Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Quantitative and Qualitative Comparison of Generative Models for Subject-Specific Gaze Synthesis: Diffusion vs GANs

T0 review · 3 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A diffusion model conditioned on a compact user identity embedding and an identity-stripped gaze signal generates more realistic, identity-preserving synthetic eye-tracking sequences than a GAN conditioned on hand-crafted quality features.

desk verdict The comparison is rigged by design: DiffEyeSyn gets the target's own trajectory as input while the GAN generates from noise, so the headline result is not established. read the letter →

arxiv 2511.09867 v2 pith:2KCA4DCM submitted 2025-11-13 cs.HC

Kamrul Hasan , Dmytro Katrychuk , Mehedi Hasan Raju , Oleg V. Komogortsev This is my paper
classification cs.HC
keywords gazesynthesisdiffusionmodelsgenerativeadversarialnetworkssubject-specificgenerationeyetrackingbiometricsidentityembeddingsyntheticdata
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to determine how best to generate synthetic eye-tracking data that carries a specific person's gaze signature, for applications from biometric authentication to AR/VR testing. It takes two generative families—diffusion and GANs—and adds explicit subject conditioning to each, then compares them head-to-head on 1000 Hz gaze recordings using spatial accuracy, spatial precision, and identity-similarity metrics. The central claim is that the diffusion-based model, conditioned on a compact 128-dimensional user embedding and an identity-removed signal, generates sequences that are closer to real gaze and better preserve the target user's identity than the GAN, which is conditioned on hand-crafted quality descriptors plus one-hot identity. If true, this points to diffusion conditioning as the more effective route to subject-specific synthetic gaze, and it suggests that smaller, identity-focused conditioning vectors interfere less with generation.

What carries the argument

The load-bearing mechanism is the conditioning pair for the denoising diffusion model: an identity-removed velocity signal v0 (the raw 1000 Hz signal downsampled to 25 Hz and upsampled back, then converted to velocity by a Savitzky–Golay filter) together with a 128-dimensional user embedding extracted by a pretrained eye-movement biometric encoder. The model is trained to predict noise with two losses—standard noise reconstruction plus cosine-similarity identity guidance between the embedding of the generated and real velocity—so that generation is steered toward the target user's oculomotor signature. The GAN baseline instead concatenates one-hot identity with hand-crafted standard-deviatio

What would settle it

Measure the identity content of the conditioning signal v0 directly by feeding it to the same biometric encoder used for evaluation; if real-vs-v0 cosine similarity is high, or if the diffusion model's metrics degrade sharply when v0 is replaced by a fixed identity-free template, the identity-removal assumption fails and the comparison is confounded.

Watch

Extended reading notes

Core claim

Using a public 1000 Hz eye-tracking dataset, the authors modified a diffusion gaze synthesizer to condition on a compact user embedding from a pre-trained eye-movement identity encoder plus a velocity signal whose identity cues are suppressed by downsampling to 25 Hz and upsampling back. They also modified a GAN gaze synthesizer to condition on hand-crafted directional data-quality features and one-hot subject identity. Evaluating on unseen subjects across seven tasks, the diffusion model achieves median spatial errors of about 3.7–4.1 degrees of visual angle versus 13.6–15.5 for the GAN, constrains worst-case fixation jitter better, and yields cosine similarities of 0.92–0.95 between synthe

Load-bearing premise

The load-bearing premise is that downsampling the gaze signal to 25 Hz and upsampling it back genuinely strips the user's identity from the conditioning signal, so the diffusion model's accuracy and identity scores come from generation rather than from leaked copies of the target user's own trace.

Editorial extensions

If this is right

  • If the central claim holds, subject-specific synthetic gaze is achievable with a compact identity embedding, removing the need for high-dimensional or hand-crafted identity features.
  • Gaze-based authentication models could be pre-trained on diffusion-generated sequences with confidence that identity-relevant features survive, since synthetic embeddings nearly match real ones.
  • Applications setting target sizes or foveated-rendering thresholds can assume smaller typical gaze error when simulation data comes from the diffusion model (~4° vs ~14° at median).
  • Worst-case accuracy (U95|E95) remains poor for both models, so neither is yet ready for reliability-critical use without further work.

