Pith. sign in

REVIEW 4 major objections 6 minor 40 references

AtomDiffuser: Time-Aware Degradation Modeling for Drift and Beam Damage in STEM Imaging

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read AtomDiffuser is a time-aware degradation model that predicts a global affine drift and a per-pixel beam-damage decay map between any two STEM frames, and it transfers from synthetic training to real cryo-STEM data.

desk verdict A clean synthetic pipeline for a genuinely new task, but the central disentanglement claim is not actually tested: the loss never supervises the predicted drift/damage factors, and the joint setting is only shown qualitatively. read the letter →

arxiv 2508.10359 v1 pith:SMTGA35X submitted 2025-08-14 cs.CV

Hao Wang , Hongkui Zheng , Kai He , Abolfazl Razi This is my paper
classification cs.CV
keywords scanningtransmissionelectronmicroscopybeamdamagedriftcorrectiondegradationmodelingdiffusionmodelscryo-STEMtime-resolvedimagingdeeplearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

AtomDiffuser tries to establish that the two main corruptions in time-resolved STEM imaging—sample drift and beam-induced signal loss—can be separated by modeling them as a structured, time-conditioned physical process rather than as generic noise. Given a clean reference frame $x_0$ and a later damaged frame $x_T$, the model predicts a global affine transformation $T_t$ and a pixel-wise attenuation map $\lambda_t$ for any intermediate time $t$, encoding the physical chronology of decay followed by drift. Trained entirely on synthetic sequences built from Perlin-noise decay, linear temporal interpolation, and acquisition noise, it reproduces ground-truth damage intensities with high accuracy and estimates drift to within a few pixels; the same model also produces smooth, interpretable degradation trajectories on real cryo-STEM frames of a beam-sensitive solid electrolyte. The motivation is practical: if drift can be corrected and damage mapped per atomic column over time, researchers could quantify radiation effects that currently hide inside blurry, misaligned time series. The paper's core claim is that this factored, explicit degradation model is both learnable and transferable.

What carries the argument

The load-bearing mechanism is the structured degradation function $D(x_0,t) = \Phi(\lambda_t \odot x_0; T_t)$, with attenuation applied before the affine warp. The final decay field $\lambda_T$ is a Perlin-noise mask, and intermediate states are generated by linear interpolation $\lambda_t = \lambda_T \cdot (t/T)$ and $T_t = T_T \cdot (t/T)$, so every training frame knows its time index. The predictor is a dual-stream U-Net with shared encoders for $x_0$ and $x_T$, a sinusoidal time embedding $\gamma(t)$ modulating the bottleneck, a fully connected head for the affine parameters $(\theta, t_x, t_y)$, and decoder skip connections that output the full-resolution decay map $\hat{\lambda}_t$. Tr

What would settle it

On a real cryo-STEM time series of a beam-sensitive material, independently measure atomic-column displacements by template matching on low-dose frames and measure elemental loss by electron energy-loss spectroscopy; if the model's predicted affine parameters disagree with the independently measured motion beyond its reported error, or if predicted $\lambda_t$ does not track the measured intensity loss under cumulative dose, the disentanglement claim is falsified.

Watch

Extended reading notes

Core claim

The central discovery is that inter-frame STEM degradation can be factored as $x_t = \Phi(\lambda_t \odot x_0; T_t)$, where $\odot$ is pixel-wise multiplication by a spatial survival map and $\Phi$ is an affine warp. A compact dual-stream U-Net conditioned on time $t$ can invert this process: given $x_0$ and the final degraded frame $x_T$, it recovers $\lambda_t$ and $T_t$ at arbitrary intermediate times. On synthetic benchmarks the factored representation produces damage-intensity curves with $R^2$ up to $0.9952$ and drift errors of roughly 2–7 pixels. On real cryo-STEM frames of Li$_7$La$_3$Zr$_2$O$_{12}$, the same model yields smooth intermediate damage states while progressively correcti

Load-bearing premise

The load-bearing premise is that the synthetic degradation generator—linear-in-time Perlin-noise decay followed by global affine drift, plus Poisson, jitter, and readout noise—faithfully represents how real cryo-STEM samples lose signal and move; if real physics follows a different time schedule, spatial pattern, or ordering, the model's real-data predictions of $\lambda_t$ and $T_t$ are not trustworthy.

