Pith. sign in

REVIEW 5 major objections 4 minor 1 cited by

Optical Physics-Based Generative Models

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that six optical equations—Helmholtz, dissipative wave, Eikonal, and their nonlinear extensions—can be reformulated as generative density flows, and that the nonlinear versions dramatically outperform the linear ones.

desk verdict The paper's central density-flow derivations for the dissipative wave and Eikonal equations are algebraically wrong, so the claimed optical-physics-to-generative-model mapping does not hold; the experimental numbers are also internally inconsistent. read the letter →

arxiv 2506.04357 v1 pith:6XA7HYYD submitted 2025-06-04 physics.optics cs.LG

Amirreza Ahmadnejad , Somayyeh Koohi This is my paper
classification physics.opticscs.LG MSC 35Q6078A0568T07
keywords opticalphysicsgenerativemodelsnonlinearopticsHelmholtzequationdissipativewaveEikonalKerreffectdensityflows-generativePDE
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that the mathematics of light propagation is a generative model in waiting. Its central claim is that six optical equations—Helmholtz, dissipative wave, Eikonal, and their Kerr, cubic-quintic, and intensity-dependent extensions—can each be rewritten as a probability density flow satisfying two conditions: the density stays nonnegative, and high-frequency components decay faster than low-frequency ones. If that holds, a neural network trained on the flow's velocity and birth/death fields can generate samples by running the flow backward. The paper reports that the nonlinear versions are not just valid but stronger: the nonlinear Helmholtz model reaches FID 0.0089 versus 1.0909 for the linear one, with lower parameter counts, lower memory, and better mode coverage. A sympathetic reader takes away a concrete recipe for turning wave physics into a sampler.

What carries the argument

The machinery is the s-generative PDE reformulation: a generalized continuity equation $\partial p/\partial t + \nabla\cdot(p v) - R = 0$ with $p \ge 0$ (condition C1) plus a dispersion-relation criterion $\mathrm{Im}\,\omega(k) < \mathrm{Im}\,\omega(0)$ for all $\|k\|>0$ (condition C2). For each optical equation the paper supplies explicit formulas, such as $p=-\phi_t$, $v=\nabla\phi/\phi_t$, $R=k_0^2\phi$ for Helmholtz; $p=-(\phi_t+2\epsilon\phi)$, $v=\nabla\phi/(\phi_t+2\epsilon\phi)$ for the dissipative wave equation; and $p=\phi$ for the Eikonal equation. The birth/death term $R$ handles non-conservative probability mass, while the dispersion relation is what turns wave physics into a smoothing, forgetting generative process.

What would settle it

For a dissipative-wave test field, numerically compare $\partial p/\partial t + \nabla\cdot(p v)$ with zero on a grid; with $p=-(\phi_t+2\epsilon\phi)$ and $v=\nabla\phi/(\phi_t+2\epsilon\phi)$, the divergence $p v$ equals $-\Delta\phi$, so the claimed continuity equation would predict the opposite sign of the Laplacian from the original equation. A nonzero residual falsifies the claimed equivalence.

Watch

Extended reading notes

Core claim

The central claim is that each of six optical partial differential equations admits an exact density-flow representation $\partial p/\partial t + \nabla\cdot(p v) - R = 0$ with a nonnegative density and a dispersion relation that damps high frequencies faster than low ones. These two properties, called C1 and C2, make the equation an s-generative PDE, meaning it can serve as a generative model by simulating the flow backward in time from a simple prior to the data distribution. The paper derives explicit density, velocity, and birth/death fields for each optical equation, trains neural networks on those fields, and reports that the nonlinear versions improve sample quality, mode coverage, and computational efficiency over their linear counterparts. The nonlinear Helmholtz model is reported to reach FID 0.0089 versus 1.0909 for the linear version, with 40–60% parameter and memory reductions.

Load-bearing premise

The load-bearing premise is that the algebraic reformulations are correct: each stated p, v, and R must make the generalized continuity equation hold exactly whenever the original optical equation holds.

Editorial extensions

If this is right

  • Nonlinear Helmholtz with Kerr effect achieves FID 0.0089 versus 1.0909 for the linear Helmholtz model on the paper's benchmarks.
  • Kerr self-focusing cuts parameter counts by 40–60% and training time by 30–50% while preserving mode separation.
  • Cubic-quintic dissipation prevents mode collapse, retaining roughly 87% of initial mode separation compared to 34% for the cubic-only model.
  • Intensity-dependent Eikonal guidance enables conditional generation with 30–50% fewer steps than classifier-guidance methods.
  • The same density-flow framework reconstructs refractive index distributions from wave patterns with roughly 90–95% accuracy, giving a bidirectional bridge between generative AI and optical inverse problems.

