{"id":"05818bee-a0e4-4eea-b4a3-9eff3480f862","arxiv_id":"2505.06804","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Diffusion model sampling guided by SIREN Jacobian signals can place user-specified critical points at chosen locations in generated vector fields while keeping outputs close to the training distribution.","lead":"Topology guidance lets a user steer an existing diffusion model of vector fields so that generated fields contain user-specified critical points, such as sinks, sources, and saddles, at chosen locations. The value is a direct control handle for exploring simulation ensembles and comparing flow distributions without retraining the model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stability/type guidance is not well-defined for complex eigenvalues and Eq. 13's sign convention drives the wrong direction, so the method as written cannot produce the reported focus/stable configurations.","rationale":"The reader's weakest_assumption focused on the in-distribution status of \\hat z_t for the SIREN, which is a real concern and is partially acknowledged by the authors. However, I found a more definite gap: the algorithm as written is internally inconsistent for the focus and 'stable/unstable' specifications that are part of the central claim. The empirical tables are not enough to dismiss this, because the reported alignments may have been produced by an energy that is not fully described in Eqs. 7-15. This is not an accusation of bad faith; rather, it is a reproducibility issue that a careful reader cannot resolve from the text. The reader already noted the Sec. 4.3 misstatement, so we partially agree, but the reader's primary weakest assumption was different. Since the verdict was already CONDITIONAL and this concern is addressable by a precise rewriting of the energy functions, I do not recommend changing the verdict. A concrete implementation test will settle whether the stated equations alone can produce the claimed focus and stability control.","tokens_in":18315,"tokens_out":10359,"duration_ms":112265,"concrete_test":"Implement Eqs. 7-14 exactly as written, with no additional terms, and run the Tab. 3 'stable sink' and 'unstable source' configurations. Record whether the loss is real-valued at every DDIM step and whether the final decoded field has a critical point at p whose Jacobian has the prescribed real-part signs. Then run the same configurations with an additional σ(Δ) term or explicit real-part handling as Fig. 4 suggests. If the reported 82-97% alignment is only achievable in the second run, the paper's stated equations are insufficient to support its central claim, and the missing energy must be specified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim includes control of critical point type and stability. As written, Secs. 4.2-4.4 cannot implement this control for the configurations reported in Tab. 3 and Fig. 4. First, Eq. 10 classifies sinks, sources, and saddles by ordering λ1 and λ2, which is only meaningful for real eigenvalues. Focus configurations require complex conjugate eigenvalues (Δ<0 in Eq. 12); substituting complex λ into the sigmoid in Eq. 11 does not yield a real objective, so the energy is undefined for foci. Second, Sec. 4.3 misstates the meaning of Δ: Δ=Tr(J)^2-4|J| distinguishes nodes from foci, not stable from unstable; stability is determined by the real parts of the eigenvalues. Third, Eq. 13, with β_s=1 for 'unstable' and β_s=-1 for 'stable', minimizes σ(β_s·λ1). For a 'stable' point the minimization drives λ1 positive, i.e., toward an unstable source, the opposite of the standard stability convention. Eq. 14 also omits any σ(Δ) term, although Fig. 4 shows discriminant-based terms for focus configurations. Thus the energy functions stated in the paper do not yield the 'stable sink', 'unstable sink', or focus results in Tabs. 1-3; an undocumented energy (e.g., a σ(Δ) term or real-part projection) is needed. The reported alignment rates are encouraging but cannot be attributed to the algorithm as described until this gap is resolved.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces \"topology guidance,\" a training-free method for steering the DDIM sampling process of a latent diffusion model over Functa/SIREN-coded 2D vector fields. Guidance energies are defined to enforce the presence of a critical point at a user-specified location, to control its type (sink, source, saddle) through eigenvalue-sign penalties, and to control its stability through an additional eigenvalue-based term. The method is evaluated on a 2D fluid-flow ensemble, reporting alignment rates above 85% in most configurations while keeping Fréchet Distance close to the unguided baseline, together with qualitative studies of topology-conditioned distributions and a \"fill-in-the-gap\" analysis.","tokens_in":18585,"tokens_out":6898,"duration_ms":69723,"significance":"If the method works as claimed, it is a useful contribution to visualization and generative-model control: it enables users to specify topological features of generated vector fields without retraining, and the ensemble-comparison application is compelling. The paper has clear strengths: large-scale quantitative evaluation (2,500 samples per configuration in Table 1), distribution-fidelity checks via Fréchet Distance, multi-critical-point experiments, and an interesting qualitative analysis of how the diffusion model fills in the gap when the prescribed topology is moved. However, the central derivation currently contains load-bearing inconsistencies in the treatment of foci and stability, so the reported results