pith. sign in

arxiv: 2606.13191 · v1 · pith:FO33PMNAnew · submitted 2026-06-11 · 💻 cs.LG

The Geometry of Phase Transitions in Generative Dynamics via Projection Caustics

Ryosuke Sakamoto , Kotaro Sakamoto This is my paper

Pith reviewed 2026-06-27 07:10 UTC · model grok-4.3

classification 💻 cs.LG
keywords diffusion modelsphase transitionsprojection causticsgenerative dynamicscritical boundary detectorscore instabilitymode commitment
0
0 comments X

The pith

Sharp transitions in generative sampling occur at projection caustics where nearest-point projections cease to be unique.

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

This paper provides a geometric account of abrupt qualitative changes in continuous-state generative models such as diffusion and flow-matching. It frames denoising as gradient descent on a free energy landscape and locates the source of phase-transition-like behavior at projection caustics. At these points the nearest-point projection onto the data support is no longer unique. The work introduces the Critical Boundary Detector to identify regions of score-direction instability. Tests across toy examples and standard diffusion models confirm that this detector localizes mode commitment and identifies intervention-sensitive time windows.

Core claim

Sharp transitions arise near projection caustics, where the nearest-point projection onto the data support ceases to be unique. The Critical Boundary Detector acts as a practical diagnostic for score-direction instability and enables targeted control in sensitive regions of the dynamics.

What carries the argument

Projection caustics: the set of points where the nearest-point projection onto the data support ceases to be unique, serving as the geometric trigger for instability in the denoising dynamics.

If this is right

  • Mode commitment happens at these caustic boundaries.
  • CBD localises mode commitment and predicts intervention-sensitive windows.
  • Targeted control becomes possible in geometrically sensitive regions.
  • The geometric view connects data support geometry directly to generation dynamics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The caustic perspective could guide the design of sampling schedules that steer clear of unstable boundaries.
  • Similar projection-based instabilities may appear in other continuous-time generative frameworks.
  • Detecting these boundaries might improve robustness when fine-tuning diffusion models on new data.

Load-bearing premise

Denoising can be viewed as gradient descent on a free energy landscape whose geometry is governed by the data support in a way that makes projection non-uniqueness the direct cause of observed phase-transition behavior.

What would settle it

A calculation or experiment that finds sharp transitions occurring away from locations where nearest-point projections lose uniqueness would falsify the geometric account.

Figures

Figures reproduced from arXiv: 2606.13191 by Kotaro Sakamoto, Ryosuke Sakamoto.