Editorial extensions

If this is right

  • Beam damage becomes a per-pixel, time-indexed quantity (the survival map $\lambda_t$), not just a frame-averaged intensity, so damage can be localized to individual atomic columns.
  • Drift correction and damage assessment happen in one pass, so time series with no stable reference frame can still be aligned and interpreted.
  • Because the model interpolates arbitrary $t$, sparse acquisitions (roughly 2 s apart in the target data) can be turned into dense virtual time-lapse sequences.
  • The small model footprint (about 2.5M parameters) makes high-resolution inference up to $2048 \times 2048$ feasible in practice, opening a route to online or near-real-time dose monitoring.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is to reinterpret the time index $t$ as cumulative electron dose: if damage scales with dose, the predicted $\lambda_t$ could be reparameterized for dose-rate extrapolation, a test the paper does not run.
  • The affine-only drift assumption means any non-rigid component of specimen motion (scan distortion, nonlinear stage creep) must be absorbed into $\lambda_t$; comparing the residual decay map against known nonlinear distortions would show how much of what the model calls damage is actually mis-modeled motion.
  • Pairwise prediction could be chained across a multi-frame sequence to reconstruct drift trajectories and cumulative damage maps, but the paper demonstrates only a single pair, so the sequence-level behavior is an inference, not a result.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. AtomDiffuser proposes a time-conditioned U-Net that, given a clean STEM frame x0, a degraded frame xT, and a time step t, predicts an affine drift transformation T̂_t and a spatially varying decay map λ̂_t, and reconstructs the intermediate frame x̂_t = Φ(λ̂_t ⊙ x0; T̂_t). The authors introduce a synthetic degradation generator with linear temporal interpolation of Perlin-based attenuation and affine drift, plus Poisson/jitter/readout noise, and evaluate damage assessment, drift correction, and a qualitative cryo-STEM LLZO demo. The paper claims to disentangle drift and beam damage and to generalize to real-world cryo-STEM data.

Significance. If the disentanglement claim were established, the framework would be a useful interpretable tool for time-resolved STEM: it would provide explicit, spatially resolved damage maps and affine drift estimates, with a lightweight architecture (2.51M parameters) and code availability. The synthetic degradation generator and the temporal interpolation scheme are cleanly described and reproducible. However, the central scientific claim—factor-level disentanglement of drift and damage—is not quantitatively validated, and the real-data evidence is only a single qualitative two-frame demo. The contributions are therefore promising but not yet supported at the level claimed.