Reading between the lines

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

  • If the flow reformulations are taken as given, the same recipe should apply to other optical PDEs with decaying dispersion relations, turning systems such as nonlinear Schrödinger-type or Maxwell-Bloch equations into generative samplers without architectural changes.
  • A testable extension would replace the discrete birth/death branching with a continuous importance-weight update, potentially eliminating resampling and making sampler cost scale with intrinsic mode count rather than particle count.
  • The reported FID gains suggest the self-focusing inductive bias carries much of the mode-separation burden; ablating the learned velocity field and keeping only the analytic birth/death term would test how much network capacity the nonlinearity actually replaces.
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

5 major / 4 minor

Summary. The paper proposes that six optical PDEs—Helmholtz, dissipative wave, Eikonal, and their Kerr, cubic-quintic, and intensity-dependent nonlinear extensions—can be cast as exact s-generative density flows of the form ∂p/∂t+∇·(pv)−R=0, and it reports large performance gains for the nonlinear variants. The theoretical contribution is the explicit p, v, R identification for each equation plus a dispersion-relation criterion for the smoothing condition. I verified by direct substitution that the linear Helmholtz reformulation of Section 3.1 satisfies the stated continuity equation, but the dissipative-wave and Eikonal reformulations do not: the dissipative-wave flow leaves a residual of −2∇²φ, and the Eikonal 'correction term' has the wrong sign. The nonlinear extensions in Sections 4.2 and 4.3 inherit these failures. The experimental sections also contain mutually inconsistent numbers for the same quantities. The central claim is therefore not supported.

Significance. If the density-flow reformulations were correct, the paper would offer a useful bridge between nonlinear optical dynamics and generative modeling, with potentially interesting inductive biases and a checkable mathematical core. The paper deserves credit for writing down explicit p, v, R formulas rather than relying on loose analogies, and for including extensive parameter sweeps, ablation studies, and pseudocode. However, the load-bearing algebra fails for two of the three linear equations and for their nonlinear extensions, so the theoretical foundation collapses. The reported empirical results are also internally contradictory, preventing the experiments from compensating for the analytic errors. No machine-checked proofs or self-contained reproducible code are included in the manuscript.

major comments (5)
  1. [§3.2, Eqs. (38)–(41)] The claimed density-flow form for the dissipative wave equation is algebraically false. Substituting p=−(φ_t+2εφ) and v=∇φ/(φ_t+2εφ) into Eq. (1) gives ∂p/∂t+∇·(pv)=−φ_tt−2εφ_t−∇²φ. Using the original PDE φ_tt+2εφ_t−∇²φ=0, the residual is −2∇²φ, not zero. Therefore Eq. (38) does not represent the dissipative wave equation, and the claim that R=0 is incorrect. This error propagates to the cubic-quintic extension in Section 4.2, whose velocity and birth/death fields (Eqs. (77)–(78)) satisfy the same incorrect conservation identity.
  2. [§3.3, Eqs. (56)–(61)] The Eikonal reformulation does not reproduce the original equation even after the stated correction. With R=−(n²−|∇φ|²−∇²φ), Eq. (1) becomes φ_t−|∇φ|²=−n², whereas the original Eikonal equation is φ_t+|∇φ|²=n². The required birth/death term would be R=n²−|∇φ|²+∇²φ, not the expression in Eq. (61). Moreover, even the intermediate Eq. (56) is written with a sign such that it implies φ_t+|∇φ|²=n²−∇²φ, which is not the original PDE. The intensity-dependent Eikonal model of Section 4.3 inherits the same sign error.
  3. [§2 and §3.1, Eq. (7) and Eqs. (22)–(24)] The claimed equivalence between Condition C2 and Im[ω(k)]<Im[ω(0)] for all ‖k‖>0 is asserted on the basis of the authors' own simulations and deferred to an appendix that gives only a heuristic amplitude-ratio argument. More importantly, the paper's own Helmholtz dispersion analysis contradicts the criterion: for 0<‖k‖≤k0 the paper finds Imω(k)=0=Imω(0), so the strict inequality in Eq. (7) fails. Calling the Helmholtz equation 'conditionally s-generative' therefore does not follow from the stated criterion, and the smoothing condition for this model is not established.
  4. [§4.1, §6.5, §7, and Abstract] The reported FID numbers are irreconcilable. The Abstract and Section 7 report FID 0.0089 for nonlinear Helmholtz; Section 4.1, Table 2 reports the best FID as 598.332; Section 6.5.1 reports 0.8721; Section 6.5.2 reports 20.96; and Table 6 reports 0.0191. Similar contradictions appear for mode coverage: Section 6.2 reports dissipative-wave mode coverage 0.94, while Figure 34 in the same section reports 0.05. Reconstruction accuracy is 89.9% in Section 6.4 but 95% in the Abstract and Conclusion. Because the empirical claims are essential to the paper's central assertion of nonlinear superiority, these inconsistencies cannot be dismissed as presentation issues.
  5. [§2 and §6.3] Condition C1 requires p(x,t)≥0 for all x and t, but the proposed densities p=−φ_t, p=−(φ_t+2εφ), and p=φ are not guaranteed nonnegative by the PDE dynamics. The paper's own C1 verification in Figure 36 measures the 'fraction of domain with non-negative density,' which acknowledges that p becomes negative in parts of the domain. Algorithm 5 explicitly clips φ to enforce positivity, which changes the equation being solved. Thus C1, as defined, is not established for any of the three models.
minor comments (4)
  1. [§3.2] The text says the final dissipative-wave form follows by 'further algebraic manipulation (full derivation in Appendix B),' but Appendix B contains numerical implementation details and does not present the promised derivation of Eq. (38).
  2. [§3.3] The sentence 'For a linearized version of the Eikonal equation around a constant solution, we can show that the dispersion relation is ω=−i|k|²' is not backed by a derivation; Appendix A's linearization around a plane wave yields real ω, and the imaginary dispersion relation appears only for the modified equation with the Laplacian term.
  3. [§2, Eq. (7)] The phrase 'Based on our numerical simulations and theoretical analysis (see Figure 37)' is used to justify the central C2 criterion; citing a numerical figure as the basis for an equivalence is not a proof and should be replaced by a self-contained argument or by a clearly stated assumption.
  4. [§5.3.1, Eq. (110)] The sign conventions in the split-step Fourier update appear inconsistent: the linear operator is written as L=−∇²−k0², but the Fourier-domain update uses exp(−iΔt(k²−k0²)), which corresponds to L=+∇²+k0². This should be checked and unified.