Reading between the lines

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

  • The large cosine-similarity gap may partly reflect conditioning leakage: if the downsample-upsample identity removal leaves user-specific micro-signals in v0, the diffusion model receives a scaffold of the target user's own trace, so the comparison against a from-noise GAN is not purely about generative capacity. A test: replace v0 with a common template and see how much accuracy drops.
  • A fairer head-to-head would give the GAN the same 128-dimensional learned embedding instead of one-hot identity, isolating conditioning format from model family.
  • The recipe suggests a privacy-preserving deployment: distribute only identity embeddings and identity-removed signals rather than raw gaze, and decode subject-specific sequences locally.
  • The diffusion model's extra microsaccades near fixations, visible in the qualitative results, hint that identity guidance may over-emphasize small idiosyncratic movements; tuning the identity-loss weight could trade identity fidelity against event-level realism.
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 / 6 minor

Summary. The paper proposes subject-specific variants of two generative models for 1000 Hz gaze sequence synthesis: a diffusion model based on DiffEyeSyn with a 128-dimensional user embedding and a 25 Hz identity-removed conditioning signal, and a GAN based on SP-EyeGAN with a new Subject-Specific Condition Generator combining DDQFE features and one-hot identity. The authors evaluate spatial accuracy, spatial precision, and cosine similarity of EKYT embeddings on GazeBase (HSS, RAN, and five other tasks), concluding that the diffusion-based approach produces more realistic and identity-preserving gaze sequences than the GAN. The paper includes qualitative figures and discusses privacy/limitations.

Significance. If the comparison were clean, the result would be a useful data point: diffusion-based conditioning with a compact identity embedding and a low-frequency scaffold appears to produce substantially more accurate and identity-similar synthetic gaze than a GAN conditioned on handcrafted quality features. The paper also ships a concrete modification of SP-EyeGAN to make it subject-aware, and the authors commit to releasing source code, which aids reproducibility. However, the central claim is not yet established because the two models are evaluated under asymmetric conditioning protocols (see major comments).

major comments (3)
  1. [§3.1.2–3.1.3, Eq. (1)–(3) vs. §3.2.4, Eq. (11)–(12)] The head-to-head comparison is confounded by an asymmetry in conditioning. DiffEyeSyn receives v0, the identity-removed version of the target user's own gaze signal for the exact trial being generated, at every denoising step (cond = (v0, z)). Since §3.1.2 states the 0–25 Hz band retains nearly all saccade-trajectory variance, v0 carries the actual saccade timings, directions, and fixation locations. SP-EyeGAN generates from latent noise plus DDQFE summary statistics and one-hot identity, with no access to the target trial's trajectory. Thus the large accuracy gap in Table 1 (e.g., 3.73 vs. 15.45 dva at U50|E50 for HSS) and the cosine-similarity gap in Table 3 may largely reflect the scaffold (v0) rather than the generative model. The claim 'diffusion produces more realistic, identity-preserving gaze sequences than GAN' is therefore not supported by the current protocol. A fix within sco
  2. [§5.3, Table 3 vs. Eq. (9)] The cosine-similarity metric is circular for the identity-preservation claim. Eq. (9) trains DiffEyeSyn with L_id = 1 − cos(φ(v̂), φ(v)) using the pre-trained EKYT encoder φ, and Table 3 then measures the same cosine similarity between φ(synthetic) and φ(real) using the same φ. DiffEyeSyn's 0.92–0.95 scores therefore reflect how well the training loss optimized the evaluation metric; they do not independently verify identity preservation. The paper should either report a different identity metric (e.g., a separately trained biometric matcher, or authentication accuracy on the synthetic data) or clearly caveat that Table 3 is a training-loss proxy.
  3. [§4.1, Table 1] The evaluation protocol for spatial accuracy and precision may also be asymmetric. Table 1 reports error between the synthetic gaze and the actual target stimulus position; DiffEyeSyn is conditioned on v0 derived from the real positional signal p, so its output is anchored to the real per-trial trajectory. SP-EyeGAN is not. For a fair spatial-accuracy comparison, both models should be evaluated under identical information conditions. At minimum, the paper should report a no-scaffold ablation of DiffEyeSyn (e.g., conditioning only on z) for Tables 1–2.
minor comments (6)
  1. [§3.1.3, Eq. (3)] The notation 'cond = (v0, z)' is used both for the conditioning set and the tuple passed to the denoiser; clarify how v0 is injected (feature-wise modulation vs. concatenation).
  2. [§3.2.3] The dimensionality of the DDQFE feature vector and the exact concatenation order with the one-hot vector are not specified. Please state the final conditioning dimension.
  3. [§5.1, Table 1] Ground-truth U95|E95 values (30.85 and 21.80 dva) are surprisingly high for known-saccade tasks; please verify that these are computed over the same 80-ms fixation bins as the synthetic results.
  4. [§5.3, Table 3] The table reports cosine similarity between synthetic and real embeddings, but the matching protocol (e.g., one-to-one per trial? averaged over trials? which user?) is not described. Specify this in the text.
  5. [§3.1.2] The text says 'the originals velocity signals (◦/s)' and later 'writeen' (Eq. 7). Please correct these typos.
  6. [§6.4] The limitation statement that the identity-removal 'may introduce minor distortions' is arguably understated given that v0 is the main scaffold for the diffusion model. Please expand on its effect and ideally provide a sensitivity analysis of the 25 Hz cutoff.

