Pith. sign in

REVIEW 2 major objections 4 minor 31 references

Diffusion classifiers bind attributes more cleanly than CLIP-style models, yet lean harder on object size and background when they score prompts by reconstruction error.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 23:38 UTC pith:UH2XJOVZ

load-bearing objection Solid empirical diagnostic of diffusion-classifier decision rules: distinct bias profile (less attribute misbinding, more size-order and background shortcuts) with heatmaps that actually explain the scoring mechanism. the 2 major comments →

arxiv 2607.03831 v1 pith:UH2XJOVZ submitted 2026-07-04 cs.CV cs.AI

How Do Diffusion Classifiers Decide? A Bias-Centric Evaluation

Saba Fathi , Fardin Ayar , Maryam Abdolali , Ehsan Javanmardi , Manabu Tsukada , Mahdi Javanmardi This is my paper
classification cs.CV cs.AI
keywords diffusion classifierszero-shot classificationattribute bindingsize-order biasbackground dependencyreconstruction-error scoringbias evaluationtext-to-image models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Diffusion models can be turned into zero-shot classifiers by asking which text prompt lets the model reconstruct an image with the least noise-prediction error. This paper asks how that scoring rule actually decides. It introduces ASOB-Bench, three targeted probes of attribute binding, size-and-order shortcuts, and background dependence, and compares a Stable Diffusion classifier against an OpenCLIP baseline that shares the same text encoder. The classifier misbinds attributes less often across seven categories, yet collapses more sharply when the largest object or the background is a misleading cue. Pixel-level error heatmaps and cross-attention maps show why: the score simply aggregates error over the whole image, so large regions and surrounding context dominate. Because the same denoiser drives generation, the same failure modes are expected to appear when these models synthesize images.

Core claim

Diffusion classifiers exhibit a distinct bias profile from contrastive vision-language models: they are less prone to systematic attribute misbinding (lower CAB with higher two-object accuracy across seven categories) yet substantially more susceptible to size-order shortcuts (accuracy gap 0.37 versus 0.23 on ComCo) and background dependency (15–19-point drops versus roughly 2 points on ImageNet-B). Reconstruction-error heatmaps and U-Net cross-attention maps trace each bias to the pixel-aggregated reconstruction-error score itself.

What carries the argument

ASOB-Bench and the text-conditioned reconstruction-error score: the classifier picks the prompt that minimizes total noise-prediction error; heatmaps of that error and of cross-attention reveal where and why the score is hijacked by size, order, or background.

Load-bearing premise

That sharing only the text encoder with the OpenCLIP baseline is enough to isolate the effect of reconstruction-error scoring from differences in visual representation or training data.

What would settle it

Run the same three ASOB-Bench probes on a diffusion classifier and a contrastive model trained on identical image–text data and architecture; if the attribute, size-order, and background gaps reverse or disappear, the claim of a scoring-rule-driven distinct bias profile is false.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Headline accuracy alone will continue to hide systematic failures of diffusion classifiers on multi-object and out-of-context scenes.
  • Mitigations that re-weight error by object area or suppress background regions can be expected to shrink the size-order and background gaps.
  • The same single-pass diagnostics diagnose analogous attribute-leakage and background-dominance failures in text-to-image generation.
  • Practitioners can choose diffusion classifiers when attribute binding matters more than size or context robustness, and avoid them when the reverse is true.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the bias is intrinsic to total-error aggregation, any future scoring rule that normalizes error by region size or by foreground mask should systematically reduce the size-order and background gaps without retraining.
  • The color-driven cross-attention routing observed here is likely a shared bottleneck for any method that re-uses the same U-Net or DiT cross-attention for grounding or editing.
  • Extending ASOB-Bench to flow-matching and consistency models would test whether the bias profile is specific to noise-prediction objectives or common to all reconstruction-based classifiers.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper introduces ASOB-Bench, a three-axis bias evaluation of diffusion classifiers (Stable Diffusion 2, with an SD3 DiT check) that score candidate prompts by noise-prediction error. Along attribute binding (CAB and related metrics on natural/unnatural color plus five new attribute categories), size-order (ComCo Scenarios 1–2), and background dependency (ImageNet-B plus a no-background condition), it reports that diffusion classifiers are less prone to attribute misbinding than an OpenCLIP ViT-H/14 baseline sharing the same text encoder, yet substantially more susceptible to size-order shortcuts and background reliance. Reconstruction-error heatmaps, full reconstructions, and U-Net cross-attention maps are used to attribute each pattern to the pixel-aggregated scoring rule and to color-driven attention routing. The authors argue that the same diagnostics illuminate failure modes of the shared generative denoiser.