Circularity Check

1 steps flagged · score 6.0 of 10

The Eikonal models' s-generative status reduces by construction: the Laplacian that supplies the smoothing (C2) is introduced through the chosen density-flow parameters and birth/death term, not present in the Eikonal equation itself.

  1. self definitional [Section 3.3, Eqs. 52–61 and Appendix A, S1.1.3]
    "However, when we consider the full nonlinear equation with the birth/death term derived in our density flow formulation: R(x, t) = −(n2(x) − |∇φ|2 − ∇2φ). The Laplacian term ∇2φ introduces a diffusive component, leading to: φt ≈ ∇2φ + n2(x) − |∇φ|2. This modified equation has a dispersion relation with imaginary component: ω = −i|k|2, confirming the s-generative property for the full Eikonal model with the birth/death correction term."

    The original Eikonal equation (Eq. 52) has no Laplacian: φt+|∇φ|2=n2(x). The paper's density-flow representation introduces ∇2φ via the arbitrary choice p=φ, v=∇φ/φ and then places the remaining mismatch in R (Eq. 61). The 'modified equation' whose dispersion relation gives the smoothing branch ω=−i|k|2 is therefore the flow equation the authors constructed, not the Eikonal equation. Declaring the Eikonal model s-generative on this basis makes Condition C2 follow from the chosen parametrization (the added diffusive term) rather than from the optical PDE itself.

full rationale

Most of the paper is not circular. The Helmholtz density-flow mapping (Section 3.1) is algebraically exact with p=−φt, v=∇φ/φt, R=k0^2φ, and the nonlinear Helmholtz extension substitutes α|φ|^2φ directly into R; the GenPhys C1/C2 conditions are taken from an external source (Liu et al. [13]). The paper's self-citations ([15]–[19]) are background hardware/repository references and are not load-bearing. The circularity is concentrated in the Eikonal family: the original Eikonal equation contains no Laplacian, but the flow parametrization p=φ, v=∇φ/φ and the birth/death term R (Eqs. 59–61, 87) introduce ∇2φ, and the smoothing branch ω=−i|k|^2 is then read off from this modified, Laplacian-containing equation and attributed to the Eikonal model. That is Condition C2 by construction. I do not count the dissipative-wave and cubic-quintic flow derivations as circular: they are algebraically inconsistent (Eqs. 38–40 leave a −2∇2φ residual when substituted into Eq. 1, and the promised Appendix B derivation is absent), but those are correctness failures that independently damage the central claim rather than reductions to inputs. The C2/dispersion-relation equivalence is also asserted partly via the authors' own Figure 37, but an independent Fourier argument is sketched in Appendix A, so it is not scored as circular. Overall score 6 reflects one central 'prediction'—the Eikonal s-generative status—reducing by construction, while the Helmholtz/nonlinear-Helmholtz chain remains independent.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

All headline performance numbers come from configurations tuned on the evaluation set by grid search or Bayesian optimization (Section 5.2). The theoretical certification relies on the paper's own C1/C2 criteria, one of which is relaxed in verification. No new physical entities are introduced.

