{"id":"cad8ccf0-d34a-4c18-859c-e4f323f6e07b","arxiv_id":"2505.01917","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A discrete diffusion model whose forward and reverse processes are random-walk jumps of intensity particles on the spatial lattice, guaranteeing exact per-sample conservation of total intensity.","lead":"Discrete Spatial Diffusion is a new image- and microstructure-generating method in which brightness is treated as a collection of tiny particles that hop randomly between neighboring pixels, so the total brightness of every generated image is preserved exactly. It is worth reading because it gives scientists a way to generate synthetic images that obey strict mass or phase conservation, which standard diffusion models cannot do.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unless post-processing is applied, DSD's binary and ternary microstructure outputs can contain overlapping particles, which breaks the physical phase-volume interpretation and the reported scientific metrics.","rationale":"The theoretical core of DSD is sound: the forward process is a standard independent-particle jump process, the reverse rates in Eq. (4) follow from time reversal of the labeled Markov process, and total intensity per channel is preserved exactly by construction in both training and sampling. I therefore do not object to the conservation statement itself. The concern I identify is about the scientific application: for binary and ternary microstructure data, the support of the training distribution is single-phase-per-pixel, but the DSD state space and sampling algorithm do not enforce that support. The paper does not state how generated samples are converted back to valid segmented images, and the unexplained Table 3 note about 'tolerance avoids overlapping particles' is not a mechanism in Algorithm 2. If final samples contain overlapping particles, the claimed exact phase-volume conservation and the reported microstructure metrics are not physically meaningful. This is an empirical check, not a logical contradiction, and it reinforces rather than changes the reader's CONDITIONAL verdict.","tokens_in":22989,"tokens_out":20053,"duration_ms":248460,"concrete_test":"Run the exact generation pipeline for the Berea, Savonnières, Massangis, and NMC models without any post-processing, then count the fraction of pixels with intensity >1 for the rock models and the fraction of pixels with both color channels positive for the electrode model. If the fraction is not effectively zero, recompute the Appendix F and G metrics after mapping each invalid pixel to a legal single-phase label; if the PSD, two-point correlation, or TauFactor values change materially, the scientific conservation and performance claims are not supported as stated. If the overlap fraction is zero, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sec. 3.1, Eq. (1) defines independent particle hops with no exclusion rule, so the forward process maps binary phase images into states with arbitrary particle counts per pixel. Algorithm 2 samples moves per pixel via binomial/multinomial draws and never enforces the one-phase-per-pixel constraint of the training data. For the rock experiments (Sec. 4.2) and electrode experiments (Sec. 4.3), final generated images therefore may contain pixels with intensity >1 (binary rocks) or with both carbon-binder and active-material channels occupied (electrodes). The paper claims exact preservation of mass fractions, but phase volume fraction is only well-defined on valid single-phase images; if such overlaps are present, total intensity per channel no longer equals phase volume, and the PoreSpy/TauFactor metrics computed on overlapping states, or after unspecified thresholding, are not reliable evidence for the claimed transport and electrochemical fidelity. Table 3's note that a CFL 'tolerance avoids overlapping particles' is unexplained and does not follow from the τ-leaping algorithm, which controls step size but not co-occupancy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Discrete Spatial Diffusion (DSD), a continuous-time discrete-state diffusion framework in which intensity units are treated as particles performing nearest-neighbor random walks on the image lattice. The forward process conserves total particle count per color channel, and the reverse process is defined through per-particle reverse transition rates that are aggregated via survival analysis in Eq. (4). The authors train a neural network to predict reverse rates using either a rate-matching or a likelihood loss, and sample via a binomial tau-leaping scheme with an adaptive CFL-based step size. Experiments cover MNIST inpainting and class conditioning, unconditional CIFAR-10 and CelebA generation, and two scientific applications: binary rock microstructure generation and two-channel lithium-ion battery electrode generation. The central claim is that DSD exactly preserves total intensity by construction, enabling constrained generation for scientific microstructure modeling.","tokens_in":23130,"tokens_out":11544,"duration_ms":131428,"significance":"If the construction holds, DSD is a useful