Circularity Check

2 steps flagged · score 7.0 of 10

Headline identity-preservation scores are the training objective itself, and spatial accuracy is scaffolded by conditioning on the target's own low-pass gaze path.

  1. self definitional [Eq. (9) (Sec. 3.1.3) vs. Sec. 5.3/Table 3]
    "Lid = 1−cos(ϕ(ˆv), ϕ(v)) ... we leverage pre-trained EKYT model to extract embeddings and measure the cosine similarity between the embeddings of synthetic and ground-truth sequences."

    The identity-preservation claim is evaluated with the same pre-trained EKYT encoder ϕ that is used both as the diffusion conditioner (z=ϕ(v)) and inside the identity-guidance loss. Eq. (9) explicitly trains DiffEyeSyn to maximize cos(ϕ(v̂), ϕ(v)); Table 3 then reports that same cosine as an independent realism/identity metric. Since SP-EyeGAN was never trained with this objective, the comparison only shows that DiffEyeSyn optimizes its own loss better than a model that does not optimize it. The 'identity-preserving' prediction therefore reduces by construction to the training objective.

  2. fitted input called prediction [Eqs. (1)-(3) and Sec. 5.1/Table 1]
    "v0 =SGDF(IdentityRemoval(p)) ... the 0–25 Hz band retains nearly all variance in saccade trajectories ... cond = (v0, z), z=ϕ(v)"

    DiffEyeSyn's denoiser is conditioned on v0, a low-pass version of the very trial p being evaluated, and the paper's own PvAF justification admits that this band keeps nearly all saccade-trajectory variance. Thus the model is handed the target trial's spatial path (saccade directions, locations, and timing) and only needs to generate high-frequency detail. The Table 1 accuracy gap (e.g., DiffEyeSyn 3.73 vs SP-EyeGAN 15.45 dva at U50|E50 HSS) is therefore a scaffold-reconstruction advantage, not a demonstration that diffusion produces more accurate gaze from an identity prior. SP-EyeGAN, in contrast, generates from latent noise with DDQFE summary statistics and one-hot identity (Eqs. 11-12), so the head-to-head comparison is not symmetric. Even a perfect identity-removal step would not remov

full rationale

The paper's two headline advantages over SP-EyeGAN are partially forced by construction. First, 'identity preservation' is measured with the same EKYT encoder that defines the DiffEyeSyn identity-guidance loss, making the 0.92-0.95 scores a read-out of the model's own training objective rather than an independent comparison. Second, spatial accuracy is evaluated on outputs generated while conditioning on the target trial's own 25 Hz down/up-sampled signal, which the paper itself states retains nearly all saccade-trajectory structure; this hands DiffEyeSyn the spatial scaffold for free. These are not mere self-citation issues: [27] is an external empirical source, and its use alone would not raise the score. But the two reductions affect the central comparative claim that 'diffusion produces more realistic, identity-preserving gaze sequences than the GAN,' so the score is 7 rather than lower. The remaining qualitative inspection and some precision results retain independent content, which is why the score is not 8-10.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central comparison rests on four premises the paper does not establish independently: (1) the DDPM formulation (standard); (2) that 25 Hz downsampling removes identity cues — an assumption inherited from the authors' own prior work; (3) that the EKYT encoder is a valid identity oracle used both as conditioning and as the evaluation metric; (4) that one-hot subject conditioning remains meaningful for unseen test subjects. The paper's own equations (Eqs. 9-10) tie the headline metric to the training objective; the design choices (25 Hz, 128 dims, λ) are hand-picked without ablations. No new physical entities are introduced; the SCG module is an architectural contrivance whose benefit over the original SP-EyeGAN is asserted, not demonstrated.