free parameters (6)
  • k0 (Helmholtz wavenumber) = 3.5 (8-Gaussian and MNIST)
    Chosen by grid search over [0.5, 5.0] minimizing FID/MMD (Figure 25 and Table S3).
  • epsilon (dissipative wave damping) = 0.31
    Chosen by grid search over [0.25, 2.0]; optimal value for eight-Gaussian dataset (Figure 25 and Table S4).
  • nscale (Eikonal refractive index scale) = 0.7
    Chosen by grid search over [0.2, 1.0] for Gaussian bump patterns (Section 5.2 and Table S5).
  • alpha (Kerr nonlinearity coefficient) = 0.2 to 0.5 (Section 4.1, Figure 11; conclusion states 0.2)
    Optimal value selected from parameter sweep; Table 2 lists alpha=0.1 for the reported minimum FID, adding inconsistency.
  • beta (quintic nonlinearity coefficient) = -0.2 (Section 4.2) or -0.8 (Conclusion)
    Optimal quintic coefficient from parameter sweeps; the two stated values disagree.
  • chi (intensity-dependent Eikonal coupling) = 0.3 (Conclusion)
    Chosen as the intensity coupling for the nonlinear Eikonal model; no sweep details are given.
assumptions (4)
  • domain assumption Conditions C1 and C2 from GenPhys are sufficient for a PDE to define a valid generative model.
    Invoked in Section 2 to define s-generative PDEs; borrowed from Liu et al. [13] and not re-proved.
  • ad hoc to paper Condition C2 is equivalent to the dispersion inequality Im[omega(k)] < Im[omega(0)] for all ||k||>0.
    Asserted in Section 2 with only a sketch in Appendix A; the paper's own verification (Figure 37) uses this criterion.
  • domain assumption The field phi, or the derived p, can be interpreted as a nonnegative probability density.
    Sections 3.1-3.3 set p = -phi_t, -(phi_t + 2epsilon phi), or phi, none of which is guaranteed nonnegative; Figure 36 only measures the fraction of the domain with nonnegative density, silently relaxing C1.
  • ad hoc to paper For the Eikonal equation, the density-flow form with the stated R reproduces the original equation after including a correction term.
    Section 3.3 Eqs. (56)-(61) introduce R = -(n^2 - |grad phi|^2 - Laplacian phi), which does not satisfy phi_t + |grad phi|^2 = n^2; the paper acknowledges a mismatch and adds a correction term without derivation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optical Physics-Based Generative Models." pith.science (2026). https://pith.science/paper/6XA7HYYD

@misc{pith2026250604357,
  author       = {Pith},
  title        = {Pith review of: Optical Physics-Based Generative Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6XA7HYYD}},
  note         = {Machine review of arXiv:2506.04357}
}
read the original abstract

This paper establishes a comprehensive mathematical framework connecting optical physics equations to generative models, demonstrating how light propagation dynamics inspire powerful artificial intelligence approaches. We analyze six fundamental optical equations, comparing linear models (Helmholtz, dissipative wave, and Eikonal equations) with their nonlinear extensions incorporating Kerr effects, cubic-quintic nonlinearities, and intensity-dependent refractive indices. Our nonlinear optical models reveal remarkable capabilities through natural self-organization principles. The nonlinear Helmholtz model achieves 40-60% parameter reduction while maintaining superior mode separation via self-focusing phenomena. The cubic-quintic dissipative wave model prevents mode collapse through balanced attractive-repulsive interactions, enabling stable soliton formation with 20-40% improved coverage. The intensity-dependent Eikonal model creates adaptive pathways that dynamically respond to content, providing enhanced controllability in conditional generation. Experimental validation demonstrates consistent superiority over linear predecessors and traditional generative approaches. The nonlinear Helmholtz model achieves FID scores of 0.0089 versus 1.0909 for linear versions, while the cubic-quintic model reaches 0.0156 FID with exceptional stability. Memory usage drops 40-60% and training time improves 30-50% due to inherent nonlinear stability properties. The framework enables bidirectional benefits, advancing both generative AI and optical physics through novel approaches to soliton analysis, wavefront control, and refractive index reconstruction with 95% accuracy. This work reveals deep connections between physical self-organization and artificial intelligence, opening pathways toward efficient optical computing implementations.

Figures

Figures reproduced from arXiv: 2506.04357 by the authors.