contribution to discrete-state diffusion modeling: it introduces spatial correlations into the noise process while preserving an exact global invariant, which is rare among diffusion-based generative models. The reverse-rate derivation in Appendix A via survival analysis is a clean way to aggregate independent-particle rates, and the spectral solution for periodic boundaries in Appendix C is elegant. The paper also demonstrates the method on scientifically relevant datasets, which is valuable for the microstructure community. The main caveats are that the scientific validation assumes generated images are valid single-phase configurations, an assumption not enforced by the sampling algorithm, and that several implementation details in the pseudocode and appendices need correction.","major_comments":[{"comment":"The forward process in Eq. (1) and the sampling in Algorithm 2 do not enforce single-phase occupancy, so generated binary rock images can contain pixels with intensity greater than 1, and generated electrode images can have both the carbon-binder and active-material channels active in the same pixel. Because no thresholding or exclusion post-processing is described for the scientific datasets, the PoreSpy and TauFactor metrics in Figures 16 and 19 may be computed on physically invalid phase configurations. This undermines the claim that the generated microstructures are realistic under strict mass conservation, since total intensity per channel only equals phase volume fraction when the image is a valid phase map. The Table 3 note that the CFL 'tolerance avoids overlapping particles' is not supported by Algorithm 2, because the CFL condition only bounds the per-step jump probability and does not prevent a pixel from receiving particles from multiple neighbors or from multiple channels.","section":"Sec. 4.2/4.3 and Table 3"},{"comment":"In Algorithm 2, the binomial sampling step draws n_Sigma from Binom([I_t]_{x,y,c}, sum_nu r_NN_nu), where r_NN_nu are rates with units of inverse time. As written, the second argument is not a probability: it omits the factor tau that was just computed by the CFL rule, and it is not normalized by the number of particles at the pixel. A correct binomial tau-leaping step would use a probability of the order of (tau / [I_t]_{x,y,c}) times the total rate, for example p = 1 - exp(-tau * sum_nu r_NN_nu / [I_t]_{x,y,c}). This is the central sampling step, so the pseudocode must be corrected and the implementation checked against it.","section":"Algorithm 2"},{"comment":"The normalization of the discrete Fourier pair is inconsistent. With the unitary convention in Eq. (14), a single particle at the origin gives k_{m,n}(0) = 1/sqrt(N), not 1/N as stated in Eq. (21). Consequently, Eq. (22) does not satisfy p_{0,0}(0) = 1 with the stated prefactor 1/sqrt(N^3). Since these transition probabilities are used in Algorithm 1 to corrupt images and to compute the rates in Eq. (4), the normalization error must be fixed. The prefactor cancels in the rate ratios of Eq. (4), but it matters when sampling forward trajectories from pt.","section":"Appendix C, Eqs. (14), (21), (22)"},{"comment":"The FID of 20.6 reported for the 'CFL epsilon = 0.01 + filtering + classifier thresholding' row is computed on only 21,780 of the 50,000 generated images, after discarding samples with classifier confidence below 0.99. This is not comparable to standard FID scores computed on all generated samples, and the paper should either report the metric on the full set of 50,000 samples under the same filtering rule or clearly label the retained-subset score as an oracle-style upper bound. The same concern applies to the sFID value in that row.","section":"Table 1 and Sec. D.2"}],"minor_comments":[{"comment":"The schedule equation is difficult to parse: the left-hand side uses t_k but t_k is defined only implicitly by the equation, and the notation 'Phi(e^{-tau2 tk})' should be written as a function of t_k. Also, the definition of Phi should be Phi(p) = log(p/(1-p)), with parentheses, to avoid ambiguity.","section":"Sec. 3.2, Eq. (2)"},{"comment":"The sign of the likelihood loss appears inconsistent: Eq. (6) defines -log L = -E[integral(...)], while Eq. (7) states log L = E[(t_k - t_{k-1}) sum(...)]. Please check the overall sign so that the loss is nonnegative and the equations are consistent.","section":"Sec. 3.4, Eqs. (6)-(7)"},{"comment":"The step size line 'tau <- min{t, epsilon min_{...} (r_NN)^{-1}}' is correct as epsilon divided by the maximum rate, but the use of 'min' over the inverse rates is easy to misread. Writing tau = min{t, epsilon / max_{...