pith. sign in

arxiv: 2606.00377 · v1 · pith:MHX2LFGBnew · submitted 2026-05-29 · 💻 cs.CV

Score-Control for Hallucination Reduction in Diffusion Models

Mahesh Bhosale , Naresh Kumar Devulapally , Abdul Wasi , Chau Pham , Vishnu Suresh Lokhande , David Doermann This is my paper

Pith reviewed 2026-06-28 22:42 UTC · model grok-4.3

classification 💻 cs.CV
keywords diffusion modelshallucinationsscore functionlipschitz constantvariance-guided modulationimage generationgenerative models
0
0 comments X

The pith

Variance-guided modulation of the score Jacobian reduces hallucinations in diffusion models by lowering the Lipschitz constant of the learned score.

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

Diffusion models generate hallucinations when the learned score function is too smooth, placing excess probability mass on implausible outputs outside the true data distribution. The paper links this probability mass directly to the Lipschitz constant of the score and confirms the link empirically on image generation tasks. It introduces Variance-Guided Score Modulation to control the score Jacobian, which reduces smoothness and moves the learned score closer to the ground-truth score. Experiments show the method cuts hallucinations by up to 25 percent on synthetic and real datasets while preserving fidelity and diversity. The work also releases two new benchmarks with extreme semantic variation for systematic hallucination testing.

Core claim

Hallucinations arise because the learned score function is smoother than the ground-truth score; the probability mass of such hallucinations scales with the Lipschitz constant of the learned score. Variance-Guided Score Modulation controls the score Jacobian during sampling to reduce this constant, thereby decreasing hallucination mass without new artifacts.

What carries the argument

Variance-Guided Score Modulation (VSM), a strategy that adjusts the score estimate using local variance to control the Jacobian and lower the Lipschitz constant of the score function.

If this is right

  • Hallucination rate scales directly with the Lipschitz constant of the score function.
  • Reducing score smoothness improves approximation to the true data distribution.
  • New benchmarks with extreme semantic variation enable repeatable measurement of hallucination reduction.
  • The modulation can be applied during sampling without retraining the underlying model.

Where Pith is reading between the lines

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

  • The same smoothness-hallucination link may appear in diffusion models for modalities other than images.
  • Controlling the score Jacobian offers a route to reliability improvements in safety-critical image generation uses.
  • The density-based view of hallucinations could be used to diagnose other failure modes such as mode collapse.

Load-bearing premise

Lowering the Lipschitz constant of the learned score will reduce hallucination probability mass without creating new artifacts or lowering sample quality.

What would settle it

A controlled experiment in which a diffusion model with measurably lower Lipschitz constant after modulation still produces the same hallucination rate as the baseline, or produces visibly degraded samples.

Figures

Figures reproduced from arXiv: 2606.00377 by Abdul Wasi, Chau Pham, David Doermann, Mahesh Bhosale, Naresh Kumar Devulapally, Vishnu Suresh Lokhande.