Significance. If the reported profile holds, the work supplies the first systematic, mechanism-level account of how reconstruction-error classifiers decide, filling a gap that has been extensively studied for contrastive VLMs but not for diffusion classifiers. The combination of controlled probes, unnatural-color and prompt-template controls, timestep-resolved ratios, and an SD3 replication is a concrete contribution; public code and data further raise the bar for follow-up work on robust diffusion-based models and on analogous generative failure modes.

major comments (2)
  1. §3.1 and the comparative claims in Tables 1–3 rest on the premise that OpenCLIP ViT-H/14, sharing only the SD2 text encoder, isolates scoring mechanism (contrastive vs. reconstruction-error) from residual differences in visual representation and training data. The within-model diagnostics (error-on-distractor, area-weighted residual error, object/background ratio <1) already establish the absolute mechanisms without the baseline, so the premise is not load-bearing for the existence of the profile; it is load-bearing for every quantitative “less/more” statement. A short additional control (e.g., another CLIP-scale model or a frozen-encoder ablation) or an explicit caveat that the comparison is mechanism-plus-representation would make the comparative language fully defensible.
  2. §3.3 and Appendix F: ComCo (12 780) and ImageNet-B (2 625 / 525) are evaluated on compute-driven subsamples without reported standard errors, confidence intervals, or a sensitivity check against the full sets. The directional gaps (size-order 0.3675 vs 0.2291; background drops of 15–19 points) are large, yet the absence of uncertainty quantification leaves open whether the reported magnitudes are stable. Adding bootstrap intervals or a note that the full sets were scored on a smaller timestep budget would close this gap without changing the experimental design.
minor comments (4)
  1. Table 1 mixes proportions (0–1) for color with percentages for the other categories; a uniform scale would improve readability.
  2. Appendix E notes that SD3 absolute accuracies are substantially lower than SD2; a one-sentence reminder in the main text that only relative trends are claimed would prevent misreading of Table 5.
  3. Figure 1 and the later heatmap figures would benefit from a brief caption note that error maps are single-step (or full-trajectory where stated) so readers do not conflate them with multi-step generation artifacts.
  4. A few citation keys appear with slight numbering drift between the main text and the appendix (e.g., ImageNet-B / BiRefNet); a final pass would align them.

Circularity Check

0 steps flagged

No circularity: empirical bias measurements and heatmaps against external benchmarks; nothing reduces by construction to fitted inputs or self-citation.

full rationale

ASOB-Bench is an empirical evaluation paper. Its load-bearing claims are measured quantities (CAB from Tang et al., Scenario-1/2 accuracy gap on ComCo from Abbasi et al., accuracy drops on ImageNet-B from Malik et al.) plus reconstruction-error and cross-attention visualizations that diagnose the pixel-aggregated scoring rule. None of these numbers is obtained by fitting a free parameter of the authors and then re-reporting it as a prediction; the CAB formula (Eq. 3), the two ComCo scenarios, and the six ImageNet-B conditions are taken from external sources and applied unchanged (or with only the authors’ added no-background segmentation). The OpenCLIP baseline shares the text encoder but is an independent contrastive model; the within-model diagnostics (error-on-distractor, area-weighted residual, object/background ratio <1) do not even require that baseline. The single self-citation to the authors’ compositionality survey is background, not a uniqueness theorem or ansatz that forces the reported bias profile. Appendix E’s SD3 DiT replication further shows the same qualitative trends without relying on any fitted quantity of the present paper. Consequently the derivation chain is self-contained against external data and contains no circular step.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central claims rest on standard diffusion-classifier scoring (Eq. 1–2), the CAB definition taken from Tang et al., the ComCo and ImageNet-B protocols, and the modeling choice that pixel-summed reconstruction error is the operative decision rule. No free parameters are fitted to produce the bias numbers; a few design choices (timestep ranges, prompt templates, subsample sizes) are fixed by hand or by prior recommendation and affect absolute accuracies but not the directional comparisons. ASOB-Bench itself is an evaluation construct, not a physical entity.