Figure 2
Figure 2. Figure 2: SD 3.5 CBD [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Transverse-cross reproduction for K = {(x1, x2) ∈ R 2 : x1x2 = 0}. As the noise level decreases, the free energy develops a ridge-like branch-competition geometry near the projection caustic. The raw CBD fields concentrate around the corresponding switching structure shown in the proxy. centroid ( [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Projection-regular versus caustic regimes. The set C(K) is the locus where projection unique￾ness fails. Geometrically, it plays the role of a medial￾axis or cut-locus type singular set [10, 16]. Dynamically, it is the region where multiple nearest-point explana￾tions compete, so that small perturbations of x may cause abrupt changes in the dominant descent direction of the free energy (See [PITH_FULL_IMA… view at source ↗
Figure 5
Figure 5. Figure 5: Transverse-cross reproduction. Using this high-CBD region as an intervention trigger recovers [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: CIFAR-10 DDPM: CBD predicts intervention-sensitive windows. (a) The [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: From shared diagnosis to phase-aware control. (a) After per-run min–max normalisation [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: wmn denotes the case where intervention is performed over the time window (m,n). L2 and L1 indicate the L2 and L1 distances, respectively, between the baseline generation result and the intervention result. If we intervene within each time window in 0-11, cartoon-like feature appears. The intervention result switches from cartoon-like features to cinematic features in the time window 12–15. This switching … view at source ↗
Figure 9
Figure 9. Figure 9: CBD plots along the baseline trajectory at late time step [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Evolution of the empirical branch-decision field during reverse diffusion on the cusp dataset. [PITH_FULL_IMAGE:figures/full_fig_p030_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Estimated switching-band proxy for the transverse cross at [PITH_FULL_IMAGE:figures/full_fig_p033_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Branch-selection control on the transverse cross. From left to right: baseline sampling, always [PITH_FULL_IMAGE:figures/full_fig_p034_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: baseline 35 [PITH_FULL_IMAGE:figures/full_fig_p035_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Future-count profile along a baseline reverse trajectory. High-value contiguous segments [PITH_FULL_IMAGE:figures/full_fig_p036_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: We intervene within the time window 110-200 [PITH_FULL_IMAGE:figures/full_fig_p036_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: We intervene within the time window 250-350. [PITH_FULL_IMAGE:figures/full_fig_p036_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: We intervene within the time window 300-400. [PITH_FULL_IMAGE:figures/full_fig_p037_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: we intervene within the time window 50-75 [PITH_FULL_IMAGE:figures/full_fig_p037_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: we intervene within the time window 85-110 [PITH_FULL_IMAGE:figures/full_fig_p037_21.png] view at source ↗
Figure 24
Figure 24. Figure 24: Normalised TAD-ratio diagnostic along a baseline reverse trajectory on CIFAR10. The ratio [PITH_FULL_IMAGE:figures/full_fig_p040_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: TAD for another trajectory (seed = 42) 40 [PITH_FULL_IMAGE:figures/full_fig_p040_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Normalised mean ℓ2 distance and mean LPIPS distance over trajectory indices. Relation between the TAD-ratio profile and downstream effect size. We then compared the smoothed TAD-ratio curve ( [PITH_FULL_IMAGE:figures/full_fig_p041_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Qualitative intervention study on SD3.5 for the prompt “a cinematic photo of a shiba inu [PITH_FULL_IMAGE:figures/full_fig_p047_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: wmn denotes the case where intervention is performed over the time window (m,n). L2 and L1 indicate the L2 and L1 distances, respectively, between the baseline generation result and the intervention result. If we intervene within each time window in 0-11, cartoon-like feature appears. The intervention result switches from cartoon-like features to cinematic features in the time window 12–15. This switching… view at source ↗
Figure 29
Figure 29. Figure 29: CBD plots along the baseline trajectory at late time step [PITH_FULL_IMAGE:figures/full_fig_p049_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: early to late time CBD plots 50 [PITH_FULL_IMAGE:figures/full_fig_p050_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: CIFAR-10 DDPM: sliding-window TAD profile (mean [PITH_FULL_IMAGE:figures/full_fig_p051_31.png] view at source ↗
Figure 32
Figure 32. Figure 32: Stable Diffusion 3.5 Medium: 1−TAD profile for probe displacements ∆ ∈ {1, 3, 5} (mean ± 1σ, five seeds). 8 10 12 14 16 18 20 Trajectory index 0.0 0.2 0.4 0.6 0.8 1.0 Normalized L2 Normalized L2 vs trajectory index SD-3.5 seed=42 seed=0 seed=1 seed=2 seed=10 mean ±1 std 0.04 0.02 0.00 0.02 0.04 Trajectory index 0.04 0.02 0.00 0.02 0.04 Normalized LPIPS Normalized LPIPS vs trajectory index SD-3.5 seed=42 s… view at source ↗
Figure 33
Figure 33. Figure 33: Stable Diffusion 3.5 Medium: sliding-window TAD profile (mean [PITH_FULL_IMAGE:figures/full_fig_p051_33.png] view at source ↗
Figure 34
Figure 34. Figure 34: CelebaHQ face DDPM: 1 − TAD profile (mean ± 1σ, five seeds). The band structure recapitulates the three-region pattern observed for DiT-XL, EDM2, and SD-3.5 in Figure 7a. 0.0 0.2 0.4 0.6 0.8 1.0 Normalized trajectory index (0=noisy, 1=clean) 0 2 4 6 8 10 Run index (seed × sample) DiT-XL/2-256 Free energy landscape heatmap (N=10 runs, F proxy from TAD_actual) 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 Free energy pro… view at source ↗
Figure 35
Figure 35. Figure 35: DiT-XL/2 free energy landscape: heatmap (left) and three-dimensional surface (right) along [PITH_FULL_IMAGE:figures/full_fig_p052_35.png] view at source ↗
Figure 36
Figure 36. Figure 36: EDM2-XS free energy landscape (CIFAR-10): heatmap (left) and three-dimensional surface [PITH_FULL_IMAGE:figures/full_fig_p052_36.png] view at source ↗
Figure 37
Figure 37. Figure 37: Stable Diffusion 3.5 Medium free energy landscape: two-dimensional projection (left) and [PITH_FULL_IMAGE:figures/full_fig_p053_37.png] view at source ↗
Figure 38
Figure 38. Figure 38: Additional trajectory and free energy diagnostics. The panels compare TAD profiles, sliding [PITH_FULL_IMAGE:figures/full_fig_p053_38.png] view at source ↗
Figure 39
Figure 39. Figure 39: CBD-ratio grouping and intervention sensitivity. High-ratio windows induce substantially larger [PITH_FULL_IMAGE:figures/full_fig_p054_39.png] view at source ↗
Figure 40
Figure 40. Figure 40: Four-model 1 − TAD overlay including MDLM caveat. DiT-XL/2 and EDM2-XS use eight seeds each; SD-3.5 Medium uses one thousand samples; MDLM is a single trajectory and is shown for completeness only. The continuous-state models (DiT-XL, EDM2, SD-3.5) all exhibit the three-region structure predicted by the projection-caustic theory; MDLM is excluded from the universal-signal claim in §3.5 for the reasons sta… view at source ↗
read the original abstract

