Extracting Local Manifold Geometry from Pretrained Diffusion Models in One Inverse Step

Source

• Raw Markdown: paper_diffusiongeometryprobe-2026.md
• PDF: paper_diffusiongeometryprobe-2026.pdf
• OpenReview: https://openreview.net/forum?id=E7op24P6Uh
• Official GALOP accepted-paper list: https://hyperboliclearning.github.io/events/kdd2026workshop
• Second OpenReview record: https://openreview.net/forum?id=vuf1MpXfXX
• Official SPIGM @ ICML 2026 accepted-paper list: https://spigmworkshop2026.github.io/papers
• Local artifact and credibility audit: papers/diffusiongeometryprobe-2026/artifact_status.md

Status And Credibility

This is an eight-page solo-author paper by Gordei Verbii, listed by the official GALOP site as an accepted poster at the Geometric Space, Architecture and Learning Objectives for Large Pre-Trained Models Workshop at KDD 2026 on 2026-08-09. The official SPIGM @ ICML 2026 accepted-paper page also lists the same title under OpenReview record vuf1MpXfXX. These are workshop acceptances, not KDD or ICML main-track evidence. No separate archival-proceedings record was verified.

No exact-title arXiv record, official code repository, released implementation, dataset, checkpoint, or supplementary archive was found as of 2026-08-09. Automated OpenReview access was blocked by browser verification, so the user-supplied rendered PDF is the authoritative source for this ingest.

The paper is highly relevant to the wiki’s manifold-reconstruction agenda because it tries to turn a pretrained diffusion score field into a local geometry probe. Its credibility must nevertheless be calibrated carefully: the experiments are small, the two image results use one generated image each, the synthetic model is deliberately undertrained, and several advertised mathematical certificates or topology interpretations do not follow from the displayed derivations as written.

Core Claim

DiffusionGeometryProbe treats one implicit DDIM inversion step as a fixed-point problem. Around one input and selected noise levels, it queries the pretrained noise predictor and its Jacobian using Jacobian-vector products, then reports a bundle of diagnostics:

• a local fixed-point contraction bound and observed Picard iteration count;
• a proxy presented as a Smale-alpha Newton certificate;
• four local intrinsic-dimension estimates;
• score-Jacobian extremal singular values and condition number;
• a proposed Cheeger-style isoperimetric proxy;
• score-energy gradient-flow basin counts and a fitted Łojasiewicz exponent;
• a score-norm plateau;
• optional persistent homology over a batch of generated or Tweedie-projected samples.

The attractive research idea is to inspect what a pretrained diffusion model has learned without running its complete reverse chain or training a separate geometry model.

Plain-Language Picture

Imagine that realistic images occupy a thin curved surface inside an enormous pixel space. A well-trained denoising model should learn which perturbations move along that surface and which move away from it. The paper asks whether one can inspect the model’s local denoising field and infer:

• how many local directions look data-like;
• whether inversion should converge quickly;
• whether tangent and normal directions form a spectral gap;
• whether local basins or sampled point clouds preserve any topology.

That framing is useful. The paper’s individual measurements, however, have very different evidential status and should not be treated as fourteen equally certified geometric invariants.

What “One Inverse Step” Means

The title does not mean one network evaluation or one-pass runtime. The method focuses on one implicit DDIM transition rather than executing the full reverse diffusion chain, but Algorithm 1 still performs:

• power and shifted-power iterations;
• multiple Picard iterations;
• repeated noisy forward evaluations for four dimension estimators;
• sweeps over four to six noise levels;
• several score-energy flows;
• optional batch persistent homology.

Reported runtimes are 218.7 seconds for CIFAR-10 and 845.6 seconds for CelebA-HQ-256 on a Colab T4. “One inverse step” is therefore a structural scope claim, not a one-forward-pass efficiency claim.