Figure 1
Figure 1. Figure 1: Motivation: Score smoothing causes hallucinations on mixture of 1D Gaussians with means [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative examples of corrected hallucinations with VSM. Each pair shows hallucinated generations (red) and [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Categorization of generated chessboards into invalid (hallucinated), memorized (seen in train), and generalized (novel) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Increasing 𝜌 decreases hallucinations until it start increasing it back because diffusion loss gets excessively down-weighted causing suboptimal results. H% for 1D and 2D are scaled by 103 and 101 respectively. Schedule C-FID↓ FLD↓ H%↓ 𝜂 (𝑡) = 𝜌 (1 − 𝛼¯𝑡 ) 17.18 19.30 7.83 𝜂 (𝑡) = 𝜌/(1 − 𝛼¯𝑡 ) 11.05 7.61 5.00 𝜂 (𝑡) = 𝜌/ √ 1 − 𝛼¯𝑡 3.91 6.99 3.50 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ablation of time-dependent scaling schedules [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Iterative Training while appending Non￾Hallucinated Images to Ptrain We propose a way that drives the hallucination rate toward zero [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Increase in Score difference Δ𝑠 positively correlates with Hallucinations on Hands dataset. Estimating sGT: For 1D and 2D datasets, we have closed form PDFs with fixed parametrs. Therefore, ground truth score can be obtained from closed form PDF: 𝑆𝐺𝑇 (𝑥𝑡 ) = Í𝑀 𝑚=1 − 𝑥𝑡 −𝜇𝑚 𝜎2 exp − (𝑥𝑡 −𝜇𝑚) 2 2𝜎2  Í𝑀 𝑚=1 exp − (𝑥𝑡 −𝜇𝑚) 2 2𝜎2  . For image datasets, we do not have access to the groundtruth posterior induc… view at source ↗
Figure 8
Figure 8. Figure 8: Generated images marked Hallucinated for the [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 11
Figure 11. Figure 11: We observe that our method corrects the deformed [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 10
Figure 10. Figure 10: Example samples from the proposed ChessImages dataset. Top: a generated chessboard configuration. Bottom: its corresponding Forsyth–Edwards Notation (FEN) string, providing an exact symbolic representation of the board state. frequently produces deformed objects and incompletely denoised samples, resulting in images that deviate from the training data distribution. In contrast, our method mitigates these … view at source ↗
read the original abstract

Diffusion models have emerged as the backbone of modern generative AI, powering advances in vision, language, audio and other modalities. Despite their success, they suffer from hallucinations, implausible samples that lie outside the support of true data distribution, which degrade reliability and trust. In this work, we first empirically confirm previously proposed hypothesis that score smoothness causes hallucinations in Image Generation diffusion models and provide a density-based perspective. We further formalize this notion by linking the hallucinations probability mass to lipschitz constant of the learned score function. Motivated by this, we introduce a Variance-Guided Score Modulation (VSM) strategy that controls the score Jacobian, in turn reducing score smoothness and better approximating the ground truth score that decreases hallucinations. Empirical results on synthetic and real-world datasets demonstrate that our approach reduces hallucinations (up to ~25%) while maintaining high fidelity and diversity, providing a principled step toward more reliable diffusion-based image generation. We also propose two benchmark datasets with extreme semantic variation for systematic hallucination evaluation. Code and Datasets are publicly available at https://github.com/bhosalems/VSM.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper empirically confirms the hypothesis that score smoothness causes hallucinations in diffusion-based image generation and provides a density-based perspective. It formalizes the connection by linking hallucination probability mass to the Lipschitz constant of the learned score function. Motivated by this, it introduces Variance-Guided Score Modulation (VSM) to control the score Jacobian, thereby reducing smoothness and better approximating the ground-truth score. Experiments on synthetic and real-world datasets report up to ~25% hallucination reduction while preserving fidelity and diversity; two new benchmark datasets with extreme semantic variation are also proposed, with code and data released publicly.

Significance. If the formalization yields a quantitative bound and VSM demonstrably reduces the relevant Lipschitz constant without introducing new artifacts, the work would provide a principled, controllable mechanism for improving reliability of diffusion models. Public code and datasets are a clear strength that supports reproducibility and further testing of the empirical claims.

major comments (3)
  1. [Formalization (abstract and motivating sections)] The abstract states that the hallucinations probability mass is linked to the Lipschitz constant of the learned score, but provides neither an equation nor a proof sketch establishing a quantitative relationship (e.g., an inequality bounding P(hallucination) by Lip(s) or a related quantity). Without this, it is unclear whether the link is tight enough for VSM to guarantee reduction rather than merely correlate with it.
  2. [VSM strategy description] The variance-guided modulation mechanism is described only at high level; it is not shown whether the Jacobian control is uniform, local, or could increase the Lipschitz constant in other regions of the score function, potentially trading one form of hallucination for another or altering out-of-support mass.
  3. [Empirical evaluation and benchmarks] Empirical results claim up to ~25% reduction, yet the abstract supplies no information on hallucination measurement protocol, data exclusion rules, statistical error bars, or how the new benchmarks avoid circularity with the evaluation metric.
minor comments (2)
  1. [Abstract] The abstract refers to a 'previously proposed hypothesis' without a citation; adding the reference would clarify the novelty of the empirical confirmation.
  2. [Method section] Notation for the score function, Jacobian, and variance guidance should be introduced consistently before the VSM definition to aid readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive report. The comments highlight opportunities to improve clarity in the formalization, the description of VSM, and the presentation of empirical details. We address each point below and commit to revisions that strengthen the manuscript without altering its core claims.