cannot be attributed to the algorithm as written.","major_comments":[{"comment":"The statement that the sign of Δ in Eq. (12) determines stability is incorrect. Δ = Tr(J)^2 − 4|J| distinguishes nodes (real eigenvalues, Δ > 0) from foci (complex conjugate eigenvalues, Δ < 0); stability is determined by the real parts of the eigenvalues, which in 2D is governed by Tr(J). As written, the stability energy in Eq. (13) cannot separate stable from unstable foci.","section":"Sec. 4.3, Eq. (12)"},{"comment":"Minimizing E_s = σ(β_s λ_1) gives the opposite of the stated labels: for \"stable\" with β_s = −1, σ(−λ_1) is minimized as λ_1 → +∞, driving the eigenvalue positive; for \"unstable\" with β_s = 1, minimization drives λ_1 → −∞. Moreover, in the combined energy Eq. (14), the stability term σ(−λ_1) for a stable point directly conflicts with the type term σ(λ_1) for an attracting point. The sign convention in the text is reversed with respect to the stated minimization objective.","section":"Sec. 4.3, Eq. (13)"},{"comment":"Foci cannot be encoded by Eq. (10), because their eigenvalues are complex conjugates and do not admit the real ordering used there; substituting complex λ into the sigmoid in Eq. (11) does not yield a real energy. Figure 4 lists σ(Δ) terms for focus and node configurations, but Eq. (14) omits any Δ term. The focus and node results in Tables 1 and 3 therefore are not produced by the energy functions stated in the paper. The authors must state the actual implemented energies (e.g., using real parts of eigenvalues or σ(Δ) terms) and confirm that the reported evaluation uses those corrected energies.","section":"Sec. 4.2, Eqs. (10)-(11), and Fig. 4"},{"comment":"The values of the guidance strength ω and the guidance-start timestep are not specified; only Fig. 3 shows an injection at t = 600 for one example. These are free parameters that materially affect the alignment/fidelity trade-off reported in Tables 1-3, so without them Algorithm 1 is not reproducible.","section":"Sec. 5.1 and Algorithm 1"}],"minor_comments":[{"comment":"The formula for Fréchet Distance appears to be missing an opening norm bar: it should read ∥μ1 − μ2∥^2, not \"μ1− μ2∥2\".","section":"Eq. (17)"},{"comment":"The sentence \"otherwise we have negative critical point\" is a typo; it should say \"unstable critical point\".","section":"Sec. 4.3"},{"comment":"The text in Fig. 4 is garbled in several places, with missing square-root symbols and mismatched labels between configurations and energy terms; the figure should be redrawn so that each critical-point configuration and its associated constraints are unambiguous.","section":"Fig. 4"},{"comment":"The FD metric is evaluated on latent embeddings, but the embedding network is not described; please specify how the Fréchet Distance is computed on the latent distributions.","section":"Sec. 5.2"},{"comment":"The claim that the predicted clean latent ẑ_t is sufficiently in-distribution for the SIREN is plausible but not quantitatively validated; a small experiment reporting reconstruction error of ẑ_t versus t, or the norm of the guidance gradient, would strengthen this key assumption.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"The empirical evaluation appears extensive and the application is timely, but the equations in Secs. 4.2-4.4 are internally inconsistent with respect to foci and stability, as detailed in the major comments. This is fixable in revision if the authors provide corrected energy functions and confirm that the experimental results were obtained with those functions. I would not recommend rejection, but the central claim of controlling critical-point type and stability is not supported by the manuscript as written."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe headline: this is a useful idea that is not yet properly described. The authors couple a Functa SIREN with a latent diffusion model and use Jacobian eigenvalues to guide sampling toward prescribed critical point locations and types without retraining. The empirical setup on the Jakob fluid-flow ensemble is a real step forward for the visualization community, and the alignment rates above 85% are encouraging if they hold up. The \"fill in the gap\" analysis in Fig. 10 is a nice qualitative result.\n\nWhat's new is the specific mechanism—guiding a diffusion model with SIREN-derived topological signals—and the systematic evaluation across location, type, and multiple critical points. That is worth someone's attention.\n\nBut there is a load-bearing problem in the stability and focus guidance as written. Sec. 4.3 says the discriminant Δ = Tr² − 4|J| determines stability: positive means stable, negative means unstable. That is not what Δ does; it separates nodes from foci. Stability is about the real parts of the eigenvalues. Worse, Eq. 13 sets β_s = −1 for 'stable' and minimizes σ(β_s λ1); that pushes λ1 positive, which is precisely an unstable source under any normal convention. So the energy functions stated in Sec. 4.3 would generate the opposite of the reported 'stable' configurations. And for focus configurations, the eigenvalues are complex, so Eq. 11's sigmoid on λ1, λ2 is not even well-defined. Fig. 4 shows σ(Δ) terms for foci, but Eq. 14—the combined energy—omits them. So the text does not actually describe the algorithm that produced the Table 3 results. This is not a typo-level issue; it is the heart of the guidance method for two of the five critical point types.