Continuous-state generative samplers, including diffusion and flow-matching models, evolve through continuous reverse-time dynamics, yet their samples often undergo abrupt qualitative changes: trajectories commit to modes, semantic alternatives collapse, and small perturbations in narrow time windows can produce large downstream effects. This paper develops a geometric account of such phase-transition-like behaviour. We view denoising as gradient descent on a free energy landscape and show that sharp transitions arise near projection caustics, where the nearest-point projection onto the data support ceases to be unique. Motivated by this perspective, we introduce the Critical Boundary Detector (CBD), as practical diagnostics for score-direction instability. Across toy models, standard diffusion models, and latent text-to-image diffusion models, CBD localises mode commitment, predicts intervention-sensitive windows, and supports targeted control in geometrically sensitive regions. Our results connect geometry of data and dynamics of diffusion generation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper claims that abrupt qualitative changes (mode commitment, semantic collapse) in continuous reverse-time dynamics of diffusion and flow-matching models arise near projection caustics, where nearest-point projection onto the clean data support ceases to be unique. It frames denoising as gradient descent on a free-energy landscape whose geometry is dictated by the support, introduces the Critical Boundary Detector (CBD) as a diagnostic for score-direction instability, and reports that CBD localizes mode commitment, predicts intervention windows, and enables targeted control across toy, standard, and latent diffusion models.

Significance. If the geometric link between projection non-uniqueness and singularities in the reverse-time vector field can be made rigorous, the work would supply a concrete, falsifiable account of phase-transition behavior in generative samplers and a practical tool (CBD) for localizing sensitive regions. The empirical localization results across model scales would then constitute a reproducible diagnostic with potential downstream uses in controllable generation.