read point-by-point responses

  1. Referee: [Formalization (abstract and motivating sections)] The abstract states that the hallucinations probability mass is linked to the Lipschitz constant of the learned score, but provides neither an equation nor a proof sketch establishing a quantitative relationship (e.g., an inequality bounding P(hallucination) by Lip(s) or a related quantity). Without this, it is unclear whether the link is tight enough for VSM to guarantee reduction rather than merely correlate with it.

    Authors: We agree that the abstract and motivating sections would benefit from an explicit statement of the quantitative link. The density-based perspective in the paper connects hallucination mass to score smoothness via the Lipschitz constant, but we will revise to include the key inequality (P(hallucination) bounded in terms of Lip(s)) together with a short derivation sketch in both the abstract and Section 3. This will make the motivation for controlling the Jacobian via VSM more precise. revision: yes

  2. Referee: [VSM strategy description] The variance-guided modulation mechanism is described only at high level; it is not shown whether the Jacobian control is uniform, local, or could increase the Lipschitz constant in other regions of the score function, potentially trading one form of hallucination for another or altering out-of-support mass.

    Authors: The VSM modulation is constructed to act locally by scaling the score update according to per-sample variance estimates, which targets regions of high uncertainty without uniform application across the entire function. We will expand the method section with the explicit update rule, a brief analysis of the resulting local Lipschitz behavior, and additional diagnostics confirming that the global Lip constant does not increase and that out-of-support mass is not inflated. If the current experiments do not fully address this, we will add targeted ablations. revision: partial

  3. Referee: [Empirical evaluation and benchmarks] Empirical results claim up to ~25% reduction, yet the abstract supplies no information on hallucination measurement protocol, data exclusion rules, statistical error bars, or how the new benchmarks avoid circularity with the evaluation metric.

    Authors: The measurement protocol (including semantic outlier detection, exclusion criteria, and error bars from multiple seeds) is fully specified in the experimental section, and the new benchmarks use held-out semantic categories disjoint from training data to avoid circularity. We will revise the abstract to include a concise statement of the evaluation protocol and benchmark construction. This addresses the presentation concern while leaving the underlying results unchanged. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper empirically confirms a prior external hypothesis on score smoothness and hallucinations, asserts a formal link between hallucination probability mass and the Lipschitz constant of the score function, and introduces a new VSM modulation strategy motivated by that link. No equations, self-citations, or fitted parameters are shown reducing the central claims to inputs by construction. The derivation chain rests on new empirical validation and a proposed control mechanism rather than self-referential definitions or renamings, making the work self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that score smoothness is a primary driver of hallucinations and that variance-guided Jacobian control approximates the ground-truth score; no free parameters or invented entities are mentioned in the abstract.

axioms (1)
  • domain assumption Score smoothness (Lipschitz constant of learned score) causes hallucinations in diffusion models
    Stated as an empirically confirmed hypothesis that motivates the entire approach.

pith-pipeline@v0.9.1-grok · 5737 in / 1037 out tokens · 17053 ms · 2026-06-28T22:42:06.128141+00:00 · methodology

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

44 extracted references · 1 canonical work pages

  1. [1]