\n\nThere are minor problems too: no code, no baseline beyond unguided sampling, the alignment metric leaves the distance threshold and 'correct classification' unspecified, and the entire evaluation is on one dataset. The paper also cites the relevant prior work (universal guidance, Functa, SIREN) adequately.\n\nNet: the idea is worth publishing if the authors fix the math and release code. But as it stands, a reader cannot reproduce the stability and focus results from the description, and the sign error suggests the actual implementation may differ from the write-up. I'd send it to review, but with a clear expectation of major revision. If the authors can reconcile the equations with the reported results, it becomes a solid contribution to topology-driven generation.\n\nRecommendation: accept for peer review, but the referee should demand a corrected Sec. 4.3 and a reproducible implementation.","headline":"Promising idea for topology-guided diffusion, but the stability/focus guidance as written is mathematically inconsistent with the reported results—fix before trusting the numbers.","tokens_in":19118,"tokens_out":4265,"would_cite":false,"duration_ms":40574,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A training-free guidance method steers a diffusion model over vector fields so that generated fields contain user-specified critical points—sinks, sources, and saddles at chosen locations—while staying in the modeled data distribution.","keywords":["topology guidance","vector field topology","critical points","diffusion guidance","coordinate-based neural networks","simulation ensembles","generative control","fluid flow analysis"],"falsifier":"A decisive test is to sweep the timestep at which guidance is first applied, holding one noise seed fixed, and record critical-point alignment: the paper's account predicts a plateau above 85% for a range of start times, and absence of that plateau, or alignment that only appears when the critical-point detector's distance tolerance is loosened, would refute the central claim. A complementary numerical check computes the distance between DDIM-predicted clean latents at the chosen start time and the training latent distribution; a large distance would violate the in-distribution assumption stated in Sec. 4.1.","tokens_in":18068,"feed_emoji":"🌀","tokens_out":9863,"duration_ms":92961,"temperature":0.7,"pith_summary":"The paper tries to establish that a diffusion model trained on latent codes of vector fields can be steered, without any retraining, to produce fields with user-specified critical points: the right location, the right type (sink, source, or saddle), and the right stability (node or focus). The proposed method, topology guidance, turns the differentiable structure of a coordinate-based neural field into energy functions, then adds gradients of those energies to the diffusion model's denoising step. On 2D fluid-flow ensembles, the paper reports alignment rates above 85% and often near 100% for fixed noise seeds, while Fréchet distances to the real ensemble remain close to unguided sampling. If correct, this means an analyst can query a generative model by topological description instead of drawing many samples and checking them afterward.","feed_headline":"Topology guidance places sinks, sources, and saddles where users ask","feed_subtitle":"No retraining: specify a critical point's location and type, and sampled flow fields obey it while staying in the data distribution.","key_machinery":"The central object is the field-conditioned coordinate network f_phi(p; z_hat_t), mapping a spatial location p and a latent code z into a 2D vector and differentiable in both arguments. It supplies three signals that define the guidance energies: the field norm ||f_phi(p; z_hat_t)|| for critical-point existence, the eigenvalues of the Jacobian J_phi(p; z_hat_t) for sink/source/saddle type, and the discriminant $\\Delta$ = Tr(J)^2 - 4|J| for node versus focus stability. Sigmoid penalties push eigenvalue signs and $\\Delta$ toward the user-specified classification, and all penalties are added together for multiple critical points. The diffusion model provides the latent prior, while DDIM's predicted clean latent z_hat_t is what keeps the coordinate network's Jacobian meaningful at intermediate timesteps.","core_discovery":"Topology guidance claims that a diffusion model over latent field codes can be guided to satisfy a prescribed critical-point configuration by evaluating a coordinate-based neural network with sinusoidal activations on the DDIM-predicted clean latent at the specified location. The network supplies the field value, its Jacobian, and the trace-determinant discriminant, which are converted into energy terms for critical-point existence, type, and stability. Gradients of the summed energy with respect to the noisy latent are added to the model's predicted noise, steering each denoising step toward the desired topology. The paper reports alignment above 85% across single and combined specifications, higher alignment when the noise vector is fixed, and Fréchet distances comparable to unguided generation, concluding that the generated fields satisfy the specified topology while remaining in the data distribution.","pith_inferences":["Inference: the same energy-on-derivatives recipe should transfer to 3D vector fields and to scalar fields, since critical points and Morse-Smale cells are defined through the same Jacobian or Hessian machinery; the paper does not test this.","Inference: the alignment drop for two nearby critical points in Table 4 suggests the fluid-flow distribution has an effective minimum separation between same-sign critical points; comparing that spacing with vortex-core statistics in the simulation