major comments (4)
  1. [§3.3, Eq. (9)] The training loss is only the L2 reconstruction error between x̂_t and x_t. There is no direct supervision on λ̂_t or T̂_t. The factorization x̂_t = Φ(λ̂_t ⊙ x0; T̂_t) is generally non-identifiable: particularly in atomic images with large zero-background regions, many (λ̂_t, T̂_t) pairs can produce nearly identical x̂_t. Thus the network can learn to reconstruct intermediate frames without recovering the true physical degradation factors. The paper's claim of disentangling drift and damage requires factor-level identifiability or explicit factor supervision; neither is provided. Please add quantitative validation of λ̂_t and T̂_t against ground truth in a joint setting, or add supervision/constraints that enforce a unique factorization.
  2. [§4.1–§4.2] The quantitative evaluations are deliberately separated: §4.1 disables geometric transformation (decay only), and §4.2 sets λ_t=1 (drift only). The paper's actual contribution is the joint disentanglement of co-occurring drift and damage, but the joint setting is never quantitatively evaluated. The only joint real-data result in §4.3 is qualitative with no ground truth, so it cannot establish that the predicted decay maps and affine transforms correspond to real physical processes. A joint synthetic benchmark with known λ_t and T_t, reporting errors on both factors, is needed.
  3. [§3.3, §4.1, §4.2] The high damage-assessment scores are in-distribution by construction: training and test degradation are drawn from the same generator family (linear temporal interpolation λ_t = λ_T·(t/T), T_t = T_T·(t/T), Perlin noise, and the same noise types). The drift benchmark applies synthetic affine motion to real STEM images, so it does not measure real drift. Thus the strong numbers in Table 1 and Table 2 reflect recovery within the assumed generator family, not external predictive power. The abstract's claim that the model 'generalizes well to real-world cryo-STEM data' is not supported by these experiments; only the qualitative Fig. 7 addresses real data.
  4. [§4.3, Fig. 7] The real-data demonstration consists of two consecutive cryo-STEM frames with unknown true drift and damage. The inferred intermediate frames are visually plausible, but without ground-truth factors or an independent validation metric (e.g., alignment error against known fiducials, or damage maps compared with post-hoc measurements), the result is anecdotal. The claim that the model 'supports high-resolution degradation inference and drift alignment' on real data needs either quantitative assessment or clearly bounded claims about what the qualitative demo can show.
minor comments (6)
  1. [§3.2, Eq. (4)] Equation (4) is malformed: 'ˆλt,( ˆθ, ˆtx, ˆty) ={(x 0, t;xT )' should probably be a function f(x0, xT, t) returning the tuple (λ̂_t, θ̂, t̂_x, t̂_y). Please fix notation.
  2. [§4.1, Fig. 5 caption] Typo: 'propsoed' should be 'proposed'.
  3. [§4.2, Fig. 6] The side-profile visualization applies a 1D Gaussian filter, which can make alignment appear better than it is. Report unsmoothed profiles or a quantitative alignment metric such as normalized cross-correlation.
  4. [§4.2] The text says 'Unlike the synthetic degradation setting used during training, this benchmark is constructed entirely from real STEM samples'—but the drift is synthetically applied to real images. The wording is misleading; please clarify that the images are real but the motion is synthetic.
  5. [Table 3] Inference time units and hardware are unspecified. Comparing CNN at 224×224 with AtomDiffuser at 2048×2048 is not apples-to-apples; report runtime at matched input resolution and hardware.
  6. [§3.3] The phrase 'decay-before-drift ordering' is presented as physically faithful, but no evidence or citation supports this chronology for the STEM acquisition setting. Please provide justification or soften the claim.

Circularity Check

1 steps flagged · score 6.0 of 10

The damage-assessment benchmark is a transformed version of the training reconstruction loss evaluated on the same synthetic generator, so the high scores do not independently validate disentangled physical damage; the joint drift+damage case is only qualitative.

  1. fitted input called prediction [Sec. 3.3 Eq. (9); Sec. 4.1 Eqs. (10)-(11)]
    "To generate intermediate training samples, we design both the decay map and the drift transformation to be temporally inferred... λ_t = λ_T · (t/T). ... The training objective is a simple L2 reconstruction loss between the predicted and ground-truth degraded frames: L_rec = ||x̂_t − x_t||². ... We generate degradation-only sequences ... and no geometric transformation is applied. ... given a predicted decay map λ̂_t, we define the scalar damage intensity as Δ̄_t = 1/HW Σ (1 − λ̂_t(i,j))."

    In the damage-only benchmark T is the identity, so by Eq. (3) x_t = λ_t ⊙ x0. The training loss Eq. (9) becomes ||λ̂_t ⊙ x0 − λ_t ⊙ x0||², i.e., a (support-weighted) reconstruction loss on λ̂_t. The damage score Eq. (10) is a spatial average of 1 − λ̂_t, so it is a re-parameterization of the same fitted quantity, not an independent measurement of beam damage. Moreover, the ground-truth λ_t is drawn from the same Perlin-noise/linear-interpolation generator used to synthesize training targets; the metrics therefore measure in-distribution inversion of the authors' own degradation model. The joint drift+damage case is never scored against known (λ, T): Sec. 4.3 is qualitative only. Hence the quantitative damage-assessment 'prediction' reduces by construction to the reconstruction objective.