free parameters (3)
  • Identity-removal downsample rate = 25 Hz
    Chosen instead of the 20 Hz used in original DiffEyeSyn; justified by the authors' own PvAF analysis [27]. The threshold determines which signal content is treated as 'identity' and removed before conditioning in §3.1.2; no sensitivity analysis is provided.
  • User embedding dimensionality = 128
    Replaces DiffEyeSyn's 512-dim embedding; selected as 'compact' (§3.1.3) with no ablation across dimensionalities.
  • Identity-guidance weight λ = not reported
    Eq. (10) defines L = L_noise + λ L_id; λ controls the strength of the identity objective that drives the headline cosine-similarity result, but its value is never given in the paper.
assumptions (4)
  • standard math DDPM forward/reverse Markov chain with Gaussian noise (Eqs. 2, 5-7) is a valid generative model.
    Standard diffusion formulation from Ho et al. [25]; not in dispute.
  • domain assumption Identity cues in gaze signals are concentrated above ~25 Hz, so downsampling to 25 Hz and upsampling removes identity while preserving saccadic structure.
    Invoked in §3.1.2; relies on prior PvAF analysis [27] from the same group. If false, the conditioning signal v0 leaks the target's own identity and spatial structure, and the diffusion model's accuracy advantage is an artifact of the scaffold.
  • domain assumption EKYT's pre-trained 128-dim embedding φ is a reliable, identity-specific, unbiased oracle usable both as conditioning (z=φ(v)) and as the evaluation metric (cosine similarity).
    §3.1.3 Eq. (3),(9) and §5.3 Table 3. The same encoder appears in training and evaluation, making the identity metric self-referential; the single-fold model choice may also leak test-subject identity if that fold included test users.
  • ad hoc to paper One-hot subject encoding (or SCG conditioning) remains meaningful for test subjects not in the training set.
    §3.2.3; the test set has 59 unseen subjects but one-hot dimensionality equals the number of training subjects. The paper does not specify how unseen-subject conditioning is performed at inference.
invented entities (1)
  • Subject-Specific Condition Generator (SCG) with DDQFE + one-hot branches
    purpose: Adds subject identity and signal-quality conditioning to FixGAN/SacGAN in §3.2.3-3.2.4.
    A hand-designed conditioning module; its claimed benefit over the original unconditioned SP-EyeGAN is asserted (§5, §6.2) but no baseline comparison is shown, and the feature choice (std, RMS) is not validated against alternatives.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantitative and Qualitative Comparison of Generative Models for Subject-Specific Gaze Synthesis: Diffusion vs GANs." pith.science (2026). https://pith.science/paper/2KCA4DCM

@misc{pith2026251109867,
  author       = {Pith},
  title        = {Pith review of: Quantitative and Qualitative Comparison of Generative Models for Subject-Specific Gaze Synthesis: Diffusion vs GANs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2KCA4DCM}},
  note         = {Machine review of arXiv:2511.09867}
}
read the original abstract

Gaze-based biometrics has emerged as a promising approach for user authentication, but advances in this area are constrained by the limited availability of high-quality, subject-specific gaze recordings. Recent generative models have shown promise for synthesizing gaze data, yet most existing approaches rely on random noise distributions or global, predefined latent embeddings and do not explicitly model subject-specific gaze characteristics. To address this limitation, we revisit two recent generative models, diffusion and generative adversarial networks (GANs), and modify both to support subject-aware gaze synthesis. For the diffusion-based approach, we incorporate compact user embeddings to capture subject-level gaze traits. For the GAN-based approach, we introduce a subject-specific conditioning module that guides the generator to preserve idiosyncratic gaze patterns. Later, we evaluate both approaches using standard eye-movement signal quality metrics, including spatial accuracy and precision, and assess whether the generated sequences retain identity-related features relevant to biometric applications. Experimental results show that the diffusion-based approach produces more realistic, identity-preserving gaze sequences than the GAN-based approach. Overall, this work advances the understanding of synthetic gaze quality, realism, and subject specificity and supports the development of gaze-based biometric applications.

Figures

Figures reproduced from arXiv: 2511.09867 by the authors.