} r_NN} would be clearer.","section":"Algorithm 2"},{"comment":"The comparison of the rock FID of 0.9 with a literature value of 18.1 from Lee and Yun is acknowledged not to be one-to-one; this comparison should be moved to the appendix or marked as indicative rather than quantitative.","section":"Sec. F.2"},{"comment":"No code repository or seed information is provided. Given the number of algorithm choices and the pseudocode issues above, a code release with the exact sampling and post-processing steps would substantially improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The overlap issue in the scientific experiments and the Algorithm 2 binomial-probability error are the two most important concerns; both are fixable in revision. I would also ask the authors to clarify the Appendix C normalization, since that formula feeds into the forward noising step. A code release would help confirm that the corrected pseudocode matches the actual implementation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should read this paper for the central mechanism: DSD moves discrete intensity units as independent nearest-neighbor random walkers, and total particle count per channel is conserved by construction. That is genuinely new for diffusion models, and it fits physical data with hard mass/volume constraints. The reverse-time rate derivation in Appendix A is sound under the stated independence assumption, and the survival-analysis argument for multi-particle pixels is neat. The fact that rate-matching and likelihood training behave similarly is also a useful empirical data point.\n\nCredit where due: MNIST inpainting and intensity-conditioning show the exact constraint working; the rock and electrode samples look plausible; and the authors are transparent about the heuristic schedule and cost scaling. Related work is credited fairly, including their own Blackout Diffusion, which is the natural lineage.\n\nThe main soft spot is that the forward process has no exclusion rule, so a binary rock image can evolve into multi-particle pixels, and the reverse process never explicitly enforces single-phase occupancy. The generated \"binary\" microstructures may therefore contain overlapping particles, in which case total particle count no longer equals phase volume fraction. The Table 3 note that the CFL tolerance \"avoids overlapping particles\" does not follow from the algorithm; CFL only limits per-step jump probability, not co-occupancy. If outputs are thresholded before computing PoreSpy/TauFactor metrics, that has to be stated; otherwise those metrics are not reliable evidence for transport or phase fidelity. This is the key issue to fix. Second, the CIFAR-10 FID of 20.6 uses bilateral filtering and discards 56% of samples via classifier confidence, so it is not an unconditional generation score and should not be compared to standard FID numbers. Third, no code or data are released, which matters for a method whose selling point is exactness. Fourth, the abstract says \"electrochemical performance\" but the electrode metrics are morphological proxies (interface length, triple-phase boundary, relative diffusivity), so the wording overstates what was measured.\n\nThat said, the core idea holds up and is worth engaging with. I would send this to serious peer review with a request for major revision: show how binary/ternary validity is achieved (or post-process), report unconditional FID without sample selection, and release code/data. If those are addressed, this becomes a solid paper for the discrete-diffusion and microstructure-generation communities.","headline":"DSD's spatial random-walk diffusion with exact intensity conservation is a real contribution, but the no-exclusion dynamics can produce overlapping phases, and the scientific metrics are computed without showing the outputs are valid single-phase images.","tokens_in":23695,"tokens_out":4478,"would_cite":true,"duration_ms":43717,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Diffusion models can conserve total image intensity exactly by treating each intensity unit as a particle that hops across the pixel lattice, and the approach extends to rock and battery-electrode microstructures.","keywords":["discrete-state diffusion","intensity conservation","jump process","lattice random walk","mass conservation","microstructure generation","image inpainting","generative modeling"],"falsifier":"Measure the conservation claim directly by summing per-channel intensity before and after full reverse-time sampling; it should match the conditioning value to machine precision, since exactness holds by construction. To test the independence premise that carries the method, build a training set whose defining feature is an interaction between intensity units that a single-particle random walk cannot express — for example, binary images with a hard rule that occupied pixels never share a diagonal neighbor — train DSD on it, and compare the generated samples' two-point correlation with the training set's; a measurable mismatch while total intensity is exactly right would identify the independence assumption as the limiting factor.","tokens_in":22744,"feed_emoji":"🎲","tokens_out":16281,"duration_ms":149311,"temperature":0.7,"pith_summary":"The paper sets out to establish that generative diffusion models can obey a hard conservation law: the total number of intensity units (particles) in each color channel is preserved exactly, in every sample, through both the forward corruption and the reverse generation process. Earlier discrete diffusion models randomize each pixel's intensity independently, so global quantities such as mass can at best be matched on average or