full rationale

No load-bearing self-citation chain or imported uniqueness theorem is present; the only self-citations are background and baseline comparisons. The drift benchmark in Sec. 4.2 is a genuinely external check: real STEM images are warped with synthetic affine transforms, giving independent ground truth for the affine head. However, the central quantitative evidence for damage disentanglement is circular in the specific sense that the ground-truth decay maps are generated by the same λ_t = λ_T·(t/T), Perlin-noise process used to synthesize training frames, and the model is trained only to reconstruct x_t (Eq. 9). In the no-drift setting this makes the damage prediction a normalized version of the frame-reconstruction target. The joint disentanglement claim, which is the paper's main contribution, is never quantitatively evaluated with known (λ, T); only a qualitative real-data demonstration is provided. The high damage-assessment R² therefore reflects in-distribution recovery of the authors' own degradation model rather than an external validation of physically disentangled drift and damage. This is partial circularity, not complete: the affine-drift evaluation on real images retains independent content.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities; it defines an interpretable decay map and affine warp as model outputs. The load-bearing input is the synthetic generator, whose parameters are hand-chosen rather than fitted to real data.

free parameters (3)
  • Synthetic drift magnitudes = 5, 10, 15, 25, 50 px; rotation 5-15 degrees
    Chosen by hand for the drift benchmark; model performance is measured on these ranges, not on real measured drift.
  • Synthetic damage schedules = linear damage 0 to 0.9 over 10 frames
    The training and test damage use linear interpolation lambda_t = lambda_T * (t/T); this schedule is an assumption, not fit to real beam damage kinetics.
  • Network weights = 2.51M parameters
    Learned from synthetic data; the central claim depends on these weights generalizing to real data, a claim only shown qualitatively.
assumptions (4)
  • domain assumption STEM inter-frame degradation is exactly modeled by a global affine drift followed by pixel-wise attenuation, applied as decay-then-drift (Eq. 3-7).
    The entire method is built on this physical heuristic; real cryo-STEM data may involve non-rigid drift, local deformations, or nonlinear damage that violate this.
  • ad hoc to paper Temporal evolution is linear: lambda_t = lambda_T * (t/T) and T_t = T_T * (t/T).
    This linear interpolation is used to generate training samples and to infer intermediate states; no evidence is provided that real beam damage and drift accumulate linearly in time.
  • domain assumption Synthetic degradation (Perlin noise decay, Poisson noise, scan jitter, readout noise) is representative of real cryo-STEM degradation.
    The model's real-data generalization claim depends on this representativeness; the paper provides no quantitative comparison of synthetic versus real degradation statistics.
  • domain assumption Trained on TEMImageNet atom maps, which are representative of the structures of interest (e.g., LLZO).
    TEMImageNet is a general atomic image library; LLZO and other beam-sensitive materials may have different contrast and structure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AtomDiffuser: Time-Aware Degradation Modeling for Drift and Beam Damage in STEM Imaging." pith.science (2026). https://pith.science/paper/SMTGA35X

@misc{pith2026250810359,
  author       = {Pith},
  title        = {Pith review of: AtomDiffuser: Time-Aware Degradation Modeling for Drift and Beam Damage in STEM Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SMTGA35X}},
  note         = {Machine review of arXiv:2508.10359}
}
read the original abstract

Scanning transmission electron microscopy (STEM) plays a critical role in modern materials science, enabling direct imaging of atomic structures and their evolution under external interferences. However, interpreting time-resolved STEM data remains challenging due to two entangled degradation effects: spatial drift caused by mechanical and thermal instabilities, and beam-induced signal loss resulting from radiation damage. These factors distort both geometry and intensity in complex, temporally correlated ways, making it difficult for existing methods to explicitly separate their effects or model material dynamics at atomic resolution. In this work, we present AtomDiffuser, a time-aware degradation modeling framework that disentangles sample drift and radiometric attenuation by predicting an affine transformation and a spatially varying decay map between any two STEM frames. Unlike traditional denoising or registration pipelines, our method leverages degradation as a physically heuristic, temporally conditioned process, enabling interpretable structural evolutions across time. Trained on synthetic degradation processes, AtomDiffuser also generalizes well to real-world cryo-STEM data. It further supports high-resolution degradation inference and drift alignment, offering tools for visualizing and quantifying degradation patterns that correlate with radiation-induced atomic instabilities.

Figures

Figures reproduced from arXiv: 2508.10359 by the authors.