Figure 1
Figure 1. Time evolution of the Helmholtz equation-based generative model. Top row: Wave field at different times (t = [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Green’s functions for the Helmholtz equation (top row) and the dissipative wave equation (bottom row) [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Radial profiles of the Green’s functions for the Helmholtz equation (left) and the dissipative wave equation [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (51 more)
Figure 4
Figure 4. Figure 4: Helmholtz velocity field at t = 0.80 for an eight-Gaussian distribution. The arrows indicate the direction and [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Wave-diffusion transition at a center point for different damping coefficients ( [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Evolution of the dissipative wave equation for different damping coefficients ( [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Eikonal birth/death regions for a Gaussian bump refractive index at t = 0.80. Green regions indicate birth (R > [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Three different refractive index patterns used in the Eikonal generative model. Left: Constant refractive index. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Density evolution under the Eikonal equation for different refractive index patterns at times t = 0.20, 0.80, [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Evolution of wave amplitude distribution under the nonlinear Helmholtz equation at various Kerr nonlinearity [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Parameter space exploration showing the generative performance of the nonlinear Helmholtz model evaluated [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Comparison of wave amplitude evolution between linear ( [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: Visualization of soliton formation and stability evolution under nonlinear Helmholtz dynamics at different [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 14
Figure 14. Figure 14: Relative performance improvement (%) of the cubic-quintic nonlinear model over the cubic-only model [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]
Figure 15
Figure 15. Figure 15: Evolution of wave amplitude distributions comparing cubic-only (top row) and cubic-quintic (bottom [PITH_FULL_IMAGE:figures/full_fig_p022_15.png]
Figure 16
Figure 16. Figure 16: Cross-sectional comparison of wave amplitude profiles ( [PITH_FULL_IMAGE:figures/full_fig_p022_16.png]
Figure 17
Figure 17. Figure 17: Adaptive balancing visualization showing the dynamic equilibrium between diffusion and wave mechanisms [PITH_FULL_IMAGE:figures/full_fig_p023_17.png]
Figure 18
Figure 18. Figure 18: Multimodal interaction dynamics showing the temporal evolution of coupled field modes at six time [PITH_FULL_IMAGE:figures/full_fig_p024_18.png]
Figure 19
Figure 19. Figure 19: Average distance between modes as a function of time, showing the periodic separation and convergence [PITH_FULL_IMAGE:figures/full_fig_p025_19.png]
Figure 20
Figure 20. Figure 20: Comparison of convergence performance between Eikonal Model and Classifier Guidance approaches across [PITH_FULL_IMAGE:figures/full_fig_p027_20.png]
Figure 21
Figure 21. Figure 21: Comparison of intensity focusing behavior with and without intensity dependence at three evolutionary steps [PITH_FULL_IMAGE:figures/full_fig_p027_21.png]
Figure 22
Figure 22. Figure 22: Absolute difference between models with and without intensity dependence at the final simulation step. The [PITH_FULL_IMAGE:figures/full_fig_p028_22.png]
Figure 23
Figure 23. Figure 23: Conditional generation results showing the effect of varying refractive index conditions on wavefront [PITH_FULL_IMAGE:figures/full_fig_p030_23.png]
Figure 24
Figure 24. Figure 24: Comparison of guidance efficiency between Eikonal Model and Classifier Guidance approaches. Left [PITH_FULL_IMAGE:figures/full_fig_p031_24.png]
Figure 25
Figure 25. Figure 25: Optimal parameter regions for each optical physics model. Left: Helmholtz optimal [PITH_FULL_IMAGE:figures/full_fig_p032_25.png]
Figure 26
Figure 26. Figure 26: Computational efficiency comparison across models. Left: Generation time per sample (in ms). Right: [PITH_FULL_IMAGE:figures/full_fig_p034_26.png]
Figure 27
Figure 27. Figure 27: Comparison of generated samples for the eight-Gaussian mixture dataset. Top left: Ground truth distribution. [PITH_FULL_IMAGE:figures/full_fig_p038_27.png]
Figure 28
Figure 28. Figure 28: Effect of key parameters on Maximum Mean Discrepancy (MMD) for each optical physics model. Left: [PITH_FULL_IMAGE:figures/full_fig_p039_28.png]
Figure 29
Figure 29. Figure 29: Robustness assessment of optical physics models to initialization scale. Left: Helmholtz model showing [PITH_FULL_IMAGE:figures/full_fig_p039_29.png]
Figure 30
Figure 30. Figure 30: t-SNE visualization of features from the Helmholtz model (FID: 20.96) compared to real MNIST data. [PITH_FULL_IMAGE:figures/full_fig_p040_30.png]
Figure 31
Figure 31. Figure 31: Comprehensive comparison of generative models for MNIST: (a) Samples generated by the dissipative wave [PITH_FULL_IMAGE:figures/full_fig_p041_31.png]
Figure 32
Figure 32. Figure 32: Performance metrics of optical physics models: (a) Comparison of FID and MMD metrics across the three [PITH_FULL_IMAGE:figures/full_fig_p042_32.png]
Figure 33
Figure 33. Figure 33: Quality assessment visualizations: (a) Comprehensive quality assessment comparing optical physics models [PITH_FULL_IMAGE:figures/full_fig_p043_33.png]
Figure 34
Figure 34. Figure 34: Mode coverage and sample diversity analysis. Left: Standard deviation of min-distances showing that the [PITH_FULL_IMAGE:figures/full_fig_p044_34.png]
Figure 35
Figure 35. Figure 35: Visual comparison of samples from different generative approaches. Top row: Real data (8 Gaussians) [PITH_FULL_IMAGE:figures/full_fig_p044_35.png]
Figure 36
Figure 36. Figure 36: Verification of density positivity (C1 condition) for Helmholtz (left) and dissipative wave (right) equations. [PITH_FULL_IMAGE:figures/full_fig_p045_36.png]
Figure 37
Figure 37. Figure 37: Verification of the smoothing property (C2 condition) through dispersion relation analysis. Left: Helmholtz [PITH_FULL_IMAGE:figures/full_fig_p046_37.png]
Figure 38
Figure 38. Figure 38: Refractive index reconstruction using our framework. Top: Ground truth index (left), reconstructed index [PITH_FULL_IMAGE:figures/full_fig_p046_38.png]
Figure 39
Figure 39. Figure 39: Optimization progress (left) and cross-sectional comparison (right) during refractive index reconstruction. [PITH_FULL_IMAGE:figures/full_fig_p047_39.png]
Figure 40
Figure 40. Figure 40: Dispersion relations for different optical physics equations, showing both oscillatory behavior (left) and [PITH_FULL_IMAGE:figures/full_fig_p047_40.png]
Figure 41
Figure 41. Figure 41: Physics applications enabled by our framework: optical ray paths (top left), wavefront propagation (top [PITH_FULL_IMAGE:figures/full_fig_p048_41.png]
Figure 42
Figure 42. Figure 42: Comprehensive performance comparison between linear and nonlinear optical physics models. Top row [PITH_FULL_IMAGE:figures/full_fig_p049_42.png]
Figure 43
Figure 43. Figure 43: Mode coverage improvement achieved by nonlinear optical physics models. The Eikonal model shows [PITH_FULL_IMAGE:figures/full_fig_p049_43.png]
Figure 44
Figure 44. Figure 44: MNIST samples generated by linear and nonlinear optical physics models. Top row: Linear Helmholtz, [PITH_FULL_IMAGE:figures/full_fig_p050_44.png]
Figure 45
Figure 45. Figure 45: Normalized performance metrics heatmap comparing all optical physics model variants. Yellow indicates [PITH_FULL_IMAGE:figures/full_fig_p051_45.png]
Figure 46
Figure 46. Figure 46: Direct comparison of generated probability density distributions for the eight-Gaussian mixture dataset. [PITH_FULL_IMAGE:figures/full_fig_p052_46.png]
Figure 47
Figure 47. Figure 47: Parameter sensitivity analysis across the three optical physics models. Left: Helmholtz model showing FID [PITH_FULL_IMAGE:figures/full_fig_p053_47.png]
Figure 48
Figure 48. Figure 48: Comparative generation results for the eight-Gaussian mixture dataset. Left: Ground truth distribution [PITH_FULL_IMAGE:figures/full_fig_p054_48.png]
Figure 49
Figure 49. Figure 49: Direct comparison of density evolution patterns between linear and nonlinear optical models. Top row [PITH_FULL_IMAGE:figures/full_fig_p055_49.png]
Figure 50
Figure 50. Figure 50: Training time comparison across dimensionality for optical physics models. The nonlinear Helmholtz model [PITH_FULL_IMAGE:figures/full_fig_p057_50.png]
Figure 51
Figure 51. Figure 51: Memory usage scaling across different dimensionalities. Optical physics models demonstrate superior [PITH_FULL_IMAGE:figures/full_fig_p057_51.png]
Figure 52
Figure 52. Figure 52: Generation steps required for different distribution complexities. For moderate complexity distributions [PITH_FULL_IMAGE:figures/full_fig_p058_52.png]
Figure 53
Figure 53. Figure 53: Comprehensive efficiency metrics heatmap comparing optical physics models with diffusion baselines. [PITH_FULL_IMAGE:figures/full_fig_p059_53.png]
Figure 54
Figure 54. Figure 54: Relative efficiency comparison showing performance ratios relative to diffusion models across dimensionality. [PITH_FULL_IMAGE:figures/full_fig_p060_54.png]

