Pith. sign in

REVIEW 3 major objections 8 minor 45 references

HyperFLINT: Hypernetwork-based Flow Estimation and Temporal Interpolation for Scientific Ensemble Visualization

T0 review · 3 major / 8 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read HyperFLINT claims that conditioning a flow-estimation and interpolation network on simulation parameters, through a hypernetwork that generates the network's weights, improves reconstruction accuracy across scientific ensemble members and…

desk verdict HyperFLINT is a competent, well-engineered extension of FLINT that conditions flow estimation on simulation parameters, but the parameter-space generalization claim is not yet backed by a split that rules out parameter memorization. read the letter →

arxiv 2412.04095 v2 pith:VR3KATQB submitted 2024-12-05 cs.CV cs.GRcs.LG

Hamid Gadirov , Qi Wu , David Bauer , Kwan-Liu Ma , Jos Roerdink , Steffen Frey This is my paper · ORCID
classification cs.CVcs.GRcs.LG
keywords hypernetworkflowestimationtemporalinterpolationscientificensemblevisualizationparameterspaceexploration3Dtime-varyingvolumessuper-resolution
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

HyperFLINT sets out to prove that explicitly feeding simulation parameters into a deep network improves the reconstruction of scientific ensemble data. Scientific runs of cosmological structure formation and white-dwarf mergers generate such large 3D+time volumes that only a subset of timesteps can be stored, and the paper's goal is to estimate the missing flow fields and density fields at those timesteps more accurately than methods that ignore the ensemble's parameters. The claimed mechanism is a hypernetwork that turns a short parameter vector into weights for the main network, and the reported results are higher density-interpolation accuracy and lower flow endpoint error than FLINT, STSR-INR, CoordNet, and linear interpolation at 3x, 5x, and 8x temporal rates, with an ablation attributing the gain to the hypernetwork itself. A further claim is that because the network is conditioned on simulation parameters, it can synthesize plausible density and flow estimates for parameter combinations that were never simulated, enabling parameter space exploration without new simulation runs.

What carries the argument

The load-bearing object is the HyperNet, a feedforward network with an MLP followed by Conv1D layers that takes the simulation parameter vector and outputs the convolution kernels for the main FLINT* network, a three-block convolutional and deconvolutional stack that refines flow estimates and warps the two boundary scalar fields toward the target time. The hypernetwork is the only pathway by which one ensemble member differs from another, so all member-specific behavior and all parameter-space generalization must be carried by the generated weights; that is also why the paper can treat weight similarity as a proxy for data similarity. Training is driven by a loss $L = L_{rec} + 0.2\,L_{flow}$ combining L1 reconstruction loss on the interpolated density field with a RAFT-style flow loss accumulated over the blocks with exponentially increasing weights ($\gamma = 0.8$).

What would settle it

Train or test HyperFLINT on an ensemble in which two members share identical simulation parameters but are generated from different initial conditions or random seeds; if the parameter-conditioned network produces near-identical outputs for the two members while their true flow and density fields differ measurably, the premise that parameters alone determine the data is refuted. A second test is to take the Nyx 24-member configuration from Section 7, hold out an entire parameter combination such as $\Omega_m = 0.155$, $h = 0.7$ during training, and quantitatively compare HyperFLINT's synthesized density and flow against the true simulated volumes for that combination; if the endpoint error and PSNR on such genuinely unseen members are no better than the parameter-agnostic FLINT baseline, the claimed parameter-space generalization does not materialize.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a single trained model can interpolate missing timesteps and estimate flow fields across an entire scientific ensemble, including members never seen during training, provided its weights are generated by a hypernetwork conditioned on that member's simulation parameters. The hypernetwork maps a short vector of physical parameters — total matter density, baryon density, and Hubble constant for Nyx; primary and secondary white-dwarf masses for Castro — to the convolutional kernels of the main FLINT* network, so the architecture stays fixed but the weights adapt to each ensemble member. In quantitative evaluation, this parameter-conditioned design outperforms the parameter-agnostic baselines FLINT, STSR-INR, and CoordNet in density interpolation while additionally producing flow estimates the scalar-only baselines cannot, and it lowers endpoint error markedly relative to FLINT on both datasets. The paper also reports that similarity between hypernetwork-generated weights tracks similarity between the underlying data fields at 96 percent triplet correlation, and that varying only the parameters while keeping input fields fixed yields outputs visually close to the ground truth for unsimulated parameter values.

