Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Querying a diffusion network at a shifted timestep cancels the variance inflation of high-temperature sampling, turning temperature into a training-free diversity knob.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 08:45 UTC pith:N5W7WW4S

load-bearing objection Practical zero-cost diversity knob for pretrained diffusion via time-shifted temperature; clean under mode separation but main gains come from the high-noise regime the math does not cover. the 3 major comments →

arxiv 2607.10853 v1 pith:N5W7WW4S submitted 2026-07-12 cs.CV cs.LG

Diversify Diffusion with Temperature Sampling and Variance-Corrective Time Shifting

Peizhuo Li , Emre Aksan , Alexandru-Eugen Ichim , Thabo Beeler , Olga Sorkine-Hornung This is my paper
classification cs.CV cs.LG
keywords diffusion modelstemperature samplingscore scalingtime shiftingdiversityflow matchinginference-time controlmode reweighting
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Diffusion models copy their training data so faithfully that rare modes are hard to sample. Raising the temperature of the target distribution (raising p0 to a power γ less than one) would flatten common modes and lift rare ones, but simply scaling the score by γ also inflates each mode’s variance and wrecks sample quality. This paper shows that querying the pretrained score network at a carefully shifted, less-noisy timestep and then scaling by γ cancels that variance inflation while keeping the desired mode reweighting. The fix needs no retraining, works for both diffusion and flow-matching backbones, and gives users a practical dial for diversity. Applying the temperature change only in early (high-noise) steps mainly changes global composition; applying it late mainly changes local appearance.

Core claim

Naive temperature sampling reweights modes correctly but expands per-mode variance, so the reverse process no longer matches the tempered marginal and samples degrade. Under a well-separated-modes approximation, that expansion is exactly a rescaling of noise level. Querying the pretrained network at the shifted time whose noise level is γ times smaller, then multiplying the resulting score by γ, restores the original variance schedule while preserving the reweighting. The same construction carries over to variance-preserving diffusion and rectified-flow models, turning temperature into a zero-cost diversity control.

What carries the argument

Variance-corrective time shifting: query the score (or velocity) network at a shifted timestep t̃ chosen so σ̃_t² = γ σ_t² (or the VP / flow-matching analogue), then scale the output by γ. This cancels the variance expansion of direct score scaling while keeping the mode reweighting of the tempered target p0^γ.

Load-bearing premise

The method rests on modes being well separated at the noise level where temperature is applied; if modes overlap strongly, the simple time-shift formula no longer recovers the tempered distribution.

What would settle it

On a multimodal density whose modes deliberately overlap at the noise scales used for tempering, measure whether the empirical mode frequencies still match the tempered weights π_k^γ and whether sample quality remains comparable to the untempered baseline; a large mismatch would refute the claim that the time shift recovers the tempered marginal.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Any pretrained diffusion or flow-matching checkpoint can be used as a diversity-controllable generator by wrapping its score or velocity call with a one-line time-shift and γ-scale.
  • Users can trade composition-level diversity against local detail diversity by choosing which segment of the denoising trajectory receives γ < 1.
  • Condition fidelity need not collapse when diversity is increased, provided γ is not driven too low and the correction is applied.
  • Competing inference-time diversity methods that lack the variance correction will continue to produce blurrier or off-manifold samples at high temperature.

Where Pith is reading between the lines

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

  • The same time-shift identity may extend to products of guidance signals (e.g., classifier-free guidance times temperature), giving a joint knob for fidelity and diversity.
  • How much diversity headroom a prompt has may itself become a diagnostic of how peaked the learned conditional is, revealing structure in the training distribution without extra training.
  • If the well-separated-modes premise fails for dense manifolds (e.g., highly entangled video or 3D), a learned or adaptive shift schedule might still recover a usable approximation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes variance-corrective time shifting as a training-free way to sample from the high-temperature target p0^γ (0<γ<1) with pretrained diffusion and flow-matching models. Naive score scaling reweights modes but inflates per-mode variance and degrades quality; the fix queries the network at a shifted noise level (σ̃_t² = γ σ_t², with VP/flow analogues) and multiplies the score by γ. Under a well-separated Gaussian-mixture approximation this cancels the variance inflation while preserving mode reweighting. The authors further show that applying γ(t)<1 at early vs. late denoising stages yields compositional vs. local diversity. Experiments on DiT, Stable Diffusion (1.5/XL/3.5) and a motion diffusion model report gains in Vendi diversity at modest cost to VQA/accuracy and CMMD, outperforming CADS and Feynman–Kac correctors in the reported settings.