Figure 1
Figure 1. (a) Overview of the updated DiffEyeSyn architecture: The original positional signal [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Qualitative comparison between SP-EyeGAN and DiffEyeSyn for two different tasks: (a) TEX and (b) RAN. [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Privatization of Synthetic Gaze: Attenuating State Signatures in Diffusion-Generated Eye Movements

    cs.HC 2026-01 conditional novelty 4.0 of 10

    Diffusion-generated synthetic gaze shows weaker, less stable correlations with self-reported fatigue and difficulty than real gaze, suggesting partial state-level privatization.

Reference graph

Works this paper leans on

51 extracted references · 7 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Advances in eye tracking technology: theory, algorithms, and applications.Computational intelligence and neuroscience, 2016:7831469, 2016

    Hong Fu, Ying Wei, Francesco Camastra, Pietro Arico, and Hong Sheng. Advances in eye tracking technology: theory, algorithms, and applications.Computational intelligence and neuroscience, 2016:7831469, 2016

    2016

  2. [2]

    Eye-tracking in ar/vr: A technological review and future directions.IEEE Open Journal on Immersive Displays, 2024

    Xin Jin, Suyu Chai, Jie Tang, Xianda Zhou, and Kai Wang. Eye-tracking in ar/vr: A technological review and future directions.IEEE Open Journal on Immersive Displays, 2024

    2024

  3. [3]

    Gazebase, a large-scale, multi-stimulus, longitudinal eye movement dataset.Scientific Data, 8(1):184, 2021

    Henry Griffith, Dillon Lohr, Evgeny Abdulin, and Oleg Komogortsev. Gazebase, a large-scale, multi-stimulus, longitudinal eye movement dataset.Scientific Data, 8(1):184, 2021

    2021

  4. [4]

    The promise of eye-tracking methodology in organizational research: A taxonomy, review, and future avenues.Organizational Research Methods, 22(2):590–617, 2019

    Martin Meißner and Josua Oll. The promise of eye-tracking methodology in organizational research: A taxonomy, review, and future avenues.Organizational Research Methods, 22(2):590–617, 2019

    2019

  5. [5]

    Eye-tracking based classification of mandarin chinese readers with and without dyslexia using neural sequence models.arXiv preprint arXiv:2210.09819, 2022

    Patrick Haller, Andreas Säuberli, Sarah Elisabeth Kiener, Jinger Pan, Ming Yan, and Lena Jäger. Eye-tracking based classification of mandarin chinese readers with and without dyslexia using neural sequence models.arXiv preprint arXiv:2210.09819, 2022

    arXiv 2022

  6. [6]

    Eye-tracking based autism spectrum disorder diagnosis using chaotic butterfly optimization with deep learning model.Computers, Materials & Continua, 76(2), 2023

    Tamilvizhi Thanarajan, Youseef Alotaibi, Surendran Rajendran, and Krishnaraj Nagappan. Eye-tracking based autism spectrum disorder diagnosis using chaotic butterfly optimization with deep learning model.Computers, Materials & Continua, 76(2), 2023

    2023

  7. [7]

    Eye know you too: Toward viable end-to-end eye movement biometrics for user authentication.IEEE Transactions on Information Forensics and Security, 17:3151–3164, 2022

    Dillon Lohr and Oleg V Komogortsev. Eye know you too: Toward viable end-to-end eye movement biometrics for user authentication.IEEE Transactions on Information Forensics and Security, 17:3151–3164, 2022. 11 Quantitative and Qualitative Comparison of Generative Models for Subject-Specific Gaze Synthesis: Diffusion vs GAN

    2022

  8. [8]

    Gaze authentication: Factors influencing authentication performance.arXiv preprint arXiv:2509.10969, 2025

    Dillon Lohr, Michael J Proulx, Mehedi Hasan Raju, and Oleg V Komogortsev. Gaze authentication: Factors influencing authentication performance.arXiv preprint arXiv:2509.10969, 2025

    arXiv 2025

Show all 51 references
  1. [9]

    Person identification using ocular biometrics with liveness detection, July 14 2015

    Oleg V Komogortsev. Person identification using ocular biometrics with liveness detection, July 14 2015. US Patent 9,082,011

    2015

  2. [10]

    Iris print attack detection using eye movement signals

    Mehedi Hasan Raju, Dillon J Lohr, and Oleg Komogortsev. Iris print attack detection using eye movement signals. In2022 Symposium on Eye Tracking Research and Applications, pages 1–6, 2022

    2022

  3. [11]

    Towards foveated rendering for gaze-tracked virtual reality.ACM Transactions On Graphics (TOG), 35 (6):1–12, 2016

