arxiv: 2605.13448 · v1 · submitted 2026-05-13 · 📊 stat.ML · cs.LG· math.PR

Recognition: 2 theorem links

· Lean Theorem

On the Limits of Latent Reuse in Diffusion Models

Yifeng Yu , Lu Yu

Authors on Pith no claims yet

Pith reviewed 2026-05-14 18:33 UTC · model grok-4.3

classification 📊 stat.ML cs.LGmath.PR

keywords diffusion modelslatent reusedistribution shiftprincipal anglesscore errorlow-dimensional manifoldsmixed training

0 comments

The pith

Reusing a frozen source latent space for a shifted target dataset produces score error set by principal-angle misalignment and by ambient noise amplified over diffusion time.

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

The paper studies when it is safe to train a diffusion model in a low-dimensional latent space on one dataset and then reuse that same frozen latent space on a related but shifted second dataset. It shows that the resulting error in the target-domain score function is controlled by two geometric and temporal quantities: how much the source and target subspaces are misaligned (measured by principal angles) and how strongly the target's ambient noise is boosted at each diffusion time step. A reader would care because many efficient diffusion pipelines depend on exactly this kind of latent reuse to avoid retraining from scratch, yet the analysis supplies explicit conditions under which the shortcut stops working. The authors then examine mixed source-target training and show how the smallest shared latent dimension that keeps error low depends on the relative geometry of the two distributions.

Core claim

In the source-target setting where both datasets are approximately low-dimensional but may lie near different subspaces, freezing and reusing the source latent space induces a target-domain score error governed by the principal-angle misalignment between the source and target subspaces and by the target ambient noise amplified according to the diffusion time scale. The same geometric framework is then used to characterize the shared latent dimension required under mixed training.

What carries the argument

Principal angles between the source and target subspaces, which measure their misalignment, together with the diffusion time scale that amplifies ambient noise into the score error.

If this is right

Target score error increases with larger principal-angle misalignment between the source and target subspaces.
Ambient noise in the target domain contributes more to the score error at larger diffusion time steps.
The minimal shared latent dimension needed for mixed source-target training increases with greater geometric mismatch between the two distributions.
Latent reuse stays reliable mainly when the subspaces are closely aligned and diffusion schedules remain short.

Where Pith is reading between the lines

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

Developers could run PCA on source and target data first to compute principal angles and decide reuse viability before any diffusion training begins.
The same geometric bounds may apply to latent reuse in other score-based generative models such as flow-matching or continuous normalizing flows.
When principal angles exceed a threshold set by acceptable error, joint training on both datasets becomes preferable to reuse.

Load-bearing premise

The score error is governed primarily by subspace misalignment and diffusion time without other diffusion-process factors dominating the result.

What would settle it

Measure the actual target score error after reusing a source latent space on datasets whose principal angles are known, then check whether the error follows the predicted linear growth with the principal angles and the predicted increase with diffusion time step.

read the original abstract

Diffusion models are often trained in low-dimensional latent spaces, which are then reused for related but shifted datasets. In this work, we study when such latent reuse remains reliable under distribution shift. We consider a source-target setting in which both datasets are approximately low-dimensional but may lie near different subspaces. We show that freezing and reusing a source latent space induces a target-domain score error governed by two quantities: the principal-angle misalignment between the source and target subspaces, and the target ambient noise amplified by the diffusion time scale. Motivated by these limits, we further study mixed source-target training and characterize how the required shared latent dimension depends on the relative geometry of the two distributions. Our results provide theoretical guidance on when latent reuse is reliable and when learning a shared representation may be necessary.

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.