Figure 1
Figure 1. Differences between the diffusion process [ [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Degradation modeling. From left to right, the STEM [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Framework of the proposed AtomDiffuser. to x0: x_t = \sqrt {\bar {\alpha }_t} x_0 + \sqrt {1 - \bar {\alpha }_t} \, \epsilon , \quad \epsilon \sim \mathcal {N}(0, \mathbf {I}), (1) where α¯t denotes the cumulative product of the noise sched￾ule, and the noise factor ϵ is sampled from a standard Gaus￾sian distribution N . However, this formulation is limited in its ability to incorporate interpretable, structured deg… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Training process. For each iteration, we start with a [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Damage Assessment. (a) are the ground-truth samples, (b) are the ground-truth damage maps, (c) are the predicted damage maps [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Drift test on STEM samples. (a) are the flow maps of the predicted motion directions, (b) are the STEM samples and its generated [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Degradation inference test. For each iteration, we apply [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 35 canonical work pages

  1. [1]

    Cold diffusion: Inverting arbitrary im- age transforms without noise.Advances in Neural Informa- tion Processing Systems, 36:41259–41282, 2023

    Arpit Bansal, Eitan Borgnia, Hong-Min Chu, Jie Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Cold diffusion: Inverting arbitrary im- age transforms without noise.Advances in Neural Informa- tion Processing Systems, 36:41259–41282, 2023. 1, 4

    work page 2023

  2. [2]

    Improving Multislice Electron Ptychography with a Generative Prior

    Christian K Belardi, Chia-Hao Lee, Yingheng Wang, Justin Lovelace, Kilian Q Weinberger, David A Muller, and Carla P Gomes. Improving multislice electron ptychography with a generative prior.arXiv preprint arXiv:2507.17800, 2025. 3

    work page Pith review arXiv 2025

  3. [3]

    Custom-edit: Text-guided image edit- ing with customized diffusion models.arXiv preprint arXiv:2305.15779, 2023

    Jooyoung Choi, Yunjey Choi, Yunji Kim, Junho Kim, and Sungroh Yoon. Custom-edit: Text-guided image edit- ing with customized diffusion models.arXiv preprint arXiv:2305.15779, 2023. 3

    arXiv 2023

  4. [4]

    Addressing class im- balance and data limitations in advanced node semiconduc- tor defect inspection: A generative approach for sem images

    Bappaditya Dey, Vic De Ridder, Victor Blanco, Sandip Halder, and Bartel Van Waeyenberge. Addressing class im- balance and data limitations in advanced node semiconduc- tor defect inspection: A generative approach for sem images. In2024 International Symposium ELMAR, pages 141–148. IEEE, 2024. 3

    work page 2024

  5. [5]

    Radiation damage in the tem and sem.Micron, 35(6):399–409, 2004

    RF Egerton, P Li, and M Malac. Radiation damage in the tem and sem.Micron, 35(6):399–409, 2004. 2

    work page 2004

  6. [6]

    Bridging microscopy with molecular dynamics and quantum simulations: an atomai based pipeline.npj Computational Materials, 8(1):74, 2022

    Ayana Ghosh, Maxim Ziatdinov, Ondrej Dyck, Bobby G Sumpter, and Sergei V Kalinin. Bridging microscopy with molecular dynamics and quantum simulations: an atomai based pipeline.npj Computational Materials, 8(1):74, 2022. 1, 3

    work page 2022

  7. [7]

    A hybrid frequency-spatial domain model for sparse image reconstruction in scanning transmission electron microscopy

    Bintao He, Fa Zhang, Huanshui Zhang, and Renmin Han. A hybrid frequency-spatial domain model for sparse image reconstruction in scanning transmission electron microscopy. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2682–2691, 2021. 3

    work page 2021

  8. [8]

    Determination of atomic positions from time resolved high resolution transmission electron mi- croscopy images.Ultramicroscopy, 186:139–145, 2018

    Zahra Hussaini, Pin Ann Lin, Bharath Natarajan, Wenhui Zhu, and Renu Sharma. Determination of atomic positions from time resolved high resolution transmission electron mi- croscopy images.Ultramicroscopy, 186:139–145, 2018. 2, 3

    work page 2018

Show all 40 references
  1. [9]

    Deep learning-based noise filtering toward millisecond order imaging by using scanning transmission electron microscopy.Scientific reports, 12(1): 13462, 2022