free parameters (3)
  • timestep sampling range for attribute tasks = 800-999
    Restricted to 800–999 following Jeong et al.; choice affects absolute accuracy and the depth of observed misbinding (Figs. 6–7) but is not fitted to the bias metrics themselves.
  • number of timesteps for scoring = 50
    Uniform sampling over 50 timesteps; computational convenience choice that determines the noise-prediction error estimate.
  • ComCo and ImageNet-B subsample sizes = 12780 ComCo / 2625 ImageNet-B used
    Compute-driven subsets (Appendix F) rather than full datasets; exact counts listed but selection criteria beyond 'high-quality classes' for ImageNet-B are not fully specified.
axioms (4)
  • domain assumption Noise-prediction error under a text prompt approximates -log p(z|y) and is therefore a valid zero-shot classification score (Eq. 1–2).
    Taken from Li et al. and Clark et al.; foundational to treating the diffusion model as a classifier.
  • domain assumption CAB = two-object* - two-object + 1/2 correctly quantifies directional attribute misbinding.
    Adopted unchanged from Tang et al. [24]; used throughout Table 1.
  • ad hoc to paper OpenCLIP ViT-H/14 sharing the SD2 text encoder isolates the effect of scoring mechanism rather than text encoding.
    Stated in §3.1; load-bearing for all 'distinct bias profile' comparisons.
  • domain assumption Pixel-aggregated reconstruction error is the operative decision variable that produces size and background shortcuts.
    Inferred from the scoring rule and confirmed by heatmaps; central to the mechanistic claims in §§4.2–4.3.
invented entities (1)
  • ASOB-Bench independent evidence
    purpose: Multi-dimensional bias evaluation suite (attribute binding + size-order + background) with new attribute categories and unnatural-color controls.
    Evaluation construct assembled from existing protocols plus new synthetic partitions; not a physical or theoretical entity requiring independent evidence beyond the reported measurements.

pith-pipeline@v1.1.0-grok45 · 21147 in / 3086 out tokens · 27339 ms · 2026-07-11T23:38:37.768964+00:00 · methodology

0 comments
read the original abstract

Diffusion models have recently been repurposed for zero-shot classification, giving rise to diffusion classifiers that identify the best-matching text prompt by minimizing the noise-prediction error. Despite their growing adoption, how these models make classification decisions remains poorly understood. We introduce ASOB-Bench, a bias evaluation for diffusion classifiers along three dimensions: Attribute binding, Size-Order bias, and Background dependency. These dimensions serve not as an exhaustive taxonomy but as targeted probes of how the text-conditioned reconstruction-error score reaches a decision. Such a perspective is well studied for discriminative vision-language models, yet remains overlooked for diffusion classifiers. Extending an existing framework with five new attribute categories on newly constructed datasets, we find diffusion classifiers are less prone to attribute misbinding than an OpenCLIP baseline; on the established ComCo benchmark they are substantially more susceptible to size-order shortcuts; and on ImageNet-B they suffer far larger accuracy drops, revealing heavy reliance on background over foreground cues. Reconstruction-error heatmaps and U-Net cross-attention visualizations expose the mechanism behind each bias. Because diffusion classifiers share the same denoiser as text-to-image models, these single-pass diagnostics also point toward analogous failure modes in generation. Overall, diffusion classifiers exhibit a distinct bias profile from vision-language models, offering guidance for building more robust diffusion-based models.

Figures

Figures reproduced from arXiv: 2607.03831 by Ehsan Javanmardi, Fardin Ayar, Mahdi Javanmardi, Manabu Tsukada, Maryam Abdolali, Saba Fathi.

Figure 1
Figure 1. Figure 1: Overview of ASOB-Bench across three bias dimensions: attribute binding, size-and-order, and background dependency. Reconstruction-error heatmaps and cross￾attention visualizations provide mechanistic insight into each bias. 040 learned iterative denoising. A striking recent insight is that these generative 040 041 models can be repurposed for discriminative tasks: by measuring how well a dif- 041 042 fusio… view at source ↗
Figure 2
Figure 2. Figure 2: One-step reconstruction error heatmaps for natural (top) and unnatural (bottom) color pairs. On natural images, error concentrates [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cross-attention heatmaps for “carrot” under natural (top) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Full reconstruction for a table-and-clock scene. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Error ratio (query/distractor) on natural images. In the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Error heatmaps for negative prompts of Scenario 1 (a) [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Reconstruction-error heatmaps across background conditions. Even in non-adversarial settings, error concentrates on the back [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Attention ratio (query/distractor) on natural images. [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Error ratio (changed object / remaining objects) for neg [PITH_FULL_IMAGE:figures/full_fig_p012_12.png] view at source ↗
Figure 15
Figure 15. Figure 15: Error ratio (object/background) for no-background con [PITH_FULL_IMAGE:figures/full_fig_p013_15.png] view at source ↗
Figure 2
Figure 2. Figure 2: Fig2: Dataset samples: (a) Figure 16. Dataset samples. Panels (a)–(c) show [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 6 linked inside Pith