Desk Editor's Note private letter to a colleague

The paper bounds target score error under latent reuse by principal-angle misalignment plus time-scaled noise, but the bound's completeness is unclear without checks on extra kernel terms.

read the letter

The main point is a bound showing that freezing a source latent space for a shifted target dataset produces score error controlled by the principal angles between the subspaces and by ambient noise scaled up with diffusion time. This framing gives a concrete test for when reuse stays reliable versus when joint training is needed. They follow it with a mixed-training analysis that ties the required shared dimension to the same geometry. That part is useful and follows naturally from the first result. The derivation starts from standard score-matching and subspace tools, which keeps it grounded. The mixed-training section adds practical value by showing how the geometry dictates the minimal shared latent size. The soft spot is whether the bound really captures the full error. The stress-test concern is fair: the forward kernel and score-matching loss could introduce cross terms from manifold curvature or orthogonal projections that are not controlled by the stated quantities. If those terms matter, the claim that the error is governed exactly by misalignment and scaled noise does not hold. The abstract gives no error bars or numerical checks, so tightness is unknown. This is for researchers who train diffusion models across related but shifted datasets and want guidance on reuse versus retraining. A reader who works on efficient adaptation or theoretical limits will get something concrete from it. It deserves a serious referee because the geometric setup is standard and the question is relevant, even if the bound needs tightening or validation.

Referee Report

2 major / 1 minor

Summary. The manuscript analyzes limits of latent reuse in diffusion models under distribution shift. In a source-target setting where both datasets are approximately low-dimensional but may occupy different subspaces, it derives that freezing a source latent space induces target-domain score error governed by the principal-angle misalignment between subspaces and target ambient noise scaled by diffusion time. It further characterizes mixed source-target training and the dependence of required shared latent dimension on relative geometry of the two distributions.

Significance. If the derivations hold, the work supplies interpretable geometric limits on when latent reuse remains reliable versus when a shared representation must be learned. This is relevant for efficient diffusion-model training on related but shifted datasets, a frequent practical scenario. The use of principal angles to quantify misalignment is a clear strength, as is the explicit dependence on diffusion time scale.

major comments (2)

[Abstract / main theorem] Abstract and main theoretical derivation: the claim that score error is governed exactly by principal-angle misalignment plus ambient-noise amplification requires showing that forward-diffusion kernel interactions and manifold curvature introduce no uncontrolled cross terms (e.g., time-dependent projections of the score onto the orthogonal complement). The provided skeptic note indicates this control is not yet explicit; without it the governance statement is not load-bearing.
[Mixed-training section] Mixed-training analysis: the characterization of required shared latent dimension should include a concrete dependence on the relative principal angles and noise levels; if the bound reduces to a trivial function of the geometry parameters already used in the reuse case, the added value of the mixed-training section is limited.

minor comments (1)

[Notation / preliminaries] Notation for principal angles and diffusion time scale should be introduced once with a brief reminder of their definitions to aid readers unfamiliar with subspace geometry.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below and have revised the manuscript to strengthen the explicit control of cross terms and the geometric dependence in the mixed-training analysis.

read point-by-point responses

Referee: [Abstract / main theorem] Abstract and main theoretical derivation: the claim that score error is governed exactly by principal-angle misalignment plus ambient-noise amplification requires showing that forward-diffusion kernel interactions and manifold curvature introduce no uncontrolled cross terms (e.g., time-dependent projections of the score onto the orthogonal complement). The provided skeptic note indicates this control is not yet explicit; without it the governance statement is not load-bearing.

Authors: We appreciate the referee's emphasis on rigor here. In the proof of the main result (Theorem 3.1), the target score error is decomposed into the principal-angle misalignment contribution and the diffusion-time-amplified ambient noise term. Cross terms arising from the forward kernel and manifold curvature are controlled by the low-dimensional manifold assumption together with the contractivity of the Ornstein-Uhlenbeck process in the orthogonal complement; these terms are bounded by the ambient noise level times a factor that vanishes as the subspace approximation error goes to zero. To make this control fully explicit, we have added a dedicated remark following Theorem 3.1 and expanded the proof sketch in the appendix to isolate and bound each cross term. We believe this renders the governance statement load-bearing under the stated assumptions. revision: yes
Referee: [Mixed-training section] Mixed-training analysis: the characterization of required shared latent dimension should include a concrete dependence on the relative principal angles and noise levels; if the bound reduces to a trivial function of the geometry parameters already used in the reuse case, the added value of the mixed-training section is limited.