Load-bearing premise

The load-bearing premise is that the few simulation parameters fed to the hypernetwork — $\Omega_m$, $\Omega_b$, and $h$ for Nyx, and $M_P$ and $M_S$ for Castro — fully determine how an ensemble member's data look, so conditioning on those numbers alone transfers across members and to never-simulated parameter values. If other unobserved factors such as initial conditions or solver randomness also shape the data, the claimed generalization would rest on a hidden correlation rather than on the parameters themselves.

Editorial extensions

If this is right

  • Simulation data can be stored more sparsely: density and flow fields at 3x, 5x, or 8x-omitted timesteps can be reconstructed with higher fidelity than parameter-agnostic methods, directly reducing storage and I/O pressure on large runs.
  • Researchers can probe simulation outcomes for parameter combinations never run, approximating the density and flow fields a new member would produce, which is the paper's claimed parameter space exploration.
  • Hypernetwork-generated weight similarities can serve as a cheap proxy for data similarity across ensemble members, supporting member comparison and parameter-sampling decisions without materializing full volumes.
  • The same hypernetwork-plus-main-network design should transfer to other spatio-temporal ensembles without pre-training, fine-tuning, or domain-specific assumptions, since the method is not tied to the physics of Nyx or Castro.

Reading between the lines

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

  • The paper's implicit claim that parameters alone determine member behavior is only tested qualitatively; a direct experiment with members that share parameters but differ in initial conditions or random seeds would show whether the model truly generalizes across the parameter manifold.
  • If the 96 percent weight-to-data similarity correlation holds beyond astrophysics, hypernetwork weights could be reused as a general ordering metric for ensemble members in other simulation domains, but that reuse is an extrapolation beyond the two datasets tested here.
  • A quantitative held-out test over entire parameter combinations, scoring EPE and PSNR against the true simulated volumes for the unseen Nyx members shown in Section 7, would convert the paper's visual evidence of parameter-space generalization into a measurable claim.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 8 minor

Summary. HyperFLINT extends the FLINT method with a hypernetwork that takes simulation parameters (e.g., Ω_m, Ω_b, h for Nyx; M_P, M_S for Castro) as input and generates the convolutional weights of a simplified FLINT* network. The network is trained end-to-end with a reconstruction loss and an exponentially weighted flow loss to perform temporal interpolation of scalar fields and estimation of flow fields. The paper reports quantitative comparisons against FLINT, STSR-INR, CoordNet, and Linear interpolation on the Nyx and Castro 3D+time ensembles at 3×, 5×, and 8× interpolation factors, showing higher PSNR and lower EPE for HyperFLINT. An ablation removes the hypernetwork, the flow loss, and the reconstruction loss, indicating that each component contributes to the reported performance. The paper also presents qualitative parameter-space exploration experiments, including a weight-similarity analysis and parameter-driven synthesis for unseen parameter settings.

Significance. The idea of conditioning a flow/interpolation network on ensemble simulation parameters via a hypernetwork is well motivated and of clear practical interest for scientific visualization, where parameter-space exploration and reconstruction of missing timesteps are important tasks. The manuscript gives a clear description of the architecture, losses, and training setup, and the ablation in Table 3 plausibly isolates a contribution from the hypernetwork (Nyx 5× EPE drops from 0.0357 without hyper to 0.0238 with hyper). The inference-time comparison is also a useful practical result. However, the central claim that the hypernetwork enables genuine generalization to unseen parameters rests on a single qualitative experiment, and the train/test split is not documented as parameter-disjoint. As such, the evidence is incomplete, and the strengths of the architecture are not yet matched by the strength of the evaluation.