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. Strategic Alignment Patterns in National AI Policies

    cs.CY 2025-07 reject novelty 4.0 of 10

    A policy-analysis preprint scores alignment between objectives, foresight, and instruments in 15-20 national AI strategies, claiming distinct governance-based archetypes, but ships no data, figures, or code to support...

Reference graph

Works this paper leans on

47 extracted references · 28 canonical work pages · cited by 1 Pith paper

  1. [1]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401–4410, 2019

    2019

  2. [2]

    Tacnet: Temporal audio source counting network.arXiv preprint arXiv:2311.02369, 2023

    Amirreza Ahmadnejad, Ahmad Mahmmodian Darviishani, Mohmmad Mehrdad Asadi, Sajjad Saffariyeh, Pe- dram Yousef, and Emad Fatemizadeh. Tacnet: Temporal audio source counting network.arXiv preprint arXiv:2311.02369, 2023

    arXiv 2023

  3. [3]

    Language models are few-shot learners.Advances in Neural Information Processing Systems, 33:1877–1901, 2020

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Nee- lakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.Advances in Neural Information Processing Systems, 33:1877–1901, 2020

    1901

  4. [4]

    Highly accurate protein structure prediction with alphafold.Nature, 596(7873):583–589, 2021

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold.Nature, 596(7873):583–589, 2021

    2021

  5. [5]

    Learning to simulate complex physics with graph networks

    Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. InInternational Conference on Machine Learning, pages 8459–8468, 2020

    work page 2020

  6. [6]

    Dynamic control of spontaneous emission using magnetized insb higher-order- mode antennas.Journal of Physics: Photonics, 6(3):035011, 2024

    Sina Aghili, Rasoul Alaee, Amirreza Ahmadnejad, Ehsan Mobini, Mohammadreza Mohammadpour, Carsten Rockstuhl, and Ksenia Dolgaleva. Dynamic control of spontaneous emission using magnetized insb higher-order- mode antennas.Journal of Physics: Photonics, 6(3):035011, 2024

    work page 2024

  7. [7]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InInternational Conference on Machine Learning, pages 2256–2265, 2015

    work page 2015

  8. [8]

    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

Show all 47 references
  1. [10]

    Poisson flow generative models.arXiv preprint arXiv:2209.11178, 2022

    Yang Xu, Ziming Liu, Max Tegmark, and Tommi Jaakkola. Poisson flow generative models.arXiv preprint arXiv:2209.11178, 2022

    2022 arXiv

  2. [11]

    Generative adversarial nets.Advances in Neural Information Processing Systems, 27:2672– 2680, 2014

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets.Advances in Neural Information Processing Systems, 27:2672– 2680, 2014

    2014

  3. [12]

    Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    2013 arXiv

  4. [13]

    Genphys: From physical processes to generative models.arXiv preprint arXiv:2304.02637, 2023