major comments (3)
  1. [Abstract / §1] Abstract and §1: the central claim that 'sharp transitions arise near projection caustics' because 'the nearest-point projection onto the data support ceases to be unique' is asserted rather than derived. The reverse-time score ∇log p_t remains C^∞ for all t>0 under standard Gaussian convolution; no explicit map is supplied from the geometry of supp(p_0) to singularities (or even rapid changes) of the probability-flow ODE or Fokker-Planck dynamics that would survive the convolution.
  2. [Abstract / Introduction] The free-energy interpretation is presented as motivation rather than a derived equivalence. No section shows that the denoising vector field is exactly the gradient of a free-energy functional whose critical points or Hessian are governed by nearest-point projection geometry; without this step the link between projection caustics and observed phase transitions remains circular.
  3. [Experiments (toy, standard, latent models)] Empirical claims for CBD (localization of mode commitment, prediction of intervention-sensitive windows) are reported without the quantitative controls required to rule out post-hoc fitting: no ablation of the CBD definition itself, no comparison against random or gradient-norm baselines, and no statement of how many trajectories or seeds were excluded.
minor comments (2)
  1. [Notation / §2] Notation for the projection operator and the precise definition of a 'caustic' in the context of the noisy marginal should be introduced with an equation before being used in the empirical sections.
  2. [Method] The manuscript should include a short appendix or paragraph clarifying whether CBD is computed from the learned score network or from an oracle; the distinction affects reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful reading and constructive critique. We address each major comment below, indicating where revisions will be made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract / §1] Abstract and §1: the central claim that 'sharp transitions arise near projection caustics' because 'the nearest-point projection onto the data support ceases to be unique' is asserted rather than derived. The reverse-time score ∇log p_t remains C^∞ for all t>0 under standard Gaussian convolution; no explicit map is supplied from the geometry of supp(p_0) to singularities (or even rapid changes) of the probability-flow ODE or Fokker-Planck dynamics that would survive the convolution.

    Authors: We agree that the current text presents the geometric link primarily through motivation and limiting arguments rather than a full derivation from the smoothed Fokker-Planck dynamics. While the score remains smooth, the direction of the probability-flow vector field can exhibit rapid variation near caustics because the underlying distance function develops singularities that influence the gradient even after convolution. In the revised manuscript we will add a short subsection in §2 that sketches this connection using the geometry of the squared-distance function and its Hessian, making the claim less assertive and more explicit. revision: yes

  2. Referee: [Abstract / Introduction] The free-energy interpretation is presented as motivation rather than a derived equivalence. No section shows that the denoising vector field is exactly the gradient of a free-energy functional whose critical points or Hessian are governed by nearest-point projection geometry; without this step the link between projection caustics and observed phase transitions remains circular.

    Authors: The free-energy view is introduced heuristically because the score equals the gradient of log p_t; we do not claim or derive an exact variational equivalence whose Hessian is controlled by projection geometry. We will revise the introduction and §2 to label this perspective explicitly as motivational, remove any implication of derived equivalence, and note the gap between the heuristic and a rigorous free-energy formulation. revision: yes

  3. Referee: [Experiments (toy, standard, latent models)] Empirical claims for CBD (localization of mode commitment, prediction of intervention-sensitive windows) are reported without the quantitative controls required to rule out post-hoc fitting: no ablation of the CBD definition itself, no comparison against random or gradient-norm baselines, and no statement of how many trajectories or seeds were excluded.

    Authors: We will expand the experimental section to include (i) an ablation study varying the CBD threshold and kernel parameters, (ii) direct comparisons against random-location and gradient-norm baselines with the same number of interventions, and (iii) explicit reporting that all quantitative results use 100 trajectories per setting across 5 independent random seeds, with no trajectories excluded. These additions will be placed in §4 and the appendix. revision: yes

Circularity Check

0 steps flagged

No circularity: geometric claim motivated by free-energy perspective, not derived tautologically

full rationale

The provided abstract and reader summary present the free-energy landscape view as a motivating perspective from which the projection-caustic account follows, with CBD introduced as an empirical diagnostic validated on toy and diffusion models. No equations or steps are shown that reduce a claimed prediction to a fitted parameter by construction, nor does any load-bearing premise collapse to a self-citation chain or self-definitional loop. The derivation chain is therefore self-contained against external benchmarks and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Review performed on abstract alone; ledger entries are therefore limited to what is explicitly invoked in the provided text.

