted noise under the original ControlNet conditioning z_O and under the dilated z_D and reshaped z_R conditionings. It is a regularizer added to the standard reconstruction loss L_rec. The two perturbations that generate z_D and z_R are morphological dilation with a square structuring element and mask reshape (minimum-enclosing rectangle or random-mask OR). The whole framework is built on a frozen Stable Diffusion XL with a trainable ControlNet, so the loss only updates ControlNet parameters.

What would settle it

Measure the output variance of a model trained with MCR when the same masked scene is presented with several masks that cover the same object but have different shapes (e.g., a tight silhouette, its bounding rectangle, and a random blob). If the outputs differ as much as a baseline's, or if the model still redraws an object whose shape matches the mask, the claim of reduced mask-shape bias fails. A specific check: take a car-in-scene example, use a circular mask that fully covers the car, and see whether the MCR model fills it with road/background rather than a car-shaped object.

Watch

Extended reading notes

Core claim

The central claim is that mask-geometry invariance is learnable through consistency regularization: given the same image and prompt, the denoising network should output the same noise prediction whether conditioned on the original mask, a k-pixel dilated mask, or a reshaped mask (rectangular or random) that covers the original region. Dilation pushes the model to borrow context beyond the strict boundary, while reshaping discourages it from treating the mask silhouette as a semantic cue. This invariance is enforced only during training, alongside the standard reconstruction loss; at inference, the original mask is used unchanged. The result is claimed to be a substantial reduction in mask ha

Load-bearing premise

The load-bearing premise is that forcing the same noise prediction for original, dilated, and reshaped masks is a valid training signal that does not conflict with the reconstruction objective — plus that the 300-sample test set's ground truths are reliable; if the consistency term merely encourages the model to ignore the mask, or the test set is biased, the reported performance gains would not generalize.

Editorial extensions

If this is right

  • MCR reduces hallucinations and mask-shape bias compared with LaMa, BrushNet, and SDXL-I on both the BrushBench and the 300-sample removal test set.
  • Ablation studies show that using both dilation and reshape together outperforms either perturbation alone, indicating the two strategies target complementary failure modes.
  • The method attains the best LPIPS and CMMD scores on both benchmarks, suggesting better perceptual quality and lower distribution-level discrepancy with the ground truth.
  • Because the base diffusion model is frozen, MCR can be added as a plug-and-play training objective for mask-conditioned inpainting without retraining the full generative backbone.

Reading between the lines

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

  • The same invariance principle could transfer to other mask-conditioned editing tasks such as object insertion or background replacement, where output should not depend on the precise mask silhouette; the paper does not test this.
  • Since the perturbed branches carry no reconstruction loss, they are supervised only by consistency to the original branch; a plausible failure mode the paper does not analyze is the model learning a trivial, mask-ignoring prediction that satisfies consistency but sacrifices localization.
  • A direct and untested implication is that MCR-trained models should produce lower output variance when the same scene is presented with different masks covering the same object; measuring this variance would isolate the mask-shape-bias reduction.
  • The 300-sample self-constructed test set is small and mixes real, physics-engine, and hand-drawn images; re-running on a larger benchmark with ground-truth removals would test how well the reported gains generalize.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes Mask Consistency Regularization (MCR), a training-time loss for mask-conditioned diffusion inpainting/object removal. The method trains a plug-and-play ControlNet on top of a frozen SDXL, adding two mask perturbations during training: dilation and reshaping (rectangular or random OR). The consistency loss L_cons in Eq. (5) forces the noise predictions under the original, dilated, and reshaped masks to agree, and the total loss is L = L_rec + lambda_cons L_cons. The authors claim this reduces mask hallucination and mask-shape bias, leading to improved object removal. They report comparisons against LaMa, BrushNet, and SDXL-I on BrushBench and a self-constructed 300-sample removal test set, plus ablations removing either perturbation.