forced afterward by projection; DSD instead lets every unit of intensity perform a continuous-time random walk on the spatial lattice, making total intensity conserved by construction while a neural network learns the reverse hopping rates. The paper validates the approach on MNIST, CIFAR-10, and CelebA, including inpainting, class conditioning, and exact total-intensity conditioning, and on two scientific applications: porous rock microstructures and lithium-ion battery electrodes, where generated samples hit target porosity or phase volume fractions exactly and reproduce transport-relevant statistics. If correct, this gives scientists a generative model whose outputs respect conservation laws automatically rather than approximately.","feed_headline":"Exact mass preservation comes to diffusion models","feed_subtitle":"Each intensity unit is a hopping particle, so rock porosity and battery phase fractions come out exactly right.","key_machinery":"The central object is a Markov jump process on a discrete lattice in which each unit of intensity is an independent particle hopping to one of four nearest neighbors at rate $r$ (Eq. 1); under periodic boundaries its transition probabilities are computed exactly by exponentiating the Markov matrix, which the discrete Fourier transform diagonalizes (Eq. 22). The load-bearing identity is Eq. 4: the reverse-time rate for the first of $n$ particles to leave a pixel is the sum, over those $n$ particles, of the per-particle reverse rates, each proportional to the forward transition probability ratio, a result derived from survival analysis of the independent-particle system. Around this identity the paper builds a training scheme — an NCSN++ network with four directional outputs and a SoftPlus nonlinearity, trained with either rate-matching or process-likelihood loss — and a sampler based on binomial $\\tau$-leaping with an adaptive step size capped by a Courant–Friedrichs–Lewy condition, so that no transition probability exceeds a fixed tolerance at any step.","core_discovery":"At the center of the paper is the claim that a diffusion model built on a continuous-time, discrete-state jump process — each unit of image intensity is a particle hopping to a nearest neighbor on the pixel lattice at a constant rate $r$ — preserves total intensity per color channel exactly in both the forward and reverse phases, and that this exactly constrained process still learns realistic structure. The reverse-time transition rate for a pixel holding $n$ particles is the sum of per-particle rates, each proportional to a ratio of forward transition probabilities (Eq. 4), obtained by survival analysis of the independent-particle system, and a neural network is trained to predict these rates from corrupted images. On MNIST, CIFAR-10, and CelebA the paper reports coherent samples, with the conserved total intensity acting as a semantic control: raising the allowed particle count changes a thin digit into a bold one. On three rock types and NMC battery cathodes, DSD-generated microstructures reproduce two-point correlations, pore size distributions, interface length, triple-phase boundary, and relative diffusivity of the training data while exactly hitting porosity or phase-fraction targets.","pith_inferences":["Beyond the paper: the independence assumption predicts a concrete failure mode — training data whose structure lives in interactions between intensity units (for instance, a rule against adjacent occupied pixels) should be reproduced only approximately, since the forward random walk erases exactly that structure; an experiment comparing two-point statistics of real and generated samples on such da","Beyond the paper: the exact conservation law covers per-channel particle count only, so cross-channel invariants such as stoichiometric ratios or color balance are not pinned down; extending the construction to other additive conserved quantities is the natural next step that the paper's closing remark gestures toward.","Beyond the paper: in the fully corrupted limit the particle configuration is the maximally spread layout with fixed total count, so DSD is most at home when the conserved quantity is the physically dominant constraint — the reported CIFAR-10 FID near 20 (after filtering and classifier screening) plausibly reflects the price of hard conservation for texture-rich natural images."],"forward_implications":["Total intensity per color channel is conserved exactly in every generated sample, making quantities such as porosity, mass fraction, or digit stroke area hard constraints that hold identically, not merely on average.","Conditioning on total intensity is exact by construction at every intensity value, including far outside the typical training range, where the paper shows Gaussian-diffusion conditioning fails at the distribution tails.","The same machinery transfers from natural images to scientific microstructures: generated rocks and battery electrodes match training-set two-point correlations, pore size distributions, and electrochemical transport metrics while hitting exact porosity or phase volume fractions.","Computational cost scales linearly