Significance. If the construction is reliable, it supplies a simple, zero-retraining diversity knob that works across both classical diffusion and modern flow-matching backbones, including practical stage-wise control of composition vs. appearance. The identification of the variance-expansion failure mode of naive temperature sampling is useful, and the appendices give concrete, implementable score/velocity wrappers for VP and rectified-flow models. The toy-mixture reweighting check (Table 1) and the direct-tempering/TSR ablation (Fig. 6) are clear strengths. The main limitation is that the rigorous recovery argument is stated only under well-separated modes, while the primary diversity gains are obtained in the early, high-noise regime where that assumption fails; closing or clearly scoping that gap would make the contribution substantially more solid for the community.

major comments (3)
  1. [Sec. 3.1, Eqs. 8–13; App. A; Sec. 4.2] Sec. 3.1 (Eqs. 8–13) and Appendix A justify the time-shift formula only when σ_t is small enough that modes remain separated and cross terms in p_t(x)^γ are negligible. The paper’s main diversity results, however, apply the same rule in the early (high-noise) third of the trajectory (Sec. 3.3, Sec. 4.2, Figs. 2/4/5), where modes of any realistic p0 are fully mixed and the binomial expansion of App. A does not apply. There is therefore no guarantee that querying at σ̃_t = √γ σ_t recovers ∇ log p_t^(γ) in the regime used for compositional diversity. Either (i) provide analysis or diagnostics for the overlapping-mode / high-noise regime, or (ii) reframe early-stage tempering as an empirically effective heuristic whose mode-reweighting is not theoretically controlled by the current derivation, and restrict the ‘recovers the tempered marginal’ claim to the low-noise setting where the approxim
  2. [Sec. 4.2, Table 2, Fig. 4] The quantitative claim of ‘consistent gains at minimal cost to sample quality and condition fidelity’ rests on a small set of hand-chosen γ values per model (γ=0.20/0.88/0.94/0.78) selected after inspecting the Vendi–fidelity curves in Fig. 4. Table 2 reports a single operating point per model without a full Pareto comparison against CADS/FK across matched fidelity levels, and CMMD is acknowledged to penalize the intended deviation from p0. A clearer multi-γ trade-off table (or fixed-fidelity Vendi comparison) would make the superiority claim load-bearing rather than pointwise.
  3. [Sec. 4.2 (Motion generation); Fig. 8] Motion-generation evaluation is only qualitative (Fig. 8 and supplementary video). Given that the abstract and contributions list Motion Diffusion as a third modality demonstrating ‘consistent gains,’ the absence of any diversity/fidelity metric (even a simple feature-space Vendi or classifier-based condition score) leaves that part of the claim unsupported in the main text.
minor comments (5)
  1. [Sec. 4.2] Sec. 4.2: typo ‘In this seciton’ → ‘section’; also ‘controled’ → ‘controlled’ in the same paragraph.
  2. [Fig. 1] Fig. 1 (left) is a useful 1D illustration of variance inflation, but the caption does not state whether the curves are analytic or sampled; a one-line clarification would help.
  3. [Sec. 2] Related work: TSR [37] is correctly identified as suffering from variance expansion; a short explicit statement that TSR’s γ(t) schedule is a smooth direct-tempering schedule (as later used in Fig. 6) would make the distinction sharper earlier.
  4. [Sec. 3.1; App. B–C] Eq. (13) and the VP/flow wrappers (App. B–C) are clear; adding a one-line pseudocode box for the ϵ_temp / v_temp wrapper in the main text would improve reproducibility for practitioners who skip the appendices.
  5. [Fig. 7; Sec. 4.3] Fig. 7 is dense; the main-text claim that early-stage tempering usually gives the largest diversity gain would be easier to verify if the per-tier peak-γ values were tabulated.