data would be a direct check.","Inference: because guidance is applied only after a chosen start time and uses the predicted clean latent, the method implicitly trades diversity for control; measuring Fréchet distance across guidance start times would quantify that tradeoff.","Inference: any smooth function of pointwise field derivatives could replace the topology energy and yield a new training-free control, so the approach generalizes to features like vorticity or other vortex criteria that the paper only mentions as future directions."],"forward_implications":["A pretrained diffusion model of fields can be queried by topological specification without retraining or fine-tuning, so users can synthesize many fields sharing a prescribed critical-point configuration.","Topology guidance enables direct comparison of two simulation ensembles: fields from different distributions can be generated with identical prescribed topology, exposing differences in local flow behavior.","The guidance is largely local, so moving the prescribed critical point changes the field mainly in a neighborhood of that point, supporting what-if edits of individual samples.","Combining norm, eigenvalue, and stability penalties lets users specify fine distinctions such as stable versus unstable sinks and sources, or node versus focus, not just coarse point type.","Multiple critical points can be specified at once, and the achievable spacing between them is bounded by the modeled distribution, which the paper uses to estimate typical critical-point distances in the fluid ensemble."],"supporting_citations":[{"why":"supplies the latent-conditioned coordinate network whose differentiability yields the topology signals.","marker":"[10]"},{"why":"defines the sinusoidal activation architecture that makes spatial derivatives stable enough to compute Jacobians.","marker":"[40]"},{"why":"provides the diffusion-guidance update that adds energy gradients to the predicted noise.","marker":"[8]"},{"why":"defines the denoising diffusion probabilistic model over latent field codes.","marker":"[20]"},{"why":"gives the DDIM sampler and the predicted clean latent used to evaluate the coordinate network.","marker":"[41]"},{"why":"shows how to use an off-the-shelf model for guidance via predicted clean latents, the step that keeps the network in-distribution.","marker":"[1]"},{"why":"supplies the 2D fluid-flow simulation ensemble used for training and evaluation.","marker":"[24]"},{"why":"establishes the critical-point classification that the guidance energies encode.","marker":"[19]"}],"fun_headline_variants":["User-specified critical points guide diffusion sampling of flow fields","Diffusion guidance places vector field critical points on demand","Control generative models with topological specifications","Topology guidance: steer diffusion denoising to prescribed critical points","Critical point specs become diffusion sampling constraints"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the clean latent predicted by DDIM at the timesteps where guidance is active is close enough to the latent distribution the coordinate network was trained on for its Jacobian-based energy gradients to be meaningful; the paper states that raw noisy latents are out-of-distribution at large t and provides no quantitative check that the predicted latent fixes this.","fun_headline_variants_meta":{"raw":{"variants":["User-specified critical points guide diffusion sampling of flow fields","Diffusion guidance places vector field critical points on demand","Control generative models with topological specifications","Topology guidance: steer diffusion denoising to prescribed critical points","Critical point specs become diffusion sampling constraints"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000506,"raw_usage":{"total_tokens":2500,"prompt_tokens":1012,"completion_tokens":1488,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":1415}},"tokens_in":628,"tokens_out":1488,"duration_ms":10951,"temperature":1.0,"reasoning_tokens":1415,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:32:31.186336+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test is to sweep the timestep at which guidance is first applied, holding one noise seed fixed, and record critical-point alignment: the paper's account predicts a plateau above 85% for a range of start times, and absence of that plateau, or alignment that only appears when the critical-point detector's distance tolerance is loosened, would refute the central claim. A complementary numerical check computes the distance between DDIM-predicted clean latents at the chosen start time and the training latent distribution; a large distance would violate the in-distribution assumption stated in Sec. 4.1.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"establishes the critical-point classification that the guidance energies encode."},{"cited_title":"Sitzmann, J","cited_arxiv_id":null,"evidence_quote":"defines the sinusoidal activation architecture that makes spatial derivatives stable enough to compute Jacobians."},{"cited_title":"Dhariwal and A","cited_arxiv_id":null,"evidence_quote":"provides the diffusion-guidance update that adds energy gradients to the predicted noise."},{"cited_title":"Bansal, H.-M","cited_arxiv_id":null,"evidence_quote":"shows how to use an off-the-shelf model for guidance via predicted clean latents, the step that keeps the network in-distribution."},{"cited_title":"Jakob, M","cited_arxiv_id":null,"evidence_quote":"supplies the 2D fluid-flow simulation ensemble used for training and evaluation."}],"review_version":1}