axioms (1)
  • domain assumption Denoising can be viewed as gradient descent on a free energy landscape
    Explicitly stated as the perspective taken in the abstract.
invented entities (1)
  • Critical Boundary Detector (CBD) no independent evidence
    purpose: practical diagnostics for score-direction instability
    Introduced in the abstract as a new tool motivated by the caustic perspective; no independent evidence supplied.

pith-pipeline@v0.9.1-grok · 5676 in / 1294 out tokens · 22420 ms · 2026-06-27T07:10:17.171540+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 15 canonical work pages · 2 internal anchors

  1. [1]

    What regularized auto-encoders learn from the data-generating distribution.J

    Guillaume Alain and Yoshua Bengio. What regularized auto-encoders learn from the data-generating distribution.J. Mach. Learn. Res., 15(1):3563–3593, 2014. doi: 10.5555/2627435.2750359. URL https://dl.acm.org/doi/10.5555/2627435.2750359

    work page doi:10.5555/2627435.2750359 2014

  2. [2]

    Albergo, Nicholas M

    Michael S. Albergo, Nicholas M. Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions.J. Mach. Learn. Res., 26:209:1–209:80, 2025. URL https://jmlr.org/papers/v26/23-1605.html

    2025

  3. [3]

    The statistical thermodynamics of generative diffusion models: Phase transitions, symmetry breaking, and critical instability.Entropy, 27(3), 2025

    Luca Ambrogioni. The statistical thermodynamics of generative diffusion models: Phase transitions, symmetry breaking, and critical instability.Entropy, 27(3), 2025. ISSN 1099-4300. doi: 10.3390/ e27030291. URLhttps://www.mdpi.com/1099-4300/27/3/291

    2025

  4. [4]

    How Out-of-Equilibrium Phase Transitions can Seed Pattern Formation in Trained Diffusion Models

    Luca Ambrogioni. How out-of-equilibrium phase transitions can seed pattern formation in trained diffusion models.CoRR, abs/2603.20092, 2026. doi: 10.48550/ARXIV.2603.20092. URL https: //doi.org/10.48550/arXiv.2603.20092

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2603.20092 2026

  5. [5]

    Anderson

    Brian D.O. Anderson. Reverse-time diffusion equation models.Stochastic Processes and their Applications, 12(3):313–326, 1982. ISSN 0304-4149. doi: https://doi.org/10.1016/0304-4149(82) 90051-5. URLhttps://www.sciencedirect.com/science/article/pii/0304414982900515

    work page doi:10.1016/0304-4149(82 1982

  6. [6]

    Arnold, Sabir M

    Vladimir I. Arnold, Sabir M. Gusein-Zade, and Alexander N. Varchenko.Singularities of Differentiable Maps, Volume I: The Classification of Critical Points, Caustics and Wave Fronts, volume 82 ofMonographs in Mathematics. Birkhäuser, Boston, MA, 1985. ISBN 978-0-8176-3187-9. doi: 10.1007/978-1-4612-5136-5

    work page doi:10.1007/978-1-4612-5136-5 1985

  7. [7]

    Davide Barilari, Ugo Boscain, and Robert W. Neel. Small-time heat kernel asymptotics at the sub-riemannian cut locus.Journal of Differential Geometry, 92(3):373–416, Nov 2012. doi: 10.4310/ jdg/1354110195

    arXiv 2012

  8. [8]

    Generative diffusion in very large dimensions.Journal of Statistical Mechanics: Theory and Experiment, 2023(9):093402, oct 2023

    Giulio Biroli and Marc Mézard. Generative diffusion in very large dimensions.Journal of Statistical Mechanics: Theory and Experiment, 2023(9):093402, oct 2023. doi: 10.1088/1742-5468/acf8ba. URL https://doi.org/10.1088/1742-5468/acf8ba

    work page doi:10.1088/1742-5468/acf8ba 2023

  9. [9]