Authors: We agree that an explicit functional dependence strengthens the contribution. The minimal shared dimension in mixed training (Theorem 4.2) is characterized as d_shared >= max(d_s, d_t) + g(theta, sigma), where theta collects the principal angles between the two subspaces and sigma denotes the relative ambient noise levels. The function g is strictly increasing in sin(theta) and in the noise ratio, yielding a non-trivial geometric threshold that is strictly larger than the reuse-case requirement when misalignment is present. We have revised the statement of Theorem 4.2 and the discussion in Section 4 to display this dependence explicitly, clarifying how mixed training can still reduce total dimension relative to separate training while respecting the geometry. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation from standard score-matching and subspace geometry

full rationale

The central claim derives the target score error bound directly from principal-angle misalignment between subspaces and ambient noise scaled by diffusion time, using standard diffusion score-matching loss and low-dimensional manifold assumptions. No quoted step reduces the result to a self-defined parameter, fitted input renamed as prediction, or load-bearing self-citation chain. The analysis treats the geometric quantities as independent inputs and produces the error expression as output, without the target result feeding back into its own definition. This matches the reader's assessment that the quantities are derived rather than fitted by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard assumptions from diffusion theory and linear algebra; no new free parameters or invented entities are introduced in the abstract.

axioms (2)

domain assumption Data lie near low-dimensional linear subspaces
Invoked to define principal-angle misalignment between source and target.
domain assumption Diffusion score error decomposes into misalignment and noise terms
Standard diffusion score-matching assumption used to bound target error.

pith-pipeline@v0.9.0 · 5422 in / 1217 out tokens · 67787 ms · 2026-05-14T18:33:52.743488+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Theorem 2 ... Bstr(V1) ≥ ∫ [μ2(t)/h²(t) (d2 − Σ cos²θj) + σ⁴₂α⁴(t)/(h²(t)ẽh²(t)) (D−d1−d2 + Σ cos²θj)] dt
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Lemma 1 ... ∇log p_i,t(x) = A_i ∇log p^LD_i,t(Aᵀ_i x) − 1/ẽh_i(t) (I−A_i Aᵀ_i)x

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

67 extracted references · 67 canonical work pages · 6 internal anchors

[2]

3, 313--326

Brian DO Anderson, Reverse-time diffusion equation models, Stochastic Processes and their Applications 12 (1982), no. 3, 313--326

work page 1982
[5]

59, Institut Henri Poincar \'e , 2023, pp

Patrick Cattiaux, Giovanni Conforti, Ivan Gentil, and Christian L \'e onard, Time reversal of diffusion processes under a finite entropy condition, Annales de l'Institut Henri Poincar \'e (B) Probabilit \'e s et Statistiques, vol. 59, Institut Henri Poincar \'e , 2023, pp. 1844--1881

work page 2023
[6]

Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li, Pixart- : Weak-to-strong training of diffusion transformer for 4k text-to-image generation, European Conference on Computer Vision, Springer, 2024, pp. 74--91

work page 2024
[7]

Minshuo Chen, Kaixuan Huang, Tuo Zhao, and Mengdi Wang, Score approximation, estimation and distribution recovery of diffusion models on low-dimensional data, Proceedings of Machine Learning Research 202 (2023), 5327--5350

work page 2023
[8]

4735--4763

Hongrui Chen, Holden Lee, and Jianfeng Lu, Improved analysis of score-based generative modeling: User-friendly bounds under minimal smoothness assumptions, International Conference on Machine Learning, PMLR, 2023, pp. 4735--4763

work page 2023
[13]

7323--7334

Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang, Svdiff: Compact parameter space for diffusion fine-tuning, Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 7323--7334

work page 2023
[14]

Zhihan Huang, Yuting Wei, and Yuxin Chen, Denoising diffusion probabilistic models are optimally adaptive to unknown low dimensionality, Mathematics of Operations Research (2026)

work page 2026
[15]