    Shiro Ihara, Hikaru Saito, Mizumo Yoshinaga, Lavakumar Avala, and Mitsuhiro Murayama. Deep learning-based noise filtering toward millisecond order imaging by using scanning transmission electron microscopy.Scientific reports, 12(1): 13462, 2022. 2, 3

    2022

  2. [10]

    Automated and autonomous experiments in electron and scanning probe mi- croscopy.ACS nano, 15(8):12604–12627, 2021

    Sergei V Kalinin, Maxim Ziatdinov, Jacob Hinkle, Stephen Jesse, Ayana Ghosh, Kyle P Kelley, Andrew R Lupini, Bobby G Sumpter, and Rama K Vasudevan. Automated and autonomous experiments in electron and scanning probe mi- croscopy.ACS nano, 15(8):12604–12627, 2021. 2

    2021

  3. [11]

    Chia-Hao Lee, Abid Khan, Bryan Clark, and Pinshane Y Huang. Using cyclegans to generate realistic stem im- ages for machine learning and atom-by-atom analysis on the million-atom scale.Microscopy and Microanalysis, 31 (Supplement 1):ozaf048–1074, 2025. 3

    2025

  4. [12]

    Fast prediction of irradiation- induced cascade defects using denoising diffusion proba- bilistic model.Nuclear Materials and Energy, 41:101805,

    Ruihao Liao, Ke Xu, Yifan Liu, Zibo Gao, Shuo Jin, Linyun Liang, and Guang-Hong Lu. Fast prediction of irradiation- induced cascade defects using denoising diffusion proba- bilistic model.Nuclear Materials and Energy, 41:101805,

  5. [13]

    Ruoqian Lin, Rui Zhang, Chunyang Wang, Xiao-Qing Yang, and Huolin L Xin. Temimagenet training library and atom- segnet deep-learning models for high-precision atom seg- mentation, localization, denoising, and deblurring of atomic- resolution images.Scientific reports, 11(1):53...

    2021

  6. [14]

    Deep convolutional neural networks to restore single-shot electron microscopy images.npj Computational Materials, 10(1):10, 2024

    I Lobato, T Friedrich, and S Van Aert. Deep convolutional neural networks to restore single-shot electron microscopy images.npj Computational Materials, 10(1):10, 2024. 2, 3

    2024

  7. [15]

    Emdiffuse: a diffusion-based deep learning method augmenting ultrastructural imaging and volume electron microscopy.bioRxiv, pages 2023–07,

    Chixiang Lu, Kai Chen, Heng Qiu, Xiaojun Chen, Gu Chen, Xiaojuan Qi, and Haibo Jiang. Emdiffuse: a diffusion-based deep learning method augmenting ultrastructural imaging and volume electron microscopy.bioRxiv, pages 2023–07,

    2023

  8. [16]

    Diffusion-based deep learn- ing method for augmenting ultrastructural imaging and vol- ume electron microscopy.Nature Communications, 15(1): 4677, 2024

    Chixiang Lu, Kai Chen, Heng Qiu, Xiaojun Chen, Gu Chen, Xiaojuan Qi, and Haibo Jiang. Diffusion-based deep learn- ing method for augmenting ultrastructural imaging and vol- ume electron microscopy.Nature Communications, 15(1): 4677, 2024. 3

    2024

  9. [17]

    Machine learn- ing approaches for improving atomic force microscopy in- strumentation and data analytics.Frontiers in Physics, 12: 1347648, 2024

    Nabila Masud, Jaydeep Rade, Md Hasibul Hasan Hasib, Adarsh Krishnamurthy, and Anwesha Sarkar. Machine learn- ing approaches for improving atomic force microscopy in- strumentation and data analytics.Frontiers in Physics, 12: 1347648, 2024. 1

    2024

  10. [18]

    Sdedit: Guided image synthesis and editing with stochastic differential equa- tions.arXiv preprint arXiv:2108.01073, 2021

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia- jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equa- tions.arXiv preprint arXiv:2108.01073, 2021. 3

    2021 arXiv

  11. [19]

    Null-text inversion for editing real im- ages using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real im- ages using guided diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3

    2023

  12. [20]

    Colin Ophus, Jim Ciston, and Chris T Nelson. Correcting nonlinear drift distortion of scanning probe and scanning transmission electron microscopies from image pairs with orthogonal scan directions.Ultramicroscopy, 162:1–9, 2016. 2

    2016

  13. [21]

    Diffusion probabilistic models enhance variational autoen- coder for crystal structure generative modeling.Scientific Reports, 14(1):1275, 2024

    Teerachote Pakornchote, Natthaphon Choomphon- Anomakhun, Sorrjit Arrerut, Chayanon Atthapak, Sakarn Khamkaeo, Thiparat Chotibut, and Thiti Bovornratanaraks. Diffusion probabilistic models enhance variational autoen- coder for crystal structure generative modeling.Scientific Re...