  1. [1]

    Clip under the microscope: A fine-grained analysis of multi-object representation.2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9308–9317, 2025

    Reza Abbasi, Ali Nazari, Aminreza Sefid, Moham- madali Banayeeanzade, Mohammad Hossein Rohban, and Mahdieh Soleymani Baghshah. Clip under the microscope: A fine-grained analysis of multi-object representation.2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9308–9317, 2025. 3, 4, 13, 14

    2025

  2. [2]

    Compositionality in con- trastive vision-language models: A survey of methods and benchmarks

    Maryam Abdolali, Fardin Ayar, Roozbeh Tavanaei, Saba Fathi, Arshia Momtazi, Ehsan Javanmardi, Manabu Tsukada, and Mahdi Javanmardi. Compositionality in con- trastive vision-language models: A survey of methods and benchmarks. TechRxiv, 2026. 1, 3

    2026

  3. [3]

    Improving image genera- tion with better captions, 2023

    James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, Yunxin Jiao, and Aditya Ramesh. Improving image genera- tion with better captions, 2023. 4

    2023

  4. [4]

    Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, and Dominik Lorenz

    A. Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, and Dominik Lorenz. Sta- ble video diffusion: Scaling latent video diffusion models to large datasets.ArXiv, abs/2311.15127, 2023. 2

    Pith/arXiv arXiv 2023

  5. [5]

    Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis

    A. Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent dif- fusion models.2023 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 22563–22575,

    2023

  6. [6]

    Text-to-image diffusion mod- els are zero-shot classifiers.ArXiv, abs/2303.15233, 2023

    Kevin Clark and Priyank Jaini. Text-to-image diffusion mod- els are zero-shot classifiers.ArXiv, abs/2303.15233, 2023. 1, 2, 3

    Pith/arXiv arXiv 2023

  7. [7]

    Hawley, and Jordi Pons

    Zach Evans, CJ Carr, Josiah Taylor, Scott H. Hawley, and Jordi Pons. Fast timing-conditioned latent audio diffusion. ArXiv, abs/2402.04825, 2024. 2

    Pith/arXiv arXiv 2024

  8. [8]

    Geneval: An object-focused framework for evaluating text- to-image alignment

    Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text- to-image alignment. InAdvances in Neural Information Pro- cessing Systems (NeurIPS), Datasets and Benchmarks Track,

  9. [9]

    Woojung Han, Yeonkyung Lee, Chanyoung Kim, Kwanghyun Park, and Seong Jae Hwang. Spatial transport optimization by repositioning attention map for training-free text-to-image synthesis.2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18401–18410, 2025. 3

    2025

  10. [10]

    Narayana, Sugato Basu, William Yang Wang, and Xin Eric Wang

    Xuehai He, Weixi Feng, Tsu-Jui Fu, Varun Jampani, Ar- jun Reddy Akula, P. Narayana, Sugato Basu, William Yang Wang, and Xin Eric Wang. Discffusion: Discriminative diffusion models as few-shot vision and language learners. Trans. Mach. Learn. Res., 2024, 2024. 2

    2024

  11. [11]

    Jonathan Ho, Ajay Jain, and P. Abbeel. Denoising diffusion probabilistic models.ArXiv, abs/2006.11239, 2020. 1, 2

    Pith/arXiv arXiv 2006

  12. [12]

    Kaiyi Huang, Chengqi Duan, Kaiyue Sun, Enze Xie, Zhen- guo Li, and Xihui Liu. T2i-compbench++: An enhanced and comprehensive benchmark for compositional text-to-image generation.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 47:3563–3579, 2025. 2

    2025

  13. [13]

    Diffusion classifiers understand compositional- ity, but conditions apply.ArXiv, abs/2505.17955, 2025

    Yujin Jeong, Arnas Uselis, Seong Joon Oh, and Anna Rohrbach. Diffusion classifiers understand compositional- ity, but conditions apply.ArXiv, abs/2505.17955, 2025. 1, 2, 3, 6, 13

    arXiv 2025

  14. [14]

    V oleti, Christopher Joseph Pal, and Siva Reddy

    Benno Krojer, Elinor Poole-Dayan, Vikram S. V oleti, Christopher Joseph Pal, and Siva Reddy. Are diffusion mod- els vision-and-language reasoners? InAdvances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2, 3