Significance. If the central claim holds, MCR is a simple, drop-in regularizer for diffusion inpainting, and the paper would provide a useful practical contribution. The formulation is coherent, the use of frozen SDXL keeps training costs reasonable, and the comparisons include strong baselines. The paper does not ship code or data, and the experimental evidence is currently incomplete: the manuscript explicitly says experiments are in progress, there is no no-MCR baseline in the same framework, no error bars or significance tests are reported, and the self-constructed test set is not released. The core idea is worth pursuing, but the current claims substantially outrun the evidence.

major comments (4)
  1. [§3, opening sentence] The paper begins Section 3 with 'We are in the process of conducting extensive experiments', which is an explicit limitation flag. As written, the quantitative results are preliminary and cannot support the abstract/conclusion claims of significant improvement. The 'Experimental Setup' paragraph also omits the training dataset, number of training images/iterations, validation split, and how the 600-image BrushBench is partitioned between training and testing. These details are needed for reproducibility and to interpret Tables 1–4.
  2. [§3.2–3.3, Tables 1–4] The central claim that MCR itself improves object removal is unsupported without a no-consistency baseline. Tables 3 and 4 ablate only full MCR against single-perturbation variants (w/M D, w/M R) and do not include the same ControlNet+SDXL framework trained with lambda_cons=0. Consequently, the gains over LaMa/BrushNet/SDXL-I in Tables 1–2 could arise from the ControlNet backbone, training data, schedule, or other hyperparameters rather than from Eq. (5). The authors should train the identical architecture without L_cons and report it in all comparisons, with multiple seeds.
  3. [§2.3, Eq. (5)] There is a correctness risk in the formulation that should be addressed explicitly. The perturbed branches have no reconstruction loss; their only supervision is consistency with the original branch's noise prediction. A model can satisfy this consistency by becoming mask-agnostic and ignoring precise mask boundaries, which would degrade localization on the exact mask at inference. The claim in Section 2 that MCR 'decouples the generation process from spurious dependencies on mask geometry' is not the same as decoupling from mask location. Please provide evidence that exact-mask localization is preserved, e.g., by measuring boundary accuracy or comparing inference with exact vs. perturbed masks.
  4. [§3.1 Benchmark and §3.2 Results] The 300-sample removal test is self-constructed and not released; there is no breakdown of how many images come from real-world, physics-engine, or hand-drawn sources, nor how the ground-truth was produced in each case. No error bars, confidence intervals, significance tests, or number of seeds are reported anywhere in the paper. On a 300-sample set, FID/CMMD are high-variance; the word 'significantly' in the abstract and conclusion is therefore not statistically supported. Please report standard deviations/confidence intervals and, if possible, release the test set so results can be reproduced and compared.
minor comments (6)
  1. [§2.1, Eq. (1)] The noisy sample x_t is defined, but Eq. (1) writes epsilon_theta as a function of x_0 rather than x_t (or a latent z_t). Clarify the actual conditioning arguments used in the implementation.
  2. [Figure 1 caption] The caption uses z0, z1, z2 while the text uses z_O, z_D, z_R. Unify the notation.
  3. [§2.2 Mask Reshape] The random reshape strategy is said to follow LaMa [11], but the exact generation procedure, coverage distribution, and how often random masks are sampled per training step are not specified. This matters for reproducibility.
  4. [Tables 3–4] 'w/M D' and 'w/M R' are ambiguous; spell out 'with only mask dilation' and 'with only mask reshape' in the captions or text.
  5. [References] The author string in reference [11] appears garbled ('Kiyoharu Aizawa Park'). Please verify all bibliographic entries.
  6. [General] Metric values are reported with inconsistent precision across tables (e.g., 60.89 vs 30.348; 0.8661 vs 0.8666). Use consistent precision and consider adding standard deviations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MCR is a self-supervised training regularizer evaluated against external baselines; its central claim does not reduce to its own objective by construction.