    2024

  14. [22]

    Deep learning for sub-˚angstr¨om resolution imaging in uncorrected scanning transmission electron mi- croscope.National Science Review, page nwaf235, 2025

    Zanlin Qiu, Yuan Meng, Junxian Li, Yanhui Hong, Ning Li, Xiaocang Han, Yu Liang, Wing Ni Cheng, Guolin Ke, Lin- feng Zhang, et al. Deep learning for sub-˚angstr¨om resolution imaging in uncorrected scanning transmission electron mi- croscope.National Science Review, page nwaf2...

    2025

  15. [23]

    Identification and correction of tempo- ral and spatial distortions in scanning transmission electron microscopy.Ultramicroscopy, 229:113337, 2021

    Kevin M Roccapriore, Nicole Creange, Maxim Ziatdinov, and Sergei V Kalinin. Identification and correction of tempo- ral and spatial distortions in scanning transmission electron microscopy.Ultramicroscopy, 229:113337, 2021. 3

    2021

  16. [24]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn 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. 1, 3

    2022

  17. [25]

    Unsupervised deep denoising for four- dimensional scanning transmission electron microscopy.npj Computational Materials, 10(1):243, 2024

    Alireza Sadri, Timothy C Petersen, Emmanuel WC Terzoudis-Lumsden, Bryan D Esser, Joanne Etheridge, and Scott D Findlay. Unsupervised deep denoising for four- dimensional scanning transmission electron microscopy.npj Computational Materials, 10(1):243, 2024. 1

    2024

  18. [26]

    Revolving scanning transmission electron microscopy: Correcting sample drift distortion without prior knowledge.Ultramicroscopy, 138: 28–35, 2014

    Xiahan Sang and James M LeBeau. Revolving scanning transmission electron microscopy: Correcting sample drift distortion without prior knowledge.Ultramicroscopy, 138: 28–35, 2014. 2

    2014

  19. [27]

    Precision con- trolled atomic resolution scanning transmission electron mi- croscopy using spiral scan pathways.Scientific Reports, 7 (1):43585, 2017

    Xiahan Sang, Andrew R Lupini, Jilai Ding, Sergei V Kalinin, Stephen Jesse, and Raymond R Unocic. Precision con- trolled atomic resolution scanning transmission electron mi- croscopy using spiral scan pathways.Scientific Reports, 7 (1):43585, 2017. 2

    2017

  20. [28]

    Flame diffuser: Wildfire image synthesis using mask guided diffusion

    Hao Wang, Sayed Pedram Haeri Boroujeni, Xiwen Chen, Ashish Bastola, Huayu Li, Wenhui Zhu, and Abolfazl Razi. Flame diffuser: Wildfire image synthesis using mask guided diffusion. In2024 IEEE International Conference on Big Data (BigData), pages 6171–6179. IEEE, 2024. 5

    2024

  21. [29]

    Diffusion prism: Enhancing diversity and morphology consistency in mask-to-image diffusion

    Hao Wang, Xiwen Chen, Ashish Bastola, Jiayou Qin, and Abolfazl Razi. Diffusion prism: Enhancing diversity and morphology consistency in mask-to-image diffusion. InPro- ceedings of the Winter Conference on Applications of Com- puter Vision (WACV) Workshops, pages 228–237, 2025. 3

    2025

  22. [30]

    Many-mobilenet: Multi- model augmentation for robust retinal disease classification

    Hao Wang, Wenhui Zhu, Xuanzhao Dong, Yanxi Chen, Xin Li, Peijie Qiu, Xiwen Chen, Vamsi Krishna Vasa, Yujian Xiong, Oana M Dumitrascu, et al. Many-mobilenet: Multi- model augmentation for robust retinal disease classification. InMICCAI Challenge on Ultra-Widefield Fundus Imagin...