Circularity Check

0 steps flagged

No circularity: tempered target and time-shift formula are derived from first principles under a stated mixture approximation; γ is a free sampling knob, not a fitted input renamed as prediction.

full rationale

The load-bearing construction (Sec. 3.1, Eqs. 1–13; App. A–C) starts from the external definition p^(γ)_0 ∝ p_0^γ, shows that direct score scaling γ∇log p_t is not the score of the noised tempered marginal, and under the explicit well-separated-modes premise derives the variance-corrective shift σ_˜t² = γ σ_t² (and VP/flow analogues). That derivation is algebraic from the mixture approximation and Tweedie’s formula; it does not fit any parameter to the Vendi/CMMD/VQA evaluation metrics, nor does it define the tempered score in terms of those metrics. γ values used in Sec. 4 are free inference knobs chosen by grid search after the method is fixed, which is ordinary hyperparameter selection, not a fitted-input-called-prediction. Related-work citations (TSR, CADS, FK) are baselines or contrasts, not load-bearing uniqueness theorems from overlapping authors. The skeptic’s concern—that the well-separated assumption fails at the high-noise stages where the method is applied—is a correctness/regime-validity issue, not circularity: the paper does not claim the early-stage score equals ∇log p_t^(γ) by construction outside its stated premise, and the empirical gains are measured against external models and metrics. No equation reduces the claimed tempered score to a quantity already fitted to the evaluation. Score 0 is therefore the honest finding.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The central claim rests on standard diffusion score identities plus one modeling assumption (well-separated modes) and a small set of per-model free sampling parameters (γ and stage). No new physical entities are postulated; the method is an inference-time reparameterization of an existing score network.

free parameters (2)
  • inverse temperature γ (per model) = model-specific (0.20–0.94)
    Chosen by grid search on the Vendi–fidelity trade-off (Fig. 4); reported operating points are γ=0.20 (DiT), 0.88 (SD1.5), 0.94 (SDXL), 0.78 (SD3.5). Not derived from first principles.
  • tempering stage schedule γ(t) = early-stage default for Sec. 4.2
    Three equal-length early/middle/late segments; tempering applied in one segment (usually early) after empirical comparison. Piecewise schedule obtained by grid search, not theory.
axioms (3)
  • domain assumption Well-separated modes: at the noise levels of interest, mixture components of p_t remain disjoint enough that cross terms in p_t(x)^γ are exponentially negligible (Sec. 3.1, App. A).
    Load-bearing for equating the tempered noisy marginal to a variance-rescaled direct tempering; without it the time-shift formula is only heuristic.
  • standard math Standard VE/VP/rectified-flow forward kernels and Tweedie score conversion from an ε- or velocity-predictor.
    Background identities used throughout Secs. 3, B, C; not re-proved.
  • domain assumption Natural image and human-motion distributions behave sufficiently like unbalanced mixtures of well-separated modes for the approximation to be useful in practice.
    Stated as empirical justification after the toy Gaussian mixture (Sec. 3.2); not independently verified beyond the reported experiments.

pith-pipeline@v1.1.0-grok45 · 20566 in / 2785 out tokens · 34754 ms · 2026-07-14T08:45:40.269108+00:00 · methodology

0 comments
read the original abstract

Diffusion models faithfully reproduce their training distribution, but also inherit its imbalances and leave rare or under-represented modes hard to reach. A natural inference-time remedy is to sample from the high-temperature target $p^{(\gamma)}_0(x) \propto p_0(x)^{\gamma}$ for $0 < \gamma < 1$, which flattens dominant modes and lifts rare ones. However, naive score scaling while correctly reweighting modes also inflates the per-mode variance, breaking the reverse diffusion process and degrading sample quality. We introduce variance-corrective time shifting, a training-free fix that queries the network at a shifted timestep and scales the resulting score by $\gamma$, canceling the variance inflation while preserving the mode reweighting. The correction turns simple temperature sampling into a practical diversity knob for pretrained diffusion and flow-matching backbones with no retraining, and we demonstrate consistent gains at minimal cost to sample quality and condition fidelity across DiT, Stable Diffusion and Motion Diffusion models. We further show that the timing of the temperature intervention enables coarse-to-fine control: high-noise stages drive compositional diversity across modes, while low-noise stages drive local appearance variation under a fixed composition.