Method Components And Audit

Component	Paper’s interpretation	Calibrated reading
DDIM fixed-point map	Local contraction rate is $\rho_g=	B_t
Banach iteration budget	Eq. (4) predicts Picard iterations and the experiment validates $n\sim (1-{\rho }_g{)}^{-1}$.	The displayed formula is standard for a contraction, but the table’s reported predictions are not fully consistent with its displayed rounded inputs. The four-point fit also changes initial residual with guidance, so it does not isolate the contraction-rate law.
Smale-alpha proxy	Eq. (5) certifies Newton-type convergence from residual and ${\rho }_g$.	Smale’s $\gamma$ depends on all higher derivatives. A stated bound on $D^{2}F$ alone does not establish the claimed $\alpha$ upper bound for a general analytic neural network. Treat this as a heuristic regime score, not a certificate.
Stanczuk normal-bundle estimate	Score covariance rank yields codimension.	The cited method recommends at least as many perturbation samples as the normal dimension. An empirical covariance from $K\ll D$ samples has rank at most $K-1$, forcing the reported estimate close to ambient dimension at image scale.
FLIPD	Fokker–Planck estimator tends to LID at low noise.	The theoretical low-noise conditions matter; the synthetic undertrained model returns approximately ambient dimension, and CelebA reports $1.009D$ at $t=4$, outside a valid dimension range.
Yeats DSM estimate	Denoising loss lower-bounds LID and becomes exact for an optimal denoiser.	The cited theorem assumes sufficiently small noise, locally constant density, and negligible curvature. Values at high-noise timesteps cannot be interpreted as the same local asymptotic without an additional argument.
Local PCA	PCA of perturbed samples is an independent LID reality check.	As written, PCA is applied to isotropic Gaussian perturbations of one point, not denoised neighbours. Its rank is therefore driven mainly by sample count; the reported totals near 20–23 are consistent with this ceiling rather than image-manifold dimension.
Score-Jacobian condition number	Its noise sweep reveals three diffusion phases.	This is a plausible transfer from the ICLR 2025 spectral-gap work, but the paper provides mainly qualitative curves and no population-level uncertainty.
Cheeger proxy	$\hat{h}=\sqrt{2}/{\kappa }_t$ lower-bounds a local isoperimetric constant.	The identification of $J_{ϵ}^{\top }{J}_{ϵ}$ with a normalized manifold Laplacian is assumed rather than derived. Moreover, the displayed Cheeger inequality gives an upper relation on $h$ from ${\lambda }_1$, not the advertised lower bound.
Score-energy basins	Number of attraction basins lower-bounds ${\sum }_k{\beta }_k(M)$.	This is not the standard weak Morse inequality. Morse theory lower-bounds critical points of each index; attraction basins count minima. In general, $\#\{\mathrm{minima}\}\ge {\beta }_0$, not ${\sum }_k{\beta }_k$.
Łojasiewicz exponent	A fitted log-log slope measures basin degeneracy.	The inequality is converted into an approximate equality, fitted on short trajectories, and clipped to $[1/2,0.99]$. The paper reports raw sub-$1/2$ fits, so this output is exploratory rather than calibrated.
Score-norm plateau	${\Pi }_t\to D-d$ estimates codimension.	The cited singular limit is a small-noise result. The paper then uses $t=300$, where noise is larger, to infer $d\approx 417$ for CelebA. At high noise, predicting the injected Gaussian already drives $\mathbb{E}\Vert {ϵ}_{\theta }{\Vert }^2$ toward $D$, so that inference is not justified by the displayed theorem.
Persistent homology	Tweedie-projected batches reveal topology with confidence bands.	The image batches contain only 12 or 32 points in very high dimension; the paper itself says they support essentially only fragile $H_0$ conclusions and no global-topology recovery.

Two Explicit Mathematical Inconsistencies

Smale Threshold

Theorem 3 writes

$${\alpha }_0=3-2\sqrt{2}\approx \mathrm{0.15768.}$$

The formula and decimal do not match:

• $3-2\sqrt{2}\approx 0.171573$;
• $(13-3\sqrt{17})/4\approx 0.157671$.

Using the more conservative decimal does not by itself invalidate a sufficient test, but the paper combines two different published thresholds and then applies a proxy that does not compute or validly bound the full Smale $\gamma$ quantity.

Morse Basin Bound

Equation (11) claims

$$\#\{\text{basins of }{E}_t\}\ge \sum _k{\beta }_k(M).$$

A sphere is the simplest counterexample to that interpretation: a Morse function can have one minimum and one maximum, while ${\beta }_0+{\beta }_2=2$. The number of attraction basins/minima can be one. The weak Morse inequalities apply to critical points by index, not to minima alone.