    2025

  23. [31]

    Defect complexes in crsbr revealed through electron microscopy and deep learning

    Mads Weile, Sergii Grytsiuk, Aubrey Penn, Daniel G Chica, Xavier Roy, Kseniia Mosina, Zdenek Sofer, Jakob Schiøtz, Stig Helveg, Malte R ¨osner, et al. Defect complexes in crsbr revealed through electron microscopy and deep learning. Physical Review X, 15(2):021080, 2025. 1, 3

    2025

  24. [32]

    Rui Xiao, Yanzhu Zhang, and Mi Li. Automated high- throughput atomic force microscopy single-cell nanome- chanical assay enabled by deep learning-based optical image recognition.Nano Letters, 24(39):12323–12332, 2024. 1

    2024

  25. [33]

    Recovering a molecule’s 3d dynamics from liquid-phase electron microscopy movies

    Enze Ye, Yuhang Wang, Hong Zhang, Yiqin Gao, Huan Wang, and He Sun. Recovering a molecule’s 3d dynamics from liquid-phase electron microscopy movies. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 10767–10777, 2023. 3

    2023

  26. [34]

    Soda: Spectral or- thogonal decomposition adaptation for diffusion models

    Xinxi Zhang, Song Wen, Ligong Han, Felix Juefei-Xu, Akash Srivastava, Junzhou Huang, Vladimir Pavlovic, Hao Wang, Molei Tao, and Dimitris Metaxas. Soda: Spectral or- thogonal decomposition adaptation for diffusion models. In 2025 IEEE/CVF Winter Conference on Applications of C...

    2025

  27. [35]

    In situ transmission electron microscopy and artificial intelligence enabled data analytics for energy materials.Journal of Energy Chemistry, 68:454–493, 2022

    Hongkui Zheng, Xiner Lu, and Kai He. In situ transmission electron microscopy and artificial intelligence enabled data analytics for energy materials.Journal of Energy Chemistry, 68:454–493, 2022. 1, 3

    2022

  28. [36]

    Convolutional neural networks for evaluation of sequential beam damage of beam-sensitive solid electrolytes, 2024

    Hongkui Zheng, Xiwen Chen, Abolfazl Razi, and Kai He. Convolutional neural networks for evaluation of sequential beam damage of beam-sensitive solid electrolytes, 2024. 1, 2, 5, 6, 8

    2024

  29. [37]

    Gaussian differential assessment of sequen- tial stem radiation damage in beam-sensitive materials.Mi- croscopy and Microanalysis, 31(Supplement 1):ozaf048– 1065, 2025

    Hongkui Zheng, Hao Wang, Xiwen Chen, Abolfazl Razi, and Kai He. Gaussian differential assessment of sequen- tial stem radiation damage in beam-sensitive materials.Mi- croscopy and Microanalysis, 31(Supplement 1):ozaf048– 1065, 2025. 3

    2025

  30. [38]

    Denoising diffusion bridge models.arXiv preprint arXiv:2309.16948, 2023

    Linqi Zhou, Aaron Lou, Samar Khanna, and Stefano Er- mon. Denoising diffusion bridge models.arXiv preprint arXiv:2309.16948, 2023. 3

    2023 arXiv

  31. [39]

    nnmobilenet: Rethinking cnn for retinopathy research

    Wenhui Zhu, Peijie Qiu, Xiwen Chen, Xin Li, Natasha Lep- ore, Oana M Dumitrascu, and Yalin Wang. nnmobilenet: Rethinking cnn for retinopathy research. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2285–2294, 2024. 5

    2024

  32. [40]

    Atomai framework for deep learning anal- ysis of image and spectroscopy data in electron and scan- ning probe microscopy.Nature Machine Intelligence, 4(12): 1101–1112, 2022

    Maxim Ziatdinov, Ayana Ghosh, Chun Yin Wong, and Sergei V Kalinin. Atomai framework for deep learning anal- ysis of image and spectroscopy data in electron and scan- ning probe microscopy.Nature Machine Intelligence, 4(12): 1101–1112, 2022. 3

    2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.