Figures

Figures reproduced from arXiv: 2607.10853 by Alexandru-Eugen Ichim, Emre Aksan, Olga Sorkine-Hornung, Peizhuo Li, Thabo Beeler.

Figure 1
Figure 1. Figure 1: Left: The effect of increasing temperature ( [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The effect of applying temperature scaling at different denoising stages. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison across model architectures. FK is not compatible with flow [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The effect of γ on diversity (Vendi, solid) and fidelity (dashed) for each model, averaged across all condition tiers. γ = 0.90 γ = 0.92 γ = 0.94 γ = 0.96 γ = 1 (baseline) Early Middle Late [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative effect of γ across denoising stages. Prompt: A bicycle leaning against a wall (SDXL). 4.3 Ablation study Choice of γ and stage. We sweep γ over a range of values for each model and compare three denoising stages (early, mid and late), and different conditioning text and class tiers (loose, medium, tight), depending on the diversity headroom. Full setup and conditioning sets are given in Sec. E.… view at source ↗
Figure 6
Figure 6. Figure 6: Effect of variance-corrective time shifting on DDIM across tempering methods and [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Per-tier breakdown of the γ sweep (full version of [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison on the motion diffusion model [ [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗

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

37 extracted references · 13 linked inside Pith

  1. [1]

    Universal guidance for diffusion models

    Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 843–852, 2023

    2023

  2. [2]

    Particle guidance: non-I.I.D

    Gabriele Corso, Yilun Xu, Valentin De Bortoli, Regina Barzilay, and Tommi Jaakkola. Particle guidance: non-I.I.D. diverse sampling with diffusion models. InThe Twelfth International Conference on Learning Representations (ICLR), 2024

    2024

  3. [3]

    Imagenet: A large- scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large- scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009

    2009

  4. [4]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780–8794, 2021

    2021

  5. [5]

    Scaling rectified flow trans- formers 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, et al. Scaling rectified flow trans- formers for high-resolution image synthesis. InForty-first international conference on machine learning, 2024

    2024

  6. [6]

    The vendi score: A diversity evaluation metric for machine learning.arXiv preprint arXiv:2210.02410, 2022

    Dan Friedman and Adji Bousso Dieng. The vendi score: A diversity evaluation metric for machine learning.arXiv preprint arXiv:2210.02410, 2022

    Pith/arXiv arXiv 2022

  7. [7]

    Generative adversarial nets.Advances in neural information processing systems, 27, 2014

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets.Advances in neural information processing systems, 27, 2014

    2014

  8. [8]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems, 30, 2017

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems, 30, 2017

    2017

  9. [9]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015

    Pith/arXiv arXiv 2015

  10. [10]

    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

  11. [11]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  12. [12]

    The curious case of neural text degeneration.arXiv preprint arXiv:1904.09751, 2019

    Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration.arXiv preprint arXiv:1904.09751, 2019

    Pith/arXiv arXiv 1904

  13. [13]

    T2i-compbench: A compre- hensive benchmark for open-world compositional text-to-image generation.Advances in Neural Information Processing Systems, 36:78723–78747, 2023

    Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A compre- hensive benchmark for open-world compositional text-to-image generation.Advances in Neural Information Processing Systems, 36:78723–78747, 2023

    2023

  14. [14]

    Rethinking fid: Towards a better evaluation metric for image generation

    Sadeep Jayasumana, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv Kumar. Rethinking fid: Towards a better evaluation metric for image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9307–9315, 2024

    2024

  15. [15]

    Optimization by simulated annealing

    Scott Kirkpatrick, C Daniel Gelatt Jr, and Mario P Vecchi. Optimization by simulated annealing. science, 220(4598):671–680, 1983