Jerry Yao-Chieh Hu, Weimin Wu, Zhuoru Li, Sophia Pi, Zhao Song, and Han Liu, On statistical rates and provably efficient criteria of latent diffusion transformers (dits), Advances in Neural Information Processing Systems 37 (2024), 31562--31628

work page 2024
[18]

18262--18272

Xingjian Leng, Jaskirat Singh, Yunzhong Hou, Zhenchang Xing, Saining Xie, and Liang Zheng, Repa-e: Unlocking vae for end-to-end tuning of latent diffusion transformers, Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 18262--18272

work page 2025
[19]

Gen Li and Yuling Yan, Adapting to unknown low-dimensional structures in score-based diffusion models, Advances in Neural Information Processing Systems 37 (2024), 126297--126331

work page 2024
[20]

Taehong Moon, Moonseok Choi, Gayoung Lee, Jung-Woo Ha, and Juho Lee, Fine-tuning diffusion models with limited data, NeurIPS 2022 Workshop on Score-Based Methods, 2022

work page 2022
[21]

26517--26582

Kazusato Oko, Shunta Akiyama, and Taiji Suzuki, Diffusion models are minimax optimal distribution estimators, International Conference on Machine Learning, PMLR, 2023, pp. 26517--26582

work page 2023
[22]

Yidong Ouyang, Liyan Xie, Hongyuan Zha, and Guang Cheng, Transfer learning for diffusion models, Advances in Neural Information Processing Systems 37 (2024), 136962--136989

work page 2024
[24]

8599--8608

Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov, Grad-tts: A diffusion probabilistic model for text-to-speech, International Conference on Machine Learning, PMLR, 2021, pp. 8599--8608

work page 2021
[26]

10684--10695

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer, High-resolution image synthesis with latent diffusion models, Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10684--10695

work page 2022
[28]

Zixing Song, Ziqiao Meng, and Jos \'e Miguel Hern \'a ndez-Lobato, Domain-adapted diffusion model for protac linker design through the lens of density ratio in chemical space, Forty-second International Conference on Machine Learning, 2025

work page 2025
[30]

1648--1656

Rong Tang and Yun Yang, Adaptivity of diffusion models to manifold structures, International Conference on Artificial Intelligence and Statistics, PMLR, 2024, pp. 1648--1656

work page 2024
[31]

7, 1661--1674

Pascal Vincent, A connection between score matching and denoising autoencoders, Neural Computation 23 (2011), no. 7, 1661--1674

work page 2011
[32]

18124--18133

Pingyu Wu, Kai Zhu, Yu Liu, Liming Zhao, Wei Zhai, Yang Cao, and Zheng-Jun Zha, Improved video vae for latent video diffusion model, Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 18124--18133

work page 2025
[33]

211--215

Peng Wang, Huijie Zhang, Zekai Zhang, Siyi Chen, Yi Ma, and Qing Qu, Diffusion models learn low-dimensional distributions via subspace clustering, 2025 IEEE 10th International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP), IEEE, 2025, pp. 211--215

work page 2025
[35]

4230--4239

Enze Xie, Lewei Yao, Han Shi, Zhili Liu, Daquan Zhou, Zhaoqiang Liu, Jiawei Li, and Zhenguo Li, Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning, Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4230--4239

work page 2023
[36]

Ruofeng Yang, Bo Jiang, Cheng Chen, Ruinan Jin, Baoxiang Wang, and Shuai Li, Few-shot diffusion models escape the curse of dimensionality, Advances in Neural Information Processing Systems 37 (2024), 68528--68558

work page 2024
[38]

Yifeng Yu and Lu Yu, Advancing wasserstein convergence analysis of score-based models: Insights from discretization and second-order acceleration, Advances in Neural Information Processing Systems 38 (2026), 138411--138465

work page 2026
[39]

Score Approximation, Estimation and Distribution Recovery of Diffusion Models on Low-Dimensional Data

Minshuo Chen and Kaixuan Huang and Tuo Zhao and Mengdi Wang. Score Approximation, Estimation and Distribution Recovery of Diffusion Models on Low-Dimensional Data. Proceedings of Machine Learning Research. 2023

work page 2023
[40]