    Ziming Liu, Di Luo, Yang Xu, Tommi Jaakkola, and Max Tegmark. Genphys: From physical processes to generative models.arXiv preprint arXiv:2304.02637, 2023

    2023 arXiv

  5. [14]

    Deep learning meets nanophotonics: A generalized accurate predictor for near fields and far fields of arbitrary 3d nanostructures.Nano Letters, 20(1):329–338, 2022

    Peter R Wiecha and Otto L Muskens. Deep learning meets nanophotonics: A generalized accurate predictor for near fields and far fields of arbitrary 3d nanostructures.Nano Letters, 20(1):329–338, 2022

    2022

  6. [15]

    Free-space optical spiking neural network.PloS one, 19(12):e0313547, 2024

    Reyhane Ahmadi, Amirreza Ahmadnejad, and Somayyeh Koohi. Free-space optical spiking neural network.PloS one, 19(12):e0313547, 2024. 62 Optical Physics-Based Generative ModelsA PREPRINT

    2024

  7. [16]

    Nontrapping tunable topological photonic memory.arXiv preprint arXiv:2502.19398, 2025

    Amirreza Ahmadnejad, Somayyeh Koohi, and Abolhassan Vaezi. Nontrapping tunable topological photonic memory.arXiv preprint arXiv:2502.19398, 2025

    2025 arXiv

  8. [17]

    Training large-scale optical neural networks with two-pass forward propagation.arXiv preprint arXiv:2408.08337, 2024

    Amirreza Ahmadnejad and Somayyeh Koohi. Training large-scale optical neural networks with two-pass forward propagation.arXiv preprint arXiv:2408.08337, 2024

    2024 arXiv

  9. [18]

    All-optical doubly resonant cavities for relu function in nanophotonic deep learning.arXiv preprint arXiv:2504.19692, 2025

    Amirreza Ahmadnejad, Mohmmad Mehrdad Asadi, and Somayyeh Koohi. All-optical doubly resonant cavities for relu function in nanophotonic deep learning.arXiv preprint arXiv:2504.19692, 2025

    2025 arXiv

  10. [19]

    Optical physics based generative models

    Amirreza Ahmadnejad. Optical physics based generative models. https://github.com/AAhmadnejad98/ Optical-Physics-Based-Generative-Models, 2024. Accessed: 2025-04-24

    2024

  11. [20]

    Unbalanced minibatch optimal transport; applications to domain adaptation.arXiv preprint arXiv:2103.03606, 2021

    Kilian Fatras, Thibault Séjourné, Nicolas Courty, and Rémi Flamary. Unbalanced minibatch optimal transport; applications to domain adaptation.arXiv preprint arXiv:2103.03606, 2021

    2021 arXiv

  12. [21]

    Unbalanced sobolev descent.arXiv preprint arXiv:2009.14148, 2020

    Youssef Mroueh and Mattia Rigotti. Unbalanced sobolev descent.arXiv preprint arXiv:2009.14148, 2020

    2009 arXiv

  13. [22]

    Accelerating langevin sampling with birth-death.arXiv preprint arXiv:1905.09863, 2019

    Yulong Lu, Jianfeng Lu, and James Nolen. Accelerating langevin sampling with birth-death.arXiv preprint arXiv:1905.09863, 2019

    1905 arXiv

  14. [23]

    Cambridge University Press, 2016

    Richard M Martin, Lucia Reining, and David M Ceperley.Interacting Electrons: Theory and Computational Approaches. Cambridge University Press, 2016

    2016

  15. [24]

    Cambridge University Press, 7 edition, 2013

    Max Born and Emil Wolf.Principles of Optics: Electromagnetic Theory of Propagation, Interference and Diffraction of Light. Cambridge University Press, 7 edition, 2013

    2013

  16. [25]

    Green’s function for the lossy wave equation.Revista Brasileira de Ensino de Física, 30:1302.1–1302.5, 2008

    R Aleixo and E Capelas de Oliveira. Green’s function for the lossy wave equation.Revista Brasileira de Ensino de Física, 30:1302.1–1302.5, 2008

    2008

  17. [26]

    Cambridge University Press, 7 edition, 1999

    Max Born and Emil Wolf.Principles of Optics: Electromagnetic Theory of Propagation, Interference and Diffraction of Light. Cambridge University Press, 7 edition, 1999

    1999

  18. [27]

    On the partial difference equations of mathematical physics

    Richard Courant, Kurt Friedrichs, and Hans Lewy. On the partial difference equations of mathematical physics. IBM Journal of Research and Development, 11(2):215–234, 1967

    1967

  19. [28]

    Cambridge University Press, 1999

    James A Sethian.Level set methods and fast marching methods: evolving interfaces in computational geometry, fluid mechanics, computer vision, and materials science, volume 3. Cambridge University Press, 1999

    1999

  20. [29]

    Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  21. [30]

    Attention is all you need.Advances in Neural Information Processing Systems, 30:5998–6008, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in Neural Information Processing Systems, 30:5998–6008, 2017

    2017

  22. [31]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234–241, 2015

    2015

  23. [32]