major comments (3)
  1. [Section 4.2 (Datasets and Evaluation) and Tables 1–3] The split is described only as 'different ensemble members' for training, validation, and testing; the paper never states whether the test members' parameter combinations are absent from the training set. Since the hypernetwork receives exactly the parameter vector as input, the improvement of HyperFLINT over 'HyperFLINT w/o hyper' in Table 3 (e.g., Nyx 5×: EPE 0.0238 vs 0.0357, PSNR 52.70 vs 50.89) could reflect memorized parameter-to-member associations rather than adaptation to genuinely new parameter regimes. The authors should specify the parameter values of every member in each split, and, more importantly, perform an experiment with a parameter-disjoint train/test split (e.g., hold out entire parameter combinations) and report the same quantitative metrics.
  2. [Section 7.2 (Parameter-Driven Data Synthesis)] This is the only experiment aimed at demonstrating generalization to unseen parameters, but it is qualitative only: no PSNR or EPE values are reported for the held-out parameter settings. Moreover, the protocol starts from the same input volumes D_s and D_u and changes only the parameters fed to the HyperNet, then visually compares the output to the GT volume of the ensemble member with the target parameters. This comparison is meaningful only if all ensemble members share the same initial conditions; the paper does not state this, and if the members differ in initial conditions, even a perfect parameter-conditioned model could not reproduce the target member's GT. The authors should clarify the initial-condition setup and provide quantitative errors for the transferred-parameter outputs (e.g., PSNR/EPE for each target parameter in Fig. 7).
  3. [Tables 1, 2, 3, and 5 (Quantitative Evaluation)] All quantitative numbers are single runs, with no variance estimates or significance tests. The abstract's claim of 'significantly improved' performance is not supported by error bars or repeated-seed statistics; for example, the Nyx 3× PSNR difference between HyperFLINT and FLINT is only 0.15 dB (53.32 vs 53.17). At a minimum, the authors should report the mean and standard deviation over multiple training runs (or other variance measures) for the main comparisons and the ablations, since the small PSNR margins may lie within run-to-run variability.
minor comments (8)
  1. [Section 4.1/6.3] The hyperparameter search in Section 6.3 is extensive, but the paper does not state whether the best configuration was selected on the validation set and then evaluated on the test set, or whether the numbers in Tables 1–3 come from the same run used for hyperparameter selection. Please clarify the selection protocol.
  2. [Section 4.2] Please state the exact number of members used for training, validation, and testing (e.g., '18/6/12' for Nyx and '6/2/4' for Castro), and list the parameter values of each member so that parameter-disjointness is explicit.
  3. [Section 3.3 (Flow Estimation and Scalar Field Interpolation)] The notation in Eq. (1b), \hat{F}^{i+1}_t = \hat{F}^i_{t→u}, is confusing; the superscript on \hat{F}^i_t is never defined separately from \hat{F}^i_{t→u}, and the final flow \hat{F}_t = \hat{F}^{N-1}_t is ambiguous. Please rewrite this equation and define all indices.
  4. [Section 3.4 (Loss Function)] The text says the flow loss uses exponentially increasing weights from RAFT, and Equation (4) writes \gamma^{N-i} with \gamma=0.8. With N=3 this gives weights 0.64, 0.8, 1.0 for i=1,2,3, which are increasing, but 'exponentially increasing' is not immediately clear from the formula as written; consider adding a short explanation.
  5. [Section 4.2 / Appendix A] The PSNR and EPE formulas in the Appendix are given for the 3D volume, but the normalization of the flow field to [-1,1] is mentioned in Section 4.1; please clarify whether EPE is computed in normalized or physical units, since the absolute EPE values would depend on this choice.
  6. [Introduction] The abstract and introduction repeat the claim that HyperFLINT works 'without requiring domain-specific assumptions, pre-training, or fine-tuning on simplified datasets'; this is stronger than what is demonstrated, since the method does assume that the listed simulation parameters are the relevant conditioning variables. Please soften or qualify this claim.
  7. [General] The paper promises 'The HyperFLINT code will be made available publicly' but no repository link is provided; please include the URL in the final version.
  8. [Section 3.3] There is a typo in the first sentence: 'HypetNet' should be 'HyperNet'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: HyperFLINT's claimed improvements are established by held-out quantitative comparisons, not by premises that contain the target results.

full rationale

HyperFLINT's central claims are empirical: Tables 1 and 2 report PSNR and EPE against held-out ground truth for Nyx and Castro, compared with external baselines (STSR-INR, CoordNet, Linear) and the prior FLINT model. The method defines a hypernetwork that maps simulation parameters to weights of a main network, but the loss functions (Eqs. 2-4) minimize L1 distances to GT density and flow; no fitted quantity is renamed as a prediction. The ablation in Table 3 isolates the hypernetwork's contribution by removing it from the same architecture, so the reported gain is not an artifact of definition. The only self-citations are to the authors' FLINT paper [GRF24], which is used as a baseline and as architectural inspiration; FLINT's own performance is not assumed as evidence for