Journal of Machine Learning Research , volume=

Nearly-tight VC-dimension and pseudodimension bounds for piecewise linear neural networks , author=. Journal of Machine Learning Research , volume=

work page
[41]

2009 , publisher=

Neural network learning: Theoretical foundations , author=. 2009 , publisher=

work page 2009
[42]

Local rademacher complexities , author=

work page
[43]

Generative modeling with denoising auto- encoders and langevin sampling.arXiv preprint arXiv:2002.00107,

Generative modeling with denoising auto-encoders and langevin sampling , author=. arXiv preprint arXiv:2002.00107 , year=

work page arXiv 2002
[44]

Hierarchical Text-Conditional Image Generation with CLIP Latents

Hierarchical text-conditional image generation with clip latents , author=. arXiv preprint arXiv:2204.06125 , volume=

work page internal anchor Pith review Pith/arXiv arXiv
[45]

International Conference on Machine Learning , pages=

Grad-tts: A diffusion probabilistic model for text-to-speech , author=. International Conference on Machine Learning , pages=. 2021 , organization=

work page 2021
[46]

Geodiff: A geo- metric diffusion model for molecular conformation generation.arXiv preprint arXiv:2203.02923, 2022

Geodiff: A geometric diffusion model for molecular conformation generation , author=. arXiv preprint arXiv:2203.02923 , year=

work page arXiv
[47]

The intrinsic dimension of images and its impact on learning

The intrinsic dimension of images and its impact on learning , author=. arXiv preprint arXiv:2104.08894 , year=

work page arXiv
[48]

Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

work page
[49]

Advances in Neural Information Processing Systems , volume=

On statistical rates and provably efficient criteria of latent diffusion transformers (dits) , author=. Advances in Neural Information Processing Systems , volume=

work page
[50]

arXiv preprint arXiv:2507.07104 , year=

Vision-language-vision auto-encoder: Scalable knowledge distillation from diffusion models , author=. arXiv preprint arXiv:2507.07104 , year=

work page arXiv
[51]

arXiv preprint arXiv:2211.14169 , year=

Latent space diffusion models of cryo-em structures , author=. arXiv preprint arXiv:2211.14169 , year=

work page arXiv
[52]

Diffusion Transformers with Representation Autoencoders

Diffusion transformers with representation autoencoders , author=. arXiv preprint arXiv:2510.11690 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[53]

SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

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

work page internal anchor Pith review Pith/arXiv arXiv
[54]

PixArt-$\alpha$: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis

Pixart- : Fast training of diffusion transformer for photorealistic text-to-image synthesis , author=. arXiv preprint arXiv:2310.00426 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[55]

European Conference on Computer Vision , pages=

Pixart- : Weak-to-strong training of diffusion transformer for 4k text-to-image generation , author=. European Conference on Computer Vision , pages=. 2024 , organization=

work page 2024
[56]

Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

Improved video vae for latent video diffusion model , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

work page
[57]

Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

Repa-e: Unlocking vae for end-to-end tuning of latent diffusion transformers , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

work page
[58]

Advances in Neural Information Processing Systems , volume=

Adapting to unknown low-dimensional structures in score-based diffusion models , author=. Advances in Neural Information Processing Systems , volume=

work page
[59]

Advances in Neural Information Processing Systems , volume=

Transfer learning for diffusion models , author=. Advances in Neural Information Processing Systems , volume=

work page
[60]

arXiv preprint arXiv:2502.04491 , year=

Provable sample-efficient transfer learning conditional diffusion models via representation learning , author=. arXiv preprint arXiv:2502.04491 , year=

work page arXiv
[61]

Advances in Neural Information Processing Systems , volume=

Few-shot diffusion models escape the curse of dimensionality , author=. Advances in Neural Information Processing Systems , volume=

work page
[62]

Neural Computation , volume=

A Connection Between Score Matching and Denoising Autoencoders , author=. Neural Computation , volume=. 2011 , publisher=

work page 2011
[63]