Experimental Evidence

Synthetic Point-Cloud DDPM

The paper samples five known point-cloud families in ${\mathbb{R}}^3$: sphere, torus, Swiss roll, trefoil knot, and two disjoint spheres. Each generated object contains 128 points, so the flattened ambient dimension is 384. A Luo–Hu point-cloud DDPM is trained for 50 epochs and deliberately left undertrained, with reported Chamfer reconstruction error from 0.45 to 0.85.

The useful result is a fixed-point stress test over four guidance scales. Reported local norm bounds rise from 0.0217 to 0.3258 and observed Picard iterations rise from 3 to 9. This supports the qualitative statement that stronger local feedback slows fixed-point convergence.

Using Equation (4), the table’s rounded ${\rho }_g$, initial residuals, and $\epsilon =10^{-4}$ gives ceiling iteration counts $\{2,4,5,9\}$, not the reported predicted/observed sequence $\{3,4,6,9\}$. Hidden unrounded values could move a near-integer boundary, but they do not explain the first case’s large difference. The claimed $R^2>0.99$ fit also uses only four guidance settings while the initial residual changes, so it is not an independent validation of $n\sim (1-{\rho }_g{)}^{-1}$.

The geometry recovery result is weak:

• FLIPD returns approximately 3 dimensions per point for every class, including the one-dimensional trefoil;
• most basin counts collapse to one;
• only the sphere produces six counted basins against a reported Betti sum of two;
• the paper interprets these failures as diagnostics of the undertrained model rather than demonstrating recovery of the known manifolds.

CIFAR-10 DDPM

The paper probes one generated image from google/ddpm-cifar10-32 at $t=4$:

• ambient dimension: 3,072;
• local contraction bound: 0.121;
• observed Picard iterations: 4;
• FLIPD: $0.833D$;
• Stanczuk: $1.000D$;
• Yeats: $0.959D$ at $t=4$, falling to $0.034D$ at $t=300$;
• local PCA: $0.0075D$, approximately 23 dimensions;
• runtime: 218.7 seconds on a Colab T4.

The large estimator disagreement is more credible as a warning about regime and protocol than as a recovered CIFAR manifold dimension.

CelebA-HQ-256 DDPM

The paper probes one generated face from google/ddpm-celebahq-256:

• ambient dimension: 196,608;
• local contraction bound: 0.959;
• observed Picard iterations: 19;
• Smale proxy: 238.8, labelled not certified;
• FLIPD: $1.009D$ at $t=4$;
• Stanczuk: $1.000D$ at $t=4$;
• Yeats: $0.063D$ at $t=4$ and $0.002D$ at $t=300$;
• local PCA: approximately $0.0001D$;
• runtime: 845.6 seconds on a Colab T4.

The paper says Yeats and local PCA agree in order of magnitude. The displayed ratios do not support that statement: at $t=4$ they imply roughly 12,386 versus roughly 20 dimensions. At $t=300$, Yeats implies roughly 393, still far from the local-PCA value. The reported $d\approx 417$ from the score plateau is numerically close to the high-noise Yeats value, but both depend on a noise regime different from the small-noise limits cited for local manifold interpretation.

What Is Genuinely Useful

Despite the problems above, the paper contributes several useful research prompts:

1. Treat inversion conditioning as a measurable model property. A Jacobian-norm bound can warn when a chosen implicit DDIM step is poorly conditioned, even if it is not the exact convergence rate.
2. Report estimator disagreement rather than one dimension number. FLIPD, normal-bundle, DSM-loss, and neighbourhood estimates can fail for different reasons.
3. Inspect geometry across noise scales. Spectral phase curves are potentially more informative than one arbitrarily selected timestep.
4. Keep local and global claims separate. A local score-field probe at one image cannot recover dataset topology.
5. Use controlled known-geometry benchmarks. The paper’s own undertrained synthetic model shows why ground truth and negative controls are essential.