    2025.Adobe Firefly: The next evolution of creative AI is here

    Adobe. 2025.Adobe Firefly: The next evolution of creative AI is here. https://blog.adobe.com/en/publish/2025/04/24/adobe-firefly-next- evolution-creative-ai-is-here Adobe reports 22B+ Firefly-generated assets world- wide

    2025

  2. [2]

    Mahmoud Afifi. 2019. 11K Hands: Gender recognition and biometric identifica- tion using a large dataset of hand images.Multimedia Tools Appl.78, 15 (Aug. 2019), 20835–20854. doi:10.1007/s11042-019-7424-8

    work page doi:10.1007/s11042-019-7424-8 2019

  3. [3]

    Lipton, and J

    Sumukh K Aithal, Pratyush Maini, Zachary C. Lipton, and J. Zico Kolter. 2024. Understanding Hallucinations in Diffusion Models through Mode Interpolation. InAdvances in Neural Information Processing Systems, A. Globerson, L. Mackey, Figure 11: We observe that our method corrects the deformed objects, incompletely denoised images on the ImageNet-1K dataset...

    2024

  4. [4]

    Ahmed Alaa, Boris Van Breugel, Evgeny S Saveliev, and Mihaela Van Der Schaar

  5. [5]

    InInternational conference on machine learning

    How faithful is your synthetic data? sample-level metrics for evaluating and auditing generative models. InInternational conference on machine learning. PMLR, 290–306

  6. [6]

    Nick Alger, Tucker Hartland, Noemi Petra, and Omar Ghattas. 2024. Point spread function approximation of high-rank Hessians with locally supported nonnegative integral kernels.SIAM Journal on Scientific Computing46, 3 (2024), A1658–A1689

    2024

  7. [7]

    Mahesh Bhosale, Abdul Wasi, Yuanhao Zhai, Yunjie Tian, Samuel Border, Nan Xi, Pinaki Sarder, Junsong Yuan, David Doermann, and Xuan Gong. 2025. PathDiff: Histopathology Image Synthesis with Unpaired Text and Mask Conditions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 22415–22424

    2025

  8. [8]

    Bradley C. A. Brown, Anthony L. Caterini, Brendan Leigh Ross, Jesse C. Cresswell, and Gabriel Loaiza-Ganem. 2023. Verifying the Union of Manifolds Hypothesis for Image Data. arXiv:2207.02862 [stat.ML] https://arxiv.org/abs/2207.02862

    arXiv 2023

  9. [9]

    Zhengdao Chen. 2025. On the Interpolation Effect of Score Smoothing.arXiv preprint arXiv:2502.19499(2025)

    Pith/arXiv arXiv 2025

  10. [10]

    Naresh Kumar Devulapally, Mingzhen Huang, Vishal Asnani, Shruti Agarwal, Siwei Lyu, and Vishnu Suresh Lokhande. 2025. Your Text Encoder Can Be An Object-Level Watermarking Controller. arXiv:2503.11945 [cs.CV] https: //arxiv.org/abs/2503.11945

    arXiv 2025

  11. [11]

    Gerald B. Folland. 1999.Real Analysis: Modern Techniques and Their Applications (2nd ed.). John Wiley & Sons, New York

    1999

  12. [12]

    Zhiye Guo, Jian Liu, Yanli Wang, Mengrui Chen, Duolin Wang, Dong Xu, and Jianlin Cheng. 2024. Diffusion models in bioinformatics and computational biology.Nature reviews bioengineering2, 2 (2024), 136–154

    2024

  13. [13]

    Susan Hao, Piyush Kumar, Sarah Laszlo, Shivani Poddar, Bhaktipriya Radharapu, and Renee Shelby. 2023. Safety and fairness for content moderation in generative models.arXiv preprint arXiv:2306.06135(2023)

    arXiv 2023

  14. [14]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in neural information processing systems33 (2020), 6840–6851

    2020

  15. [15]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models.NeurIPS(2020). https://arxiv.org/abs/2006.11239 Score-Control for Hallucination Reduction in Diffusion Models

    Pith/arXiv arXiv 2020

  16. [16]