    Biroli, T

    Giulio Biroli, Tony Bonnaire, Valentin de Bortoli, and Marc Mézard. Dynamical regimes of diffusion models.Nature Communications, 15(1):9957, Nov 2024. ISSN 2041-1723. doi: 10.1038/s41467-024-54281-3. URLhttps://doi.org/10.1038/s41467-024-54281-3

    work page doi:10.1038/s41467-024-54281-3 2024

  10. [10]

    A Transformation for Extracting New Descriptors of Shape

    Harry Blum. A Transformation for Extracting New Descriptors of Shape. In Weiant Wathen-Dunn, editor,Models for the Perception of Speech and Visual Form, pages 362–380. MIT Press, Cambridge, 1967. 13

    1967

  11. [11]

    lambda-medial axis

    Frédéric Chazal and André Lieutier. The “lambda-medial axis”.Graph. Models, 67(4):304–331, July

  12. [12]

    doi: 10.1016/j.gmod.2005.01.002

    ISSN 1524-0703. doi: 10.1016/j.gmod.2005.01.002. URLhttps://doi.org/10.1016/j.gmod. 2005.01.002

    work page doi:10.1016/j.gmod.2005.01.002 2005

  13. [13]

    Asymptotic analysis of oscillatory integrals via the Newton polyhedra of the phase and the amplitude.Journal of the Mathematical Society of Japan, 65 (2):521 – 562, 2013

    Koji Cho, Joe Kamimoto, and Toshihiro Nose. Asymptotic analysis of oscillatory integrals via the Newton polyhedra of the phase and the amplitude.Journal of the Mathematical Society of Japan, 65 (2):521 – 562, 2013. doi: 10.2969/jmsj/06520521. URLhttps://doi.org/10.2969/jmsj/06520521

    work page doi:10.2969/jmsj/06520521 2013

  14. [14]

    Bronstein, and Avishek Joey Bose

    Oscar Davis, Samuel Kessler, Mircea Petrache, İsmail İlkan Ceylan, Michael M. Bronstein, and Avishek Joey Bose. Fisher flow matching for generative modeling over discrete data. In Amir Globersons, Lester Mackey, Danielle Belgrave, Angela Fan, Ulrich Paquet, Jakub M. Tomczak, and Cheng Zhang, editors,Advances in Neural Information Processing Systems 37: An...

    2024

  15. [15]

    Diffusion models beat gans on image synthe- sis

    Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthe- sis. In Marc’Aurelio Ranzato, Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jen- nifer Wortman Vaughan, editors,Advances in Neural Information Processing Systems 34: An- nual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14...

    2021

  16. [16]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In Ruslan Salakhutdinov, Zico Kolter, Katherine A. Heller, Adrian Weller, Nu...

    2024

  17. [17]

    Curvature measures.Transactions of the American Mathematical Society, 93(3): 418–491, 1959

    Herbert Federer. Curvature measures.Transactions of the American Mathematical Society, 93(3): 418–491, 1959. ISSN 00029947. URLhttp://www.jstor.org/stable/1993504

    arXiv 1959

  18. [18]

    Prompt- to-prompt image editing with cross-attention control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt- to-prompt image editing with cross-attention control. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. URL https://openreview.net/forum?id=_CDixzkzeyb

    2023

  19. [19]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Pith/arXiv arXiv 2022

  20. [20]

    Denoising diffusion probabilistic mod- els

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic mod- els. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors,Advances in Neural Information Processing Systems 33: An- nual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, Decem- ber 6-12, 2020, virtua...