    Practical bayesian optimization of machine learning algorithms

    Jasper Snoek, Hugo Larochelle, and Ryan P Adams. Practical bayesian optimization of machine learning algorithms. InAdvances in Neural Information Processing Systems, volume 25, pages 2951–2959, 2012

    2012

  24. [33]

    Jax: composable transforma- tions of python+numpy programs

    James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. Jax: composable transforma- tions of python+numpy programs. 2018

    2018

  25. [34]

    Neural ordinary differential equations

    Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. InAdvances in Neural Information Processing Systems, volume 31, 2018

    2018

  26. [35]

    Ffjord: Free-form continuous dynamics for scalable reversible generative models.arXiv preprint arXiv:1810.01367, 2018

    Will Grathwohl, Ricky TQ Chen, Jesse Bettencourt, Ilya Sutskever, and David Duvenaud. Ffjord: Free-form continuous dynamics for scalable reversible generative models.arXiv preprint arXiv:1810.01367, 2018

    2018 arXiv

  27. [36]

    Pfgm++: Unlocking the potential of physics-inspired generative models.arXiv preprint arXiv:2302.04265, 2023

    Yilun Xu, Ziming Liu, Yonglong Tian, Shangyuan Tong, Max Tegmark, and Tommi Jaakkola. Pfgm++: Unlocking the potential of physics-inspired generative models.arXiv preprint arXiv:2302.04265, 2023

    2023 arXiv

  28. [37]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in Neural Information Processing Systems, 30, 2017

    Martin Heusel, Huber Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in Neural Information Processing Systems, 30, 2017

    2017

  29. [38]

    A kernel two-sample test.The Journal of Machine Learning Research, 13:723–773, 2012

    Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bernhard Schölkopf, and Alexander Smola. A kernel two-sample test.The Journal of Machine Learning Research, 13:723–773, 2012. 63 Optical Physics-Based Generative ModelsA PREPRINT

    2012

  30. [39]

    Elucidating the design space of diffusion-based generative models.arXiv preprint arXiv:2206.00364, 2022

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models.arXiv preprint arXiv:2206.00364, 2022

    2022 arXiv

  31. [40]

    Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 1998

    Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 1998

    1998

  32. [41]

    Visualizing data using t-sne.Journal of Machine Learning Research, 9:2579–2605, 2008

    Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne.Journal of Machine Learning Research, 9:2579–2605, 2008

    2008

  33. [42]

    Diffusion schrödinger bridge with applications to score-based generative modeling.arXiv preprint arXiv:2106.01357, 2021

    Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schrödinger bridge with applications to score-based generative modeling.arXiv preprint arXiv:2106.01357, 2021

    2021 arXiv

  34. [43]

    Learning approach to optical tomography.Optica, 2(6):517–522, 2015

    Ulugbek S Kamilov, Ioannis N Papadopoulos, Morteza H Shoreh, Alexandre Goy, Cedric V onesch, Michael Unser, and Demetri Psaltis. Learning approach to optical tomography.Optica, 2(6):517–522, 2015

    2015

  35. [44]

    Photonics for artificial intelligence and neuromorphic computing.Nature Photonics, 15(2):102–114, 2021

    Bhavin J Shastri, Alexander N Tait, Thomas Ferreira de Lima, Wolfram HP Pernice, Harish Bhaskaran, C David Wright, and Paul R Prucnal. Photonics for artificial intelligence and neuromorphic computing.Nature Photonics, 15(2):102–114, 2021

    2021

  36. [45]

    Inference in artificial intelligence with deep optics and photonics.Nature, 588(7836):39–47, 2020

    Gordon Wetzstein, Aydogan Ozcan, Sylvain Gigan, Shanhui Fan, Dirk Englund, Marin Soljaˇci´c, Cornelia Denz, David AB Miller, and Demetri Psaltis. Inference in artificial intelligence with deep optics and photonics.Nature, 588(7836):39–47, 2020

    2020

  37. [46]

    All-optical machine learning using diffractive deep neural networks.Science, 361(6406):1004–1008, 2018

    Xing Lin, Yair Rivenson, Nezih T Yardimci, Muhammed Veli, Yi Luo, Mona Jarrahi, and Aydogan Ozcan. All-optical machine learning using diffractive deep neural networks.Science, 361(6406):1004–1008, 2018

    2018

  38. [47]

    Spatiotemporal light control with active metasurfaces.Science, 364(6441), 2019

    Amr M Shaltout, Vladimir M Shalaev, and Mark L Brongersma. Spatiotemporal light control with active metasurfaces.Science, 364(6441), 2019. 64 Optical Physics-Based Generative ModelsA PREPRINT Appendices A Detailed Derivations S1.1 Green’s Function Derivations S1.1.1 Helmholtz ...

    2019

  39. [48]

    0", "6", and

    ˜G(k, t;x′) =δ(t)e −ik·x′ (S2) Fort >0, the general solution is: ˜G(k, t;x′) = ( A(k)ei √ k2 0−|k|2t +B(k)e −i √ k2 0−|k|2t for|k|< k 0 C(k)e − √ |k|2−k2 0t +D(k)e √ |k|2−k2 0t for|k|> k 0 (S3) The boundedness condition as t→ ∞requires B(k) = 0 for |k|< k0 and D(k) = 0 for |k|...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.