    Chawla, Jian Pei, Jianfeng Gao, Michael Backes, Philip S

    Yue Huang, Chujie Gao, Siyuan Wu, Haoran Wang, Xiangqi Wang, Yujun Zhou, Yanbo Wang, Jiayi Ye, Jiawen Shi, Qihui Zhang, Yuan Li, Han Bao, Zhaoyi Liu, Tianrui Guan, Dongping Chen, Ruoxi Chen, Kehan Guo, Andy Zou, Bryan Hooi Kuen-Yew, Caiming Xiong, Elias Stengel-Eskin, Hongyang Zhang, Hongzhi Yin, Huan Zhang, Huaxiu Yao, Jaehong Yoon, Jieyu Zhang, Kai Shu,...

    Pith/arXiv arXiv 2025

  17. [17]

    Marco Jiralerspong, Joey Bose, Ian Gemp, Chongli Qin, Yoram Bachrach, and Gauthier Gidel. 2023. Feature likelihood divergence: evaluating the generalization of generative models using samples.Advances in Neural Information Processing Systems36 (2023), 33095–33119

    2023

  18. [18]

    Seunghoi Kim, Chen Jin, Tom Diethe, Matteo Figini, Henry F. J. Tregidgo, Asher Mullokandov, Philip Teare, and Daniel C. Alexander. 2024. Tack- ling Structural Hallucination in Image Translation with Local Diffusion. arXiv:2404.05980 [cs.CV] https://arxiv.org/abs/2404.05980

    arXiv 2024

  19. [19]

    Saksham Singh Kushwaha, Jianbo Ma, Mark RP Thomas, Yapeng Tian, and Avery Bruni. 2025. Diff-SAGe: End-to-End Spatial Audio Generation Using Diffusion Models. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 1–5

    2025

  20. [20]

    Tuomas Kynkäänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. 2019. Improved Precision and Recall Metric for Assessing Generative Models. arXiv:1904.06991 [stat.ML] https://arxiv.org/abs/1904.06991

    arXiv 2019

  21. [21]

    Lecun, L

    Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. 1998. Gradient-based learning applied to document recognition.Proc. IEEE86, 11 (1998), 2278–2324. doi:10. 1109/5.726791

    1998

  22. [22]

    Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. 2022. Diffusion-lm improves controllable text generation.Advances in neural information processing systems35 (2022), 4328–4343

    2022

  23. [23]

    Rui Lu, Runzhe Wang, Kaifeng Lyu, Xitai Jiang, Gao Huang, and Mengdi Wang

  24. [24]

    InThe Thirteenth International Conference on Learning Represen- tations

    Towards understanding text hallucination of diffusion models via local generation bias. InThe Thirteenth International Conference on Learning Represen- tations

  25. [25]

    Shweta Mahajan, Tanzila Rahman, Kwang Moo Yi, and Leonid Sigal. 2024. Prompting hard or hardly prompting: Prompt inversion for text-to-image diffu- sion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6808–6817

    2024

  26. [26]

    Chenlin Meng, Yang Song, Wenzhe Li, and Stefano Ermon. 2021. Estimating High Order Gradients of the Data Distribution by Denoising. InNeurIPS. 12477–12488

    2021

  27. [27]

    Alexander Quinn Nichol and Prafulla Dhariwal. 2021. Improved denoising diffusion probabilistic models. InInternational conference on machine learning. PMLR, 8162–8171

    2021

  28. [28]

    Trevine Oorloff, Yaser Yacoob, and Abhinav Shrivastava. 2025. Mitigating Hallu- cinations in Diffusion Models through Adaptive Attention Modulation.arXiv preprint arXiv:2502.16872(2025)

    arXiv 2025

  29. [29]

    Zaki, Luca Ambrogioni, and Dmitry Krotov

    Bao Pham, Gabriel Raya, Matteo Negri, Mohammed J. Zaki, Luca Ambrogioni, and Dmitry Krotov. 2025. Memorization to Generalization: Emergence of Diffusion Models from Associative Memory.arXiv preprint arXiv:2505.21777(2025)

    arXiv 2025

  30. [30]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10684–10695

    2022

  31. [31]

    1976.Principles of Mathematical Analysis(3rd ed.)