Relation To Aionoscope Manifold Reconstruction

DiffusionGeometryProbe and the Aionoscope Manifold Reconstruction Benchmark inspect different objects:

Question	DiffusionGeometryProbe	Aionoscope manifold benchmark
Model requirement	DDIM-compatible diffusion noise predictor	Any frozen time-series encoder with an adapter
Object inspected	Local geometry of a generative score field in observation space	Geometry of known latent process factors after encoding
Ground truth	Known only for the synthetic point-cloud test	Known by construction from controlled latent variables and sweeps
Main measurements	Score Jacobian, diffusion-noise LID estimators, inversion conditioning	Recoverability, topology, local tangent, metric and density distortion, factor coupling
Main risk	Asymptotic theorem used outside its regime; small-sample rank ceilings	Probe or metric can reward a beautiful but downstream-useless representation

The strongest transfer is to use Aionoscope as a falsification harness for score-field probes. A diffusion-based time-series encoder or generator could be tested on latent factors with known dimension, circle/torus topology, rank-loss boundaries, and density. This would reveal whether each score-derived output tracks actual process geometry or only noise level, sample count, and model conditioning.

The paper does not itself evaluate numeric time series, multivariate process state, event streams, actions, control inputs, interventions, or action-conditioned world models.

Relation To Beckmann Transport Models

Beckmann Transport Models ask whether an autonomous field or direct map transports probability mass onto a target support with correct regional mass. DiffusionGeometryProbe instead inspects local derivatives of a pretrained diffusion score field. Neither method reconstructs an inverse chart of the data manifold.

For the Aionoscope agenda, the two are complementary optional probes:

• score-field diagnostics ask whether local tangent/normal and dimension signatures are visible;
• BTM-style diagnostics ask whether off-manifold perturbations return to the correct support and region.

Both must be tested against Aionoscope’s known latent geometry before their outputs are interpreted as representation fidelity.

Foundation TSFM Relevance

Agenda slot	Verdict	Evidence	Missing pieces
Representation quality	adjacent	Provides a score-field recipe for local tangent/normal and dimension diagnostics.	No time-series encoder, no latent-state utility test, and several estimators are not calibrated in the reported regime.
Benchmark level	warning	Shows why estimator disagreement, noise-scale sweeps, ground truth, and negative controls are necessary.	Needs repeated seeds, sufficient perturbation rank, valid asymptotic regimes, and known-geometry pass/fail criteria.
Generation and editing	adjacent	Local inversion conditioning may warn when a diffusion transition is difficult.	No numeric time-series generation, editing, exact-value fidelity, or end-to-end quality comparison.
Latent-state prediction	insufficient evidence	None; the paper probes an image-generation score field.	Must test whether geometry scores predict accessible process state or future-state quality.
Control and counterfactuals	insufficient evidence	No action or intervention channel.	Add typed control inputs or interventions and test action-sensitive manifold families.

Limitations And Gotchas

• “One inverse step” still requires many network and Jacobian-vector evaluations.
• The local contraction number is an operator-norm bound, not necessarily the exact rate.
• The Smale-alpha quantity is a heuristic proxy and should not be labelled a mathematical certificate.
• Two of the four LID outputs are structurally dominated by perturbation sample count in high ambient dimension under the described implementation.
• Several cited limits require low noise, while the paper draws its strongest image-dimension conclusion at high noise.
• The Cheeger and Morse-theory interpretations are not justified as written.
• The synthetic benchmark demonstrates failure diagnosis more than successful manifold recovery.
• CIFAR and CelebA each use one generated image, with no uncertainty over inputs, seeds, checkpoints, or hyperparameters.
• No public implementation was verified.
• Results concern image observation space, not time-series representation space or latent state.

Links Into The Wiki

• Aionoscope
• Aionoscope Manifold Reconstruction Benchmark
• Beckmann Transport Models
• Time-Series Benchmark Hygiene
• Latent-Space Predictive Learning
• Foundation Time-Series Model Research Agenda

Open Questions

• Can each proposed invariant recover known local dimension, tangent directions, topology, and density on a fully trained synthetic diffusion model across seeds?
• What perturbation count is required before the normal-bundle estimator is identifiable when $D-d$ is large?
• Does local PCA become meaningful only after Tweedie projection or another denoising map?
• Can a valid bound on all derivatives needed by Smale’s $\gamma$ be computed with practical Jacobian-vector and higher-order products?
• Which score-spectrum diagnostics predict downstream latent-state accessibility rather than image realism?
• Can Aionoscope distinguish score-field geometry that tracks a true periodic or regime variable from geometry driven only by noise schedule?
• Do the spectral-phase curves survive on multivariate time-series diffusion models under irregular sampling and regime changes?