full rationale

The paper's central claim is that adding Mask Consistency Regularization (Eq. 5) to the reconstruction loss (Eq. 1) reduces mask hallucination and mask-shape bias in object removal. This is an empirical claim about training dynamics, not a derivation that presupposes the result. The consistency loss is an auxiliary term, not a fitted re-expression of the evaluation metrics (FID, PSNR, SSIM, LPIPS, CMMD). No parameter is fit to a subset of the evaluation data and then 'predicted' on a closely related quantity. The mask perturbations in Eqs. 2–4 are straightforward morphological operations, and L_cons in Eq. 5 is a regularizer; it does not by definition guarantee improved inpainting, since it could in principle be satisfied by a mask-agnostic model that still fails reconstruction. The paper compares against external baselines (LaMa, BrushNet, SDXL-I) on an external benchmark (BrushBench) and a self-constructed test set that provides ground truth. There are no load-bearing self-citations and no imported uniqueness claims. The explicit limitation 'We are in the process of conducting extensive experiments' (Section 3) and the absence of a no-consistency (λ_cons=0) baseline are evidential weaknesses, not circularity. The reader's concern about a vacuous solution to L_cons is a training-failure risk, not a logical circularity in the method's formulation.

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

The method introduces no new physical entities. It relies on standard pretrained components (SDXL, ControlNet), a small set of hand-chosen hyperparameters, and a few domain assumptions about the validity of mask invariance and the reliability of the self-constructed benchmark. The most consequential free parameter is the consistency weight lambda_cons, and the unstated dilation kernel size is a gap in reproducibility.

free parameters (5)
  • lambda_cons = 2
    Consistency loss weight in Eq. 6, chosen by hand; no sensitivity analysis is reported.
  • dilation kernel size k
    Defines M_dil^k in Eq. 2 but no value or sensitivity is reported in the paper.
  • reshape probability = 0.5
    Equal probability of rectangular vs random reshape, chosen without justification in Section 2.2.
  • learning rate = 5e-5
    Training hyperparameter for the ControlNet module, stated in Section 3.1.
  • batch size = 2
    Training hyperparameter, small due to GPU memory limits; stated in Section 3.1.
assumptions (5)
  • domain assumption The noise prediction of the diffusion model should be invariant to mask perturbations; consistency across different masks is a valid training signal.
    Introduced in Eq. 5; the entire method rests on this assumption, which is not theoretically justified.
  • domain assumption Frozen SDXL with a trainable ControlNet can be steered to object removal via mask conditioning.
    Adopted from prior work [15, 16]; the paper relies on these components working as described.
  • domain assumption The self-constructed 300-sample removal test provides reliable ground truth for object removal.
    Section 3.1; the dataset is not released and its construction is only briefly described.
  • domain assumption Morphological dilation and rectangular/random reshapes are valid perturbations that preserve the removal task.
    Section 2.2; the perturbations change the mask, and it is assumed the model should produce the same output.
  • domain assumption BrushBench masks and captions are suitable training data for this setup.
    Section 3.1; only 600 images are used, which is small for diffusion training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mask Consistency Regularization in Object Removal." pith.science (2026). https://pith.science/paper/XJXVST7G

@misc{pith2026250910259,
  author       = {Pith},
  title        = {Pith review of: Mask Consistency Regularization in Object Removal},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XJXVST7G}},
  note         = {Machine review of arXiv:2509.10259}
}
read the original abstract

Object removal, a challenging task within image inpainting, involves seamlessly filling the removed region with content that matches the surrounding context. Despite advancements in diffusion models, current methods still face two critical challenges. The first is mask hallucination, where the model generates irrelevant or spurious content inside the masked region, and the second is mask-shape bias, where the model fills the masked area with an object that mimics the mask's shape rather than surrounding content. To address these issues, we propose Mask Consistency Regularization (MCR), a novel training strategy designed specifically for object removal tasks. During training, our approach introduces two mask perturbations: dilation and reshape, enforcing consistency between the outputs of these perturbed branches and the original mask. The dilated masks help align the model's output with the surrounding content, while reshaped masks encourage the model to break the mask-shape bias. This combination of strategies enables MCR to produce more robust and contextually coherent inpainting results. Our experiments demonstrate that MCR significantly reduces hallucinations and mask-shape bias, leading to improved performance in object removal.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 6 linked inside Pith