Score-Based Generative Modeling through Stochastic Differential Equations

Score-based generative modeling through stochastic differential equations , author=. arXiv preprint arXiv:2011.13456 , year=

work page internal anchor Pith review Pith/arXiv arXiv 2011
[64]

International Conference on Machine Learning , pages=

Diffusion models are minimax optimal distribution estimators , author=. International Conference on Machine Learning , pages=. 2023 , organization=

work page 2023
[65]

Convergence of diffusion models under the manifold hypothesis in high-dimensions.arXiv preprint arXiv:2409.18804,

Convergence of diffusion models under the manifold hypothesis in high-dimensions , author=. arXiv preprint arXiv:2409.18804 , year=

work page arXiv
[66]

International Conference on Artificial Intelligence and Statistics , pages=

Adaptivity of diffusion models to manifold structures , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2024 , organization=

work page 2024
[67]

2025 IEEE 10th International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP) , pages=

Diffusion models learn low-dimensional distributions via subspace clustering , author=. 2025 IEEE 10th International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP) , pages=. 2025 , organization=

work page 2025
[68]

arXiv preprint arXiv:2504.06566 , year=

Diffusion factor models: Generating high-dimensional returns with factor structure , author=. arXiv preprint arXiv:2504.06566 , year=

work page arXiv
[69]

Generalization error bound for denoising score matching under relaxed manifold assumption.arXiv preprint arXiv:2502.13662,

Generalization error bound for denoising score matching under relaxed manifold assumption , author=. arXiv preprint arXiv:2502.13662 , year=

work page arXiv
[70]

arXiv preprint arXiv:2501.12982 , year=

Low-dimensional adaptation of diffusion models: Convergence in total variation , author=. arXiv preprint arXiv:2501.12982 , year=

work page arXiv
[71]

Mathematics of Operations Research , year=

Denoising diffusion probabilistic models are optimally adaptive to unknown low dimensionality , author=. Mathematics of Operations Research , year=

work page
[72]

arXiv preprint arXiv:2410.08727 , year=

Losing dimensions: Geometric memorization in generative diffusion , author=. arXiv preprint arXiv:2410.08727 , year=

work page arXiv
[73]

Forty-second International Conference on Machine Learning , year=

Domain-Adapted Diffusion Model for PROTAC Linker Design Through the Lens of Density Ratio in Chemical Space , author=. Forty-second International Conference on Machine Learning , year=

work page
[74]

Guided Transfer Learning for Discrete Diffusion Models

Guided Transfer Learning for Discrete Diffusion Models , author=. arXiv preprint arXiv:2512.10877 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[75]

NeurIPS 2022 Workshop on Score-Based Methods , year=

Fine-tuning diffusion models with limited data , author=. NeurIPS 2022 Workshop on Score-Based Methods , year=

work page 2022
[76]

Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

work page
[77]

Proceedings of the IEEE/CVF international conference on computer vision , pages=

Svdiff: Compact parameter space for diffusion fine-tuning , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

work page
[78]

Stochastic Processes and their Applications , volume=

Reverse-time diffusion equation models , author=. Stochastic Processes and their Applications , volume=. 1982 , publisher=

work page 1982
[79]

Annales de l'Institut Henri Poincar

Time reversal of diffusion processes under a finite entropy condition , author=. Annales de l'Institut Henri Poincar. 2023 , organization=

work page 2023
[80]

Advances in Neural Information Processing Systems , volume=

Advancing wasserstein convergence analysis of score-based models: Insights from discretization and second-order acceleration , author=. Advances in Neural Information Processing Systems , volume=

work page
[81]

International Conference on Machine Learning , pages=

Improved analysis of score-based generative modeling: User-friendly bounds under minimal smoothness assumptions , author=. International Conference on Machine Learning , pages=. 2023 , organization=

work page 2023
[82]

Convergence of denoising diffusion models under the manifold hypothesis

Convergence of denoising diffusion models under the manifold hypothesis , author=. arXiv preprint arXiv:2208.05314 , year=

work page arXiv