    2020

  21. [21]

    Estimation of non-normalized statistical models by score matching.J

    Aapo Hyvärinen. Estimation of non-normalized statistical models by score matching.J. Mach. Learn. Res., 6:695–709, 2005. URLhttps://jmlr.org/papers/v6/hyvarinen05a.html

    2005

  22. [22]

    Asymptotic expansion of oscillatory integrals with singular phases.Kyushu Journal of Mathematics, 77(2):319–329, 2023

    Joe Kamimoto and Hiromichi Mizuno. Asymptotic expansion of oscillatory integrals with singular phases.Kyushu Journal of Mathematics, 77(2):319–329, 2023. ISSN 1340-6116. doi: 10.2206/ kyushujm.77.319. URLhttps://cir.nii.ac.jp/crid/1390297814401676928

    arXiv 2023

  23. [23]

    Toric resolution of singularities in a certain class ofc∞functions and asymptotic analysis of oscillatory integrals, 2012

    Joe Kamimoto and Toshihiro Nose. Toric resolution of singularities in a certain class ofc∞functions and asymptotic analysis of oscillatory integrals, 2012. URLhttps://arxiv.org/abs/1208.3924

    arXiv 2012

  24. [24]

    Elucidating the design space of diffusion- based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion- based generative models. In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Information Processing Systems 35: Annual Conference on 14 Neural Information Processing Systems 2022, NeurIPS 2022, New Orlea...

    2022

  25. [25]

    In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 24174–24184. IEEE, 2024. doi: 10.1109/CVPR52733.2024.02282. URLhttps://doi.org/10...

    work page doi:10.1109/cvpr52733.2024.02282 2024

  26. [26]

    Applying guidance in a limited interval improves sample and distribution quality in diffusion models

    Tuomas Kynkäänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. In Amir Globersons, Lester Mackey, Danielle Belgrave, Angela Fan, Ulrich Paquet, Jakub M. Tomczak, and Cheng Zhang, editors,Advances in Neural Information Proce...

    2024

  27. [27]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. URLhttps: //openreview.net/forum?id=XVjTT1nw5z

    2023

  28. [28]

    John N. Mather. Distance from a submanifold in euclidean space. InSingularities, Part 2 (Arcata, Calif., 1981), volume 40 ofProceedings of Symposia in Pure Mathematics, pages 199–216. American Mathematical Society, Providence, RI, 1983

    1981

  29. [29]

    Sdedit: Guided image synthesis and editing with stochastic differential equations

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. InThe Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. URLhttps://openreview.net/forum?id=aBsCjcPu_tE

    2022

  30. [30]

    InProceedings of the SIGGRAPH Asia 2025 Conference Papers (SA Conference Papers ’25)

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InIEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 4172–4182. IEEE, 2023. doi: 10.1109/ICCV51070.2023.00387. URLhttps://doi.org/10.1109/ ICCV51070.2023.00387

    work page doi:10.1109/iccv51070.2023.00387 2023

  31. [31]

    Interpreting and improving diffusion models from an optimization perspective

    Frank Permenter and Chenyang Yuan. Interpreting and improving diffusion models from an optimization perspective. In Ruslan Salakhutdinov, Zico Kolter, Katherine A. Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors,Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024, Pro...

    2024

  32. [32]

    Score-based generative models detect manifolds

    Jakiw Pidstrigach. Score-based generative models detect manifolds. In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022. URLhttp://...

    2022

  33. [33]

    Spontaneous symmetry breaking in generative diffusion models

    Gabriel Raya and Luca Ambrogioni. Spontaneous symmetry breaking in generative diffusion models. In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine, editors,Advances in Neural Information Processing Systems 36: Annual Con- ference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, Dece...

    2023

  34. [34]

    A ConvNet for the 2020s

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674–10685. IEEE, 2022. doi: 10.1109/CVPR52688.2022.01042. URLhttps://doi.org/10.1109/ C...

    work page doi:10.1109/cvpr52688.2022.01042 2022

  35. [35]

    A phase transition in diffusion models reveals the hierarchical nature of data.Proceedings of the National Academy of Sciences, 122(1):e2408799121, January 2025

    Antonio Sclocchi, Alessandro Favero, and Matthieu Wyart. A phase transition in diffusion models reveals the hierarchical nature of data.Proceedings of the National Academy of Sciences, 122 (1):e2408799121, 2025. doi: 10.1073/pnas.2408799121. URLhttps://www.pnas.org/doi/abs/10. 1073/pnas.2408799121

    work page doi:10.1073/pnas.2408799121 2025

  36. [36]

    Weiss, Niru Maheswaranathan, and Surya Ganguli

    Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Francis R. Bach and David M. Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, JMLR Workshop and Conference Proceedings, pages 2256–22...