  1. [1]

    INTRODUCTION Image inpainting, particularly for object removal, aims to recon- struct missing regions of an image with visually realistic and se- mantically coherent content [1, 2]. Early deep learning methods adopted convolutional encoder–decoder frameworks and generative adversarial networks (GANs) that leveraged attention mechanisms or partial convolut...

    arXiv 2025

  2. [2]

    METHOD Our approach frames diffusion-based object removal as a consistency- driven learning problem. Instead of relying solely on reconstruction objectives, we systematically introduce perturbations to the input masks and require the model to produce stable and coherent pre- dictions across these variations, as illustrated in Fig. 1. This design explicitl...

  3. [3]

    In this section, we outline our experimental setup and show experimental results

    EXPERIMENTS We are in the process of conducting extensive experiments to evalu- ate the proposed method. In this section, we outline our experimental setup and show experimental results. We compared our method with previous approaches on multiple datasets using several metrics, and these results reflect the effectiveness of the proposed method. Ad- dition...

    arXiv

  4. [4]

    By introducing mask dila- tion and mask reshaping strategies, MCR encourages the model to generate more contextually coherent and shape-agnostic inpainting results

    CONCLUSION In this paper, we proposed the Mask Consistency Regularization framework to address the challenges of mask hallucination and mask-shape bias in object removal. By introducing mask dila- tion and mask reshaping strategies, MCR encourages the model to generate more contextually coherent and shape-agnostic inpainting results. Experimental results ...

  5. [5]

    Context encoders: Feature learn- ing by inpainting,

    Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros, “Context encoders: Feature learn- ing by inpainting,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 2536– 2544

    2016

  6. [6]

    Kris-bench: Benchmarking next-level intelli- gent image editing models,

    Yongliang Wu, Zonghui Li, Xinting Hu, Xinyu Ye, Xianfang Zeng, Gang Yu, Wenbo Zhu, Bernt Schiele, Ming-Hsuan Yang, and Xu Yang, “Kris-bench: Benchmarking next-level intelli- gent image editing models,”arXiv preprint arXiv:2505.16707, 2025

    arXiv 2025

  7. [7]

    Generative image inpainting with contex- tual attention,

    Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang, “Generative image inpainting with contex- tual attention,” inProceedings of the IEEE conference on com- puter vision and pattern recognition, 2018, pp. 5505–5514

    2018

  8. [8]

    Image inpainting for ir- regular holes using partial convolutions,

    Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro, “Image inpainting for ir- regular holes using partial convolutions,” inECCV, 2018, pp. 85–100

    2018

Show all 29 references
  1. [9]

    Repaint: Inpainting us- ing denoising diffusion probabilistic models,

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool, “Repaint: Inpainting us- ing denoising diffusion probabilistic models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11461–11471

    2022

  2. [10]

    High-resolution image synthesis with latent diffusion models,

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer, “High-resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10684–10695

    2022

  3. [11]

    Dynamic prompt optimizing for text-to- image generation,

    Wenyi Mo, Tianyu Zhang, Yalong Bai, Bing Su, Ji-Rong Wen, and Qing Yang, “Dynamic prompt optimizing for text-to- image generation,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2024, pp. 26627–26636

    2024

  4. [12]

    Decoupling training-free guided diffu- sion by admm,

    Youyuan Zhang, Zehua Liu, Zenan Li, Zhaoyu Li, James J Clark, and Xujie Si, “Decoupling training-free guided diffu- sion by admm,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 23292–23302

    2025

  5. [13]

    Imagen editor: Text-guided image editing with diffusion models,

    Hila Chefer, Ron Mokady, Rinon Gal, Oron Ashual, Tyler Hayes, Ido Mosseri, Ariel Shamir, Yonatan Belinkov, and Ohad Fried, “Imagen editor: Text-guided image editing with diffusion models,” inProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, 2023, pp. 4085–4096

    2023

  6. [14]

    In- structpix2pix: Learning to follow image editing instructions,

    Tim Brooks, Aleksander Holynski, and Alexei A. Efros, “In- structpix2pix: Learning to follow image editing instructions,” inProceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, 2023, pp. 18392–18402

    2023

  7. [15]

    Resolution-robust large mask inpainting with fourier convolutions,

    Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Alena Remizova, Arsenii Ashukha, Alexander Silvestrov, Xin Kong, Hiromu Goka, Kiyoharu Aizawa Park, and Victor Lem- pitsky, “Resolution-robust large mask inpainting with fourier convolutions,” inProceedings of the IEEE/CVF W...

    2022

  8. [16]

    Smartbrush: Text and shape guided object inpainting with dif- fusion models,

    Yifan Liu, Chuanxia Yang, Qiang Wu, and Chen Change Loy, “Smartbrush: Text and shape guided object inpainting with dif- fusion models,”ACM Transactions on Graphics, vol. 42, no. 4, pp. 1–12, 2023

    2023

  9. [17]

    Fixmatch: Simplifying semi- supervised learning with consistency and confidence,

    Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li, “Fixmatch: Simplifying semi- supervised learning with consistency and confidence,”Ad- vances in neural information processing systems, ...

    2020

  10. [18]

    Pierre Soille,Morphological Image Analysis: Principles and Applications, Springer, 1999

    1999

  11. [19]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis,

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M¨uller, Joe Penna, and Robin Rombach, “Sdxl: Improving latent diffusion models for high-resolution image synthesis,”arXiv preprint arXiv:2307.01952, 2023

    2023 arXiv

  12. [20]

    Adding conditional control to text-to-image diffusion models,

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala, “Adding conditional control to text-to-image diffusion models,” inPro- ceedings of the IEEE/CVF international conference on com- puter vision, 2023, pp. 3836–3847

    2023

  13. [21]

    Rafael C Gonzalez,Digital image processing, Pearson educa- tion india, 2009

    2009

  14. [22]

    Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion,

    Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu, “Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 150–168

    2024

  15. [23]

    Omnipaint: Mastering object-oriented editing via disentangled insertion-removal inpainting,

    Yongsheng Yu, Ziyun Zeng, Haitian Zheng, and Jiebo Luo, “Omnipaint: Mastering object-oriented editing via disentangled insertion-removal inpainting,”arXiv preprint arXiv:2503.08677, 2025

    2025 arXiv

  16. [24]

    Omnieraser: Remove objects and their effects in images with paired video-frame data,

    Runpu Wei, Zijin Yin, Shuo Zhang, Lanxiang Zhou, Xueyi Wang, Chao Ban, Tianwei Cao, Hao Sun, Zhongjiang He, Kongming Liang, et al., “Omnieraser: Remove objects and their effects in images with paired video-frame data,”arXiv preprint arXiv:2501.07397, 2025

    2025 arXiv

  17. [25]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bern- hard Nessler, and Sepp Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” Advances in neural information processing systems, vol. 30, 2017

    2017

  18. [26]

    Image quality assessment: from error visibility to structural similarity,

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE transactions on image processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  19. [27]

    The unreasonable effectiveness of deep fea- tures as a perceptual metric,

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang, “The unreasonable effectiveness of deep fea- tures as a perceptual metric,” inProceedings of the IEEE con- ference on computer vision and pattern recognition, 2018, pp. 586–595

    2018

  20. [28]

    Re- thinking fid: Towards a better evaluation metric for image gen- eration,

    Sadeep Jayasumana, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv Kumar, “Re- thinking fid: Towards a better evaluation metric for image gen- eration,” inProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2024,...

    2024

  21. [29]

    Stable diffusion xl 1.0 inpainting 0.1,

    Diffusers team, “Stable diffusion xl 1.0 inpainting 0.1,”https://huggingface.co/diffusers/ stable-diffusion-xl-1.0-inpainting-0.1, 2023, Model card; initialized from textttstable-diffusion-xl-base-1.0, trained with 40k steps at 1024×1024 resolution (arXiv:2112.10752)

    2023 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.