with the total intensity of the image, which makes the method inexpensive for binary and low-bit-depth scientific data and increasingly costly for high-intensity natural images.","Because the forward process is a plain lattice random walk, the authors state that extension to 3D requires only implementation changes, and they argue the framework opens the door to noise processes that exploit other conservation laws and symmetries."],"supporting_citations":[{"why":"Supplies the reverse-time formalism for discrete-state Markov processes and the observation-time schedule construction that DSD adapts.","marker":"[63]"},{"why":"Establishes the continuous-time discrete-diffusion framework whose reverse-time process and likelihood loss DSD builds on.","marker":"[11]"},{"why":"Provides the DDPM training paradigm and the heuristic-versus-likelihood loss comparison that motivates the two DSD losses.","marker":"[36]"},{"why":"Supplies the NCSN++ architecture and class-conditioning mechanism used in every DSD experiment.","marker":"[69]"},{"why":"Introduces the heat-dissipation corruption model that DSD generalizes; the random walk is its microscopic description.","marker":"[61]"},{"why":"Shows spatially structured (blurring) diffusion reaches competitive generative quality, supporting DSD's spatial-noise design.","marker":"[38]"},{"why":"Provides tau-leaping, the stochastic simulation algorithm on which DSD's reverse-time sampler is based.","marker":"[31]"},{"why":"Supplies the Courant–Friedrichs–Lewy condition used to fix the adaptive reverse-time step size.","marker":"[21]"},{"why":"Defines the SSIM metric whose even degradation across observation times shapes the DSD noise schedule.","marker":"[76]"}],"fun_headline_variants":["Particle-conserving diffusion: exact mass, real microstructures","Diffusion with exact mass: particles hop, structure emerges","Mass-preserving jumps: diffusion that conserves intensity","Exact intensity conservation from discrete spatial diffusion","Battery and rock microstructures from conserved diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each unit of intensity moves independently of all the others, so that the reverse-time rate for a pixel holding many particles (Eq. 4) is just a sum of single-particle rates; any interaction structure in the data that this independence cannot express must be learned only approximately by the network.","fun_headline_variants_meta":{"raw":{"variants":["Particle-conserving diffusion: exact mass, real microstructures","Diffusion with exact mass: particles hop, structure emerges","Mass-preserving jumps: diffusion that conserves intensity","Exact intensity conservation from discrete spatial diffusion","Battery and rock microstructures from conserved diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000721,"raw_usage":{"total_tokens":3240,"prompt_tokens":956,"completion_tokens":2284,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":2221}},"tokens_in":572,"tokens_out":2284,"duration_ms":17188,"temperature":1.0,"reasoning_tokens":2221,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:07:59.525600+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the conservation claim directly by summing per-channel intensity before and after full reverse-time sampling; it should match the conditioning value to machine precision, since exactness holds by construction. To test the independence premise that carries the method, build a training set whose defining feature is an interaction between intensity units that a single-particle random walk cannot express — for example, binary images with a hard rule that occupied pixels never share a diagonal neighbor — train DSD on it, and compare the generated samples' two-point correlation with the training set's; a measurable mismatch while total intensity is exactly right would identify the independence assumption as the limiting factor.","supporting_citations":[{"cited_title":"Santos, Zachary R","cited_arxiv_id":null,"evidence_quote":"Supplies the reverse-time formalism for discrete-state Markov processes and the observation-time schedule construction that DSD adapts."},{"cited_title":"A continuous time framework for discrete denoising models","cited_arxiv_id":null,"evidence_quote":"Establishes the continuous-time discrete-diffusion framework whose reverse-time process and likelihood loss DSD builds on."},{"cited_title":"Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole","cited_arxiv_id":null,"evidence_quote":"Supplies the NCSN++ architecture and class-conditioning mechanism used in every DSD experiment."},{"cited_title":"Blurring Diffusion Models","cited_arxiv_id":null,"evidence_quote":"Shows spatially structured (blurring) diffusion reaches competitive generative quality, supporting DSD's spatial-noise design."},{"cited_title":"Gillespie","cited_arxiv_id":null,"evidence_quote":"Provides tau-leaping, the stochastic simulation algorithm on which DSD's reverse-time sampler is based."},{"cited_title":"Bovik, H.R","cited_arxiv_id":null,"evidence_quote":"Defines the SSIM metric whose even degradation across observation times shapes the DSD noise schedule."}],"review_version":1}