Diffusion models for eye-gaze trajectory generation using position and velocity representations
Laxman Basnet, Alexander Szorkovszky, Pedro G. Lind, Anis Yazidi, Shailendra Bhandari
cs.CV · cs.AI · cs.NE
Quick verdict
Two unconditional DDPMs generate realistic 8s gaze trajectories (JS 0.016/0.0065), beating Markov baselines >10x; position-vs-velocity comparison confounded.
Summary
The paper trains two denoising diffusion probabilistic models (DDPMs) to unconditionally generate 8-second eye-gaze trajectory windows (250 Hz, L=2,000) from a public Where's-Waldo visual-search dataset, sharing one FiLM-conditioned 1D U-Net (19.35M parameters) but differing in representation: one emits raw 2D gaze positions, the other two-component velocities. Each is trained with three seeds; the position model uses 3,265 training windows from 28 participants, the velocity model only 425. The position model matches nine kinematic feature distributions with mean JS divergence 0.016±0.004 and achieves a Fréchet Gaze Distance of 13.79±2.97, more than an order of magnitude below the positional-Markov baseline (467.49) and naive baselines; I-VT event statistics reproduce fixation duration within 2% but undercount saccades by roughly 15%. A TSTR protocol yields R²=0.66±0.02 versus 0.80±0.05 for real-to-real fixation-count prediction. The velocity model matches velocity marginals, speed, log-speed, and turning angle (turning-angle JS 0.0003) with spectral distances around 10⁻⁴, but reconstructed path length (JS 0.207) is its main gap. Main limitations: the position-versus-velocity comparison is explicitly not a controlled ablation — data partitions, sizes, augmentation, loss weighting, and sampling budgets all differ; event-level and ablation analyses rest on a single reference run; no unconditional deep generative baseline is trained on the same data; no code release is stated, though the dataset is public.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Unconditional DDPMs reproduce the local kinematic structure of visual-search gaze: the position model matches nine feature distributions at mean JS 0.016±0.004 and beats the kinematic Markov baseline on every feature (Table IV).
- The position-space DDPM's FGD (13.79±2.97) is more than an order of magnitude below the strongest Markov baseline (467.49).
- Event-level statistics are close but imperfect: fixation duration within 2%, saccade count underestimated ~15% — comparable to DiffGaze's reported 13% undercount.
- Synthetic-only TSTR training recovers R²=0.66±0.02, 82.7% of the real-data point estimate, demonstrating substantial downstream utility.
- Integrated long-range structure (saccade counts, cumulative path length), not local signed or directional statistics, is the principal residual difficulty for both representations.
Strengths
Honest protocol boundaries
The paper explicitly refuses to read cross-model differences as a controlled ablation (Secs. II and IV), labels single-reference-run analyses (Tables VI, VII) separately from seed averages, and reports when validation loss fails to rank generative quality (seed 123 had the worst loss but best mean JS, Sec. V.A).
Seed-level uncertainty on headline metrics
All headline results are mean±SD over three seeds; seed sensitivity is quantified (FGD relative SD ≈22%, mean JS ≈24%, mean KS ≈9%) and the FGD ordering versus baselines is stated to hold at every seed (Sec. V.A).
Clear margins over statistical baselines
Ablations with large, practical effects
Downstream utility tested, not assumed
Weaknesses
The title's comparison is confounded
Event-level claims rest on one run
FGD depends on a bespoke feature extractor
No unconditional deep-generative baseline
Velocity-space data selection under-documented
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper trains two denoising diffusion probabilistic models (DDPMs) that *unconditionally* generate 8-second eye-gaze windows (L = 2,000 samples at 250 Hz) from a public Where's-Waldo visual-search dataset (Mathema et al. [28]). Both share one FiLM-conditioned 1D U-Net with bottleneck self-attention (19,350,914 parameters; Sec. III), differing in representation: raw (x, y) position versus two-component velocity from first-order differencing. The stated goal (Sec. I) is to establish how much intrinsic gaze kinematics diffusion can recover without stimulus or participant conditioning, and to provide 'stable, non-adversarial reference values' for future conditional models. Central claims: (1) mean JS divergence 0.016±0.004 across nine position-space kinematic features (Table I); (2) FGD 13.79±2.97, more than 10× below the strongest Markov baseline (467.49); (3) I-VT fixation duration within 2% of real, with a ~15% saccade undercount comparable to DiffGaze's 13% (Sec. VI); (4) TSTR R² = 0.66±0.02, i.e. 82.7% of the real-data point estimate; (5) the velocity model matches marginals, speed, log-speed and turning angle (JS ≤ 0.015) but not reconstructed path length (JS 0.207±0.020).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does this paper generate?
It trains two diffusion models (DDPMs) that generate synthetic 8-second eye-gaze trajectories (250 Hz) unconditionally — without any stimulus image, task, or participant cue — from a public Where's-Waldo visual-search eye-tracking dataset. One model outputs raw x/y gaze positions; the other outputs two-component gaze velocities.
How realistic are the generated gaze trajectories?
Agreement is high for local kinematics: mean Jensen-Shannon divergence of 0.016±0.004 over nine position-space features and 0.0065 across velocity-space quantities, with a Fréchet Gaze Distance of 13.79±2.97 versus 467.49 for the strongest Markov baseline. Known gaps: the position model undercounts saccades by about 15%, and the velocity model's reconstructed path length is inaccurate (JS 0.207).
Which representation works better, position or velocity?
The paper deliberately avoids ranking them: the two models use different data partitions (3,265 vs 425 training windows), augmentation, loss weighting, and sampling budgets, so the comparison is confounded and explicitly not a controlled ablation. Findings are complementary: the position model captures fixation timing well; the velocity model excels on speed and turning-angle marginals; both struggle with long-range integrated structure.
Is code or data available for this paper?
The underlying eye-tracking dataset is publicly available (Mathema et al., Scientific Data, 2026). The paper states dataset availability but does not mention releasing training or evaluation code, nor trained model checkpoints.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- PDF extraction garbles math typography throughout (decimal points, ± signs, and accents render as artifacts, e.g. 'Fr´echet', '$0.016\pm0.004$'); all numbers quoted in this review were verified legible despite this.
- Appendix Table III cell text is partially garbled by extraction (e.g., split-seed values truncated to a bare 'split'); hyperparameter details cited here were cross-checked against the Sec. III body text.
- Figures appear only as captions and embedded axis values in the extracted text; no visual inspection of the distribution plots (Figs. 3-7) was possible.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.