    Anjul Patney, Marco Salvi, Joohwan Kim, Anton Kaplanyan, Chris Wyman, Nir Benty, David Luebke, and Aaron Lefohn. Towards foveated rendering for gaze-tracked virtual reality.ACM Transactions On Graphics (TOG), 35 (6):1–12, 2016

    2016

  4. [12]

    Improving user experience of eye tracking-based interaction: Introspecting and adapting interfaces.ACM Transactions on Computer-Human Interaction (TOCHI), 26(6):1–46, 2019

    Raphael Menges, Chandan Kumar, and Steffen Staab. Improving user experience of eye tracking-based interaction: Introspecting and adapting interfaces.ACM Transactions on Computer-Human Interaction (TOCHI), 26(6):1–46, 2019

    2019

  5. [13]

    Brockmole, and Sidney K

    Stephen Hutt, Kristina Krasich, James R. Brockmole, and Sidney K. D’Mello. Breaking out of the lab: Mitigating mind wandering with gaze-based attention-aware technology in classrooms. InProceedings of the 2021 CHI conference on human factors in computing systems, pages 1–14, 2021

    2021

  6. [14]

    A new comprehensive eye-tracking test battery concurrently evaluating the pupil labs glasses and the eyelink 1000.PeerJ, 7:e7086, 2019

    Benedikt V Ehinger, Katharina Groß, Inga Ibs, and Peter König. A new comprehensive eye-tracking test battery concurrently evaluating the pupil labs glasses and the eyelink 1000.PeerJ, 7:e7086, 2019

    2019

  7. [15]

    Biometric verification via complex eye movements: The effects of environment and stimulus

    Corey D Holland and Oleg V Komogortsev. Biometric verification via complex eye movements: The effects of environment and stimulus. In2012 IEEE Fifth International Conference on Biometrics: Theory, Applications and Systems (BTAS), pages 39–46. IEEE, 2012

    2012

  8. [16]

    Dillon Lohr, Henry Griffith, and Oleg V Komogortsev. Eye know you: Metric learning for end-to-end biometric authentication using eye movements from a longitudinal dataset.IEEE Transactions on Biometrics, Behavior, and Identity Science, 4(2):276–288, 2022

    2022

  9. [17]

    Supreyes: Super resolutin for eyes using implicit neural representation learning

    Chuhan Jiao, Zhiming Hu, Mihai Bâce, and Andreas Bulling. Supreyes: Super resolutin for eyes using implicit neural representation learning. InProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, pages 1–13, 2023

    2023

  10. [18]

    A survey of advances in vision-based vehicle re-identification.Computer Vision and Image Understanding, 182:50–63, 2019

    Sultan Daud Khan and Habib Ullah. A survey of advances in vision-based vehicle re-identification.Computer Vision and Image Understanding, 182:50–63, 2019

    2019

  11. [19]

    Privacy-aware eye tracking using differential privacy

    Julian Steil, Inken Hagestedt, Michael Xuelin Huang, and Andreas Bulling. Privacy-aware eye tracking using differential privacy. InProceedings of the 11th ACM Symposium on Eye Tracking Research & Applications, pages 1–9, 2019

    2019

  12. [20]

    Preserving privacy in healthcare: A systematic review of deep learning approaches for synthetic data generation.Computer Methods and Programs in Biomedicine, 260: 108571, 2025

    Yintong Liu, U Rajendra Acharya, and Jen Hong Tan. Preserving privacy in healthcare: A systematic review of deep learning approaches for synthetic data generation.Computer Methods and Programs in Biomedicine, 260: 108571, 2025

    2025

  13. [21]

    Eyes alive

    Sooha Park Lee, Jeremy B Badler, and Norman I Badler. Eyes alive. InProceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 637–644, 2002

    2002

  14. [22]

    Eyesyn: Psychology-inspired eye movement synthesis for gaze-based activity recognition

    Guohao Lan, Tim Scargill, and Maria Gorlatova. Eyesyn: Psychology-inspired eye movement synthesis for gaze-based activity recognition. In2022 21st ACM/IEEE international conference on information processing in sensor networks (IPSN), pages 233–246. IEEE, 2022

    2022

  15. [23]

    Generative adversarial networks.Communications of the ACM, 63(11):139–144, 2020

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks.Communications of the ACM, 63(11):139–144, 2020

    2020

  16. [24]

    Sp-eyegan: Generating synthetic eye movement data with generative adversarial networks