    2015

  37. [37]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d’Alché-Buc, Emily B. Fox, and Roman Garnett, editors,Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, Decem...

    2019

  38. [38]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7,

    2021

  39. [39]

    URLhttps://openreview.net/forum?id=PxTIG12RRHS

    OpenReview.net, 2021. URLhttps://openreview.net/forum?id=PxTIG12RRHS

    2021

  40. [40]

    Diffusion models encode the intrinsic dimension of data manifolds

    Jan Stanczuk, Georgios Batzolis, Teo Deveney, and Carola-Bibiane Schönlieb. Diffusion models encode the intrinsic dimension of data manifolds. In Ruslan Salakhutdinov, Zico Kolter, Katherine A. Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors,Forty-first International Conference on Machine Learning, ICML 2024, Vienna, ...

    2024

  41. [41]

    The information dynamics of generative diffusion.Entropy, 28(2), 2026

    Dejan Stančević and Luca Ambrogioni. The information dynamics of generative diffusion.Entropy, 28(2), 2026. ISSN 1099-4300. doi: 10.3390/e28020195. URLhttps://www.mdpi.com/1099-4300/ 28/2/195

    work page doi:10.3390/e28020195 2026

  42. [42]

    Manifolds, random matrices and spectral gaps: The geometric phases of generative diffusion

    Enrico Ventura, Beatrice Achilli, Gianluigi Silvestri, Carlo Lucibello, and Luca Ambrogioni. Manifolds, random matrices and spectral gaps: The geometric phases of generative diffusion. InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net, 2025. URLhttps://openreview.net/forum?id=KlN00vQEY2

    2025

  43. [43]

    The discrete and continuous brain: From decisions to movement—and back again,

    Pascal Vincent. A connection between score matching and denoising autoencoders.Neural Comput., 23(7):1661–1674, 2011. doi: 10.1162/NECO\_A\_00142. URLhttps://doi.org/10.1162/NECO_ a_00142

    work page doi:10.1162/neco 2011

  44. [44]

    Cambridge University Press, 2009

    Sumio Watanabe.Algebraic Geometry and Statistical Learning Theory. Cambridge University Press, 2009

    2009

  45. [45]

    SIAM, 2001

    Roderick Wong.Asymptotic approximations of integrals, volume 34 ofClassics in applied mathematics. SIAM, 2001. ISBN 978-0-89871-497-5

    2001

  46. [46]

    The geometry of phase transitions in diffusion models: Tubular neighbourhoods and singularities.Trans

    Manato Yaguchi, Kotaro Sakamoto, Ryosuke Sakamoto, Masato Tanabe, Masatomo Akagawa, Yusuke Hayashi, Masahiro Suzuki, and Yutaka Matsuo. The geometry of phase transitions in diffusion models: Tubular neighbourhoods and singularities.Trans. Mach. Learn. Res., 2025, 2025. URL https://openreview.net/forum?id=ahVFKFLYk2

    2025

  47. [47]

    Concurrence of Symmetry Breaking and Nonlocality Phase Transitions in Diffusion Models

    Yi-Fan Zhang, Fangjun Hu, Guangkuo Liu, Mert Okyay, and Xun Gao. Concurrence of symmetry breaking and nonlocality phase transitions in diffusion models.CoRR, abs/2605.04830, 2026. doi: 10.48550/ARXIV.2605.04830. URLhttps://doi.org/10.48550/arXiv.2605.04830. 16 A Additional Theoretical Details A.1 Settings in Subsection 2.3 In this subsection, we explain t...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2605.04830 2026