    Walter Rudin. 1976.Principles of Mathematical Analysis(3rd ed.). McGraw-Hill, New York. Extreme Value Theorem: a continuous function on a compact set attains a minimum and maximum

    1976

  32. [32]

    Berg, and Li Fei-Fei

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexan- der C. Berg, and Li Fei-Fei. 2015. ImageNet Large Scale Visual Recognition Challenge. arXiv:1409.0575 [cs.CV] https://arxiv.org/abs/1409.0575

    Pith/arXiv arXiv 2015

  33. [33]

    Soumadeep Saha, Saptarshi Saha, and Utpal Garain. 2023. VALUED–Vision and Logical Understanding Evaluation Dataset.arXiv preprint arXiv:2311.12610 (2023)

    arXiv 2023

  34. [34]

    Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. 2022. Palette: Image-to-image diffusion models. InACM SIGGRAPH 2022 conference proceedings. 1–10

    2022

  35. [35]

    Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. 2023. Finetuning text-to-image diffusion models for fairness.arXiv preprint arXiv:2311.07604(2023)

    arXiv 2023

  36. [36]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2020. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502(2020)

    Pith/arXiv arXiv 2020

  37. [37]

    Yang Song and Stefano Ermon. 2019. Generative modeling by estimating gradi- ents of the data distribution.Advances in neural information processing systems 32 (2019)

    2019

  38. [39]

    Kingma, Abhishek Kumar, Ste- fano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Ste- fano Ermon, and Ben Poole. 2021. Score-Based Generative Modeling through Stochastic Differential Equations.ICLR(2021). https://arxiv.org/abs/2011.13456

    Pith/arXiv arXiv 2021

  39. [40]

    2024.Introducing Stable Diffusion 3.5

    Stability AI. 2024.Introducing Stable Diffusion 3.5. https://stability.ai/news/ introducing-stable-diffusion-3-5 Official announcement of the SD 3.5 model family

    2024

  40. [41]

    2025.The 2025 AI Index Report

    Stanford HAI. 2025.The 2025 AI Index Report. https://hai.stanford.edu/ai- index/2025-ai-index-report Reports 78% of organizations using AI in 2024

    2025

  41. [42]

    Chrysos, and Dimitris Samaras

    Kostas Triaridis, Alexandros Graikos, Aggelina Chatziagapi, Grigorios G. Chrysos, and Dimitris Samaras. 2025. Mitigating Diffusion Model Hallucinations with Dynamic Guidance. arXiv:2510.05356 [cs.CV] https://arxiv.org/abs/2510. 05356

    Pith/arXiv arXiv 2025

  42. [43]

    Christopher Wewer, Bart Pogodzinski, Bernt Schiele, and Jan Eric Lenssen. 2025. Spatial reasoning with denoising models.arXiv preprint arXiv:2502.21075(2025)

    arXiv 2025

  43. [44]

    Tong Wu, Zhihao Fan, Xiao Liu, Hai-Tao Zheng, Yeyun Gong, Jian Jiao, Juntao Li, Jian Guo, Nan Duan, Weizhu Chen, et al. 2023. Ar-diffusion: Auto-regressive diffusion model for text generation.Advances in Neural Information Processing Systems36 (2023), 39957–39974

    2023

  44. [45]

    Zhantao Yang, Ruili Feng, Han Zhang, Yujun Shen, Kai Zhu, Lianghua Huang, Yifei Zhang, Yu Liu, Deli Zhao, Jingren Zhou, and Fan Cheng. 2024. Lipschitz Singularities in Diffusion Models. arXiv:2306.11251 [cs.CV] https://arxiv.org/ abs/2306.11251

    arXiv 2024