    Paul Prasse, David Robert Reich, Silvia Makowski, Seoyoung Ahn, Tobias Scheffer, and Lena A Jäger. Sp-eyegan: Generating synthetic eye movement data with generative adversarial networks. InProceedings of the 2023 symposium on eye tracking research and applications, pages 1–9, 2023

    2023

  17. [25]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  18. [26]

    Diffeyesyn: Diffusion-based user-specific eye movement synthesis.arXiv preprint arXiv:2409.01240, 2024

    Chuhan Jiao, Guanhua Zhang, Yeonjoo Cho, Zhiming Hu, and Andreas Bulling. Diffeyesyn: Diffusion-based user-specific eye movement synthesis.arXiv preprint arXiv:2409.01240, 2024

    2024 arXiv

  19. [27]

    Determining which sine wave frequencies correspond to signal and which correspond to noise in eye-tracking time-series.Journal of Eye Movement Research, 14(3):16, 2021

    Mehedi H Raju, Lee Friedman, Troy M Bouman, and Oleg V Komogortsev. Determining which sine wave frequencies correspond to signal and which correspond to noise in eye-tracking time-series.Journal of Eye Movement Research, 14(3):16, 2021. 12 Quantitative and Qualitative Comparis...

    2021

  20. [28]

    Evaluation of eye tracking signal quality for virtual reality applications: A case study in the meta quest pro

    Samantha Aziz, Dillon J Lohr, Lee Friedman, and Oleg Komogortsev. Evaluation of eye tracking signal quality for virtual reality applications: A case study in the meta quest pro. InProceedings of the 2024 Symposium on Eye Tracking Research and Applications, pages 1–8, 2024

    2024

  21. [29]

    Modeling physiologically plausible eye rotations

    Andrew T Duchowski and Sophie Jörg. Modeling physiologically plausible eye rotations. InProceedings of computer graphics international, 2015

    2015

  22. [30]

    Eye movement synthesis

    Andrew T Duchowski, Sophie Jörg, Tyler N Allen, Ioannis Giannopoulos, and Krzysztof Krejtz. Eye movement synthesis. InProceedings of the ninth biennial ACM symposium on eye tracking research & applications, pages 147–154, 2016

    2016

  23. [31]

    Natural eye motion synthesis by modeling gaze-head coupling

    Xiaohan Ma and Zhigang Deng. Natural eye motion synthesis by modeling gaze-head coupling. In2009 IEEE Virtual Reality Conference, pages 143–150. IEEE, 2009

    2009

  24. [32]

    Rendering of eyes for eye-shape registration and gaze estimation

    Erroll Wood, Tadas Baltrusaitis, Xucong Zhang, Yusuke Sugano, Peter Robinson, and Andreas Bulling. Rendering of eyes for eye-shape registration and gaze estimation. InProceedings of the IEEE international conference on computer vision, pages 3756–3764, 2015

    2015

  25. [33]

    Live speech driven head-and-eye motion generators.IEEE transactions on visualization and computer graphics, 18(11):1902–1914, 2012

    Binh H Le, Xiaohan Ma, and Zhigang Deng. Live speech driven head-and-eye motion generators.IEEE transactions on visualization and computer graphics, 18(11):1902–1914, 2012

    1902

  26. [34]

    Eye movement velocity and gaze data generator for evaluation, robustness testing and assess of eye tracking software and visualization tools.arXiv preprint arXiv:1808.09296, 2018

    Wolfgang Fuhl and Enkelejda Kasneci. Eye movement velocity and gaze data generator for evaluation, robustness testing and assess of eye tracking software and visualization tools.arXiv preprint arXiv:1808.09296, 2018

    2018 arXiv

  27. [35]

    Eye movement simulation and detector creation to reduce laborious parameter adjustments.arXiv preprint arXiv:1804.00970, 2018

    Wolfgang Fuhl, Thiago Santini, Thomas Kuebler, Nora Castner, Wolfgang Rosenstiel, and Enkelejda Kasneci. Eye movement simulation and detector creation to reduce laborious parameter adjustments.arXiv preprint arXiv:1804.00970, 2018

    2018 arXiv

  28. [36]

    Eyecatch: Simulating visuomotor coordination for object interception.ACM Transactions on Graphics (TOG), 31(4):1–10, 2012

    Sang Hoon Yeo, Martin Lesmana, Debanga R Neog, and Dinesh K Pai. Eyecatch: Simulating visuomotor coordination for object interception.ACM Transactions on Graphics (TOG), 31(4):1–10, 2012