    1983

  16. [16]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InEuropean conference on computer vision, pages 740–755. Springer, 2014. 10

    2014

  17. [17]

    Evaluating text-to-visual generation with image-to-text generation

    Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2024

    2024

  18. [18]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Pith/arXiv arXiv 2022

  19. [19]

    Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022

    Pith/arXiv arXiv 2022

  20. [20]

    Progen2: exploring the boundaries of protein language models.Cell systems, 14(11):968–978, 2023

    Erik Nijkamp, Jeffrey A Ruffolo, Eli N Weinstein, Nikhil Naik, and Ali Madani. Progen2: exploring the boundaries of protein language models.Cell systems, 14(11):968–978, 2023

    2023

  21. [21]

    Heavy-tailed diffusion models.arXiv preprint arXiv:2410.14171, 2024

    Kushagra Pandey, Jaideep Pathak, Yilun Xu, Stephan Mandt, Michael Pritchard, Arash Vahdat, and Morteza Mardani. Heavy-tailed diffusion models.arXiv preprint arXiv:2410.14171, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    2023

  23. [23]

    A self- supervised descriptor for image copy detection

    Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self- supervised descriptor for image copy detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14532–14542, 2022

    2022

  24. [24]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021

    2021

  26. [26]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  27. [27]

    Seyedmorteza Sadat, Jakob Buhmann, Derek Bradley, Otmar Hilliges, and Romann M. Weber. CADS: unleashing the diversity of diffusion models through condition-annealed sampling. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024, 2024

    2024

  28. [28]

    Generating high fidelity data from low-density regions using diffusion models

    Vikash Sehwag, Caner Hazirbas, Albert Gordo, Firat Ozgenel, and Cristian Canton. Generating high fidelity data from low-density regions using diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  29. [29]

    Long horizon temperature scaling

    Andy Shih, Dorsa Sadigh, and Stefano Ermon. Long horizon temperature scaling. InInterna- tional conference on machine learning, pages 31422–31434. PMLR, 2023

    2023

  30. [30]

    Feynman- kac correctors in diffusion: Annealing, guidance, and product of experts.arXiv preprint arXiv:2503.02819, 2025

    Marta Skreta, Tara Akhound-Sadegh, Viktor Ohanesian, Roberto Bondesan, Alán Aspuru- Guzik, Arnaud Doucet, Rob Brekelmans, Alexander Tong, and Kirill Neklyudov. Feynman- kac correctors in diffusion: Annealing, guidance, and product of experts.arXiv preprint arXiv:2503.02819, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Deep unsuper- vised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsuper- vised learning using nonequilibrium thermodynamics. InInternational conference on machine learning, pages 2256–2265. pmlr, 2015

    2015

  32. [32]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Pith/arXiv arXiv 2010

  33. [33]

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

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020. 11

    Pith/arXiv arXiv 2011

  34. [34]

    The information dynamics of generative diffusion

    Dejan Stancevic and Luca Ambrogioni. The information dynamics of generative diffusion. arXiv preprint arXiv:2508.19897, 2025

    arXiv 2025

  35. [35]

    Human motion diffusion model.arXiv preprint arXiv:2209.14916, 2022

    Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model.arXiv preprint arXiv:2209.14916, 2022

    Pith/arXiv arXiv 2022

  36. [36]

    A connection between score matching and denoising autoencoders.Neural computation, 23(7):1661–1674, 2011

    Pascal Vincent. A connection between score matching and denoising autoencoders.Neural computation, 23(7):1661–1674, 2011

    2011

  37. [37]

    Temporal score rescal- ing for temperature sampling in diffusion and flow models.arXiv preprint arXiv:2510.01184, 2025

    Yanbo Xu, Yu Wu, Sungjae Park, Zhizhuo Zhou, and Shubham Tulsiani. Temporal score rescal- ing for temperature sampling in diffusion and flow models.arXiv preprint arXiv:2510.01184, 2025. A Cross-term negligibility for Eq. (11) The mixture approximation in Sec. 3.1 relies on the claim that, under the well-separated mode assumption, the cross terms in pt(x)...

    Pith/arXiv arXiv 2025