    2023

  15. [15]

    Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak

    Alexander C. Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier.2023 IEEE/CVF International Confer- ence on Computer Vision (ICCV), pages 2206–2217, 2023. 1, 2, 3

    2023

  16. [16]

    Crepe: Can vision- language foundation models reason compositionally?2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10910–10921, 2023

    Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. Crepe: Can vision- language foundation models reason compositionally?2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10910–10921, 2023. 3

    2023

  17. [17]

    Khan, and Fahad Shahbaz Khan

    Hashmat Shadab Malik, Muhammad Huzaifa, Muzammal Naseer, Salman H. Khan, and Fahad Shahbaz Khan. Ob- jectcompose: Evaluating resilience of vision-based models on object-to-background compositional changes. InAsian Conference on Computer Vision, 2024. 3, 4, 13, 14

    2024

  18. [18]

    Dick, and Hidenori Tanaka

    Maya Okawa, Ekdeep Singh Lubana, Robert P. Dick, and Hidenori Tanaka. Compositional abilities emerge multiplica- tively: Exploring diffusion models on a synthetic task. InAd- vances in Neural Information Processing Systems (NeurIPS),

  19. [19]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InInternational Conference on Machine Learning, 2021. 1, 2

    2021

  20. [20]

    Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer

    Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, 2022. 2, 3

    2022

  21. [21]

    Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mah- davi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mah- davi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text- to-image diffusion models with deep language understand- ing.ArXiv, abs/2205.11487, 2022. 2

    Pith/arXiv arXiv 2022

  22. [22]

    Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions. 2020. 2

    2020

  23. [23]

    What the daam: Interpreting stable dif- fusion using cross attention

    Raphael Tang, Linqing Liu, Akshat Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Ture. What the daam: Interpreting stable dif- fusion using cross attention. InAnnual Meeting of the Asso- ciation for Computational Linguistics (ACL), 2023. 5, 11

    2023

  24. [24]

    When are lemons purple? the concept association bias of vision-language models

    Yingtian Tang, Yutaro Yamada, Yoyo Zhang, and Ilker Yildirim. When are lemons purple? the concept association bias of vision-language models. InConference on Empirical Methods in Natural Language Processing, 2023. 3, 4, 5, 13, 14

    2023

  25. [25]

    A good crepe needs more than just sugar: Investigating biases in compositional vision-language benchmarks, 2025

    Vishaal Udandarao, Mehdi Cherti, Shyamgopal Karthik, Je- nia Jitsev, Samuel Albanie, and Matthias Bethge. A good crepe needs more than just sugar: Investigating biases in compositional vision-language benchmarks, 2025. 1

    2025

  26. [26]

    Xingrui Wang, Wufei Ma, Tiezheng Zhang, Celso M de Melo, Jieneng Chen, and Alan Yuille. Spatial457: A di- agnostic benchmark for 6d spatial reasoning of large multi- modal models.2025 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 24669–24679,

    2025

  27. [27]

    Noise matters: Optimizing matching noise for diffusion classifiers

    Yanghao Wang and Long Chen. Noise matters: Optimizing matching noise for diffusion classifiers. InThe Thirty-ninth Annual Conference on Neural Information Processing Sys- tems, 2025. 2

    2025

  28. [28]

    Conceptmix: A compositional image generation benchmark with controllable difficulty.ArXiv, abs/2408.14339, 2024

    Xindi Wu, Dingli Yu, Yangsibo Huang, Olga Russakovsky, and Sanjeev Arora. Conceptmix: A compositional image generation benchmark with controllable difficulty.ArXiv, abs/2408.14339, 2024. 3

    Pith/arXiv arXiv 2024

  29. [29]

    Few-Shot Learner Parameterization by Diffusion Time-Steps

    Zhongqi Yue, Pan Zhou, Richang Hong, Hanwang Zhang, and Qianru Sun. Few-Shot Learner Parameterization by Diffusion Time-Steps . In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2

    2024

  30. [30]

    Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Y . Zou. When and why vision- language models behave like bags-of-words, and what to do about it? InInternational Conference on Learning Repre- sentations (ICLR), 2023. 3

    2023

  31. [31]

    in a complex scene

    Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral refer- ence for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research, 3:1–12, 2024. 4, 13, 14 Appendix This appendix provides additional details and analyses that support the findings reported in the main paper. Specifi- ...

    2024