    2012

  29. [37]

    An introduction to the kalman filter

    Greg Welch, Gary Bishop, et al. An introduction to the kalman filter. 1995

    1995

  30. [38]

    Automatic scanpath generation with deep recurrent neural networks

    Daniel Simon, Srinivas Sridharan, Shagan Sah, Raymond Ptucha, Chris Kanan, and Reynold Bailey. Automatic scanpath generation with deep recurrent neural networks. InProceedings of the ACM symposium on applied perception, pages 130–130, 2016

    2016

  31. [39]

    Pathgan: Visual scanpath prediction with generative adversarial networks

    Marc Assens, Xavier Giro-i Nieto, Kevin McGuinness, and Noel E O’Connor. Pathgan: Visual scanpath prediction with generative adversarial networks. InProceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018

    2018

  32. [40]

    Eyegan: Gaze-preserving, mask-mediated eye image synthesis

    Harsimran Kaur and Roberto Manduchi. Eyegan: Gaze-preserving, mask-mediated eye image synthesis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 310–319, 2020

    2020

  33. [41]

    Fully convolutional neural networks for raw eye tracking data segmentation, generation, and reconstruction

    Wolfgang Fuhl, Yao Rong, and Enkelejda Kasneci. Fully convolutional neural networks for raw eye tracking data segmentation, generation, and reconstruction. In2020 25th International Conference on Pattern Recognition (ICPR), pages 142–149. IEEE, 2021

    2021

  34. [42]

    Next-generation deep learning based on simulators and synthetic data.Trends in cognitive sciences, 26(2):174–187, 2022

    Celso M De Melo, Antonio Torralba, Leonidas Guibas, James DiCarlo, Rama Chellappa, and Jessica Hodgins. Next-generation deep learning based on simulators and synthetic data.Trends in cognitive sciences, 26(2):174–187, 2022

    2022

  35. [43]

    Hpcgen: Hierarchical k-means clustering and level based principal components for scan path genaration

    Wolfgang Fuhl and Enkelejda Kasneci. Hpcgen: Hierarchical k-means clustering and level based principal components for scan path genaration. In2022 Symposium on Eye Tracking Research and Applications, pages 1–7, 2022

    2022

  36. [44]

    Improved denoising diffusion probabilistic models

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

    2021

  37. [45]

    Diffgaze: A diffusion model for modelling fine-grained human gaze behaviour on 360° images.ACM Transactions on Interactive Intelligent Systems, 2025

    Chuhan Jiao, Yao Wang, Guanhua Zhang, Mihai Bâce, Zhiming Hu, and Andreas Bulling. Diffgaze: A diffusion model for modelling fine-grained human gaze behaviour on 360° images.ACM Transactions on Interactive Intelligent Systems, 2025

    2025

  38. [46]

    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. InProceedings of the IEEE/CVF international conference on computer vision, pages 3836–3847, 2023

    2023

  39. [47]

    Diffwave: A versatile diffusion model for audio synthesis.arXiv preprint arXiv:2009.09761, 2020

    Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis.arXiv preprint arXiv:2009.09761, 2020

    2009 arXiv

  40. [48]

    Smoothing and differentiation of data by simplified least squares procedures.Analytical chemistry, 36(8):1627–1639, 1964

    Abraham Savitzky and Marcel JE Golay. Smoothing and differentiation of data by simplified least squares procedures.Analytical chemistry, 36(8):1627–1639, 1964. 13 Quantitative and Qualitative Comparison of Generative Models for Subject-Specific Gaze Synthesis: Diffusion vs GAN

    1964

  41. [49]

    Identifying fixations and saccades in eye-tracking protocols

    Dario D Salvucci and Joseph H Goldberg. Identifying fixations and saccades in eye-tracking protocols. In Proceedings of the 2000 symposium on Eye tracking research & applications, pages 71–78, 2000

    2000

  42. [50]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization.CoRR, abs/1412.6980, 2014

    2014 arXiv

  43. [51]

    Evaluating the data quality of eye tracking signals from a virtual reality system: Case study using smi’s eye-tracking htc vive.arXiv preprint arXiv:1912.02083, 2019

    Dillon J Lohr, Lee Friedman, and Oleg V Komogortsev. Evaluating the data quality of eye tracking signals from a virtual reality system: Case study using smi’s eye-tracking htc vive.arXiv preprint arXiv:1912.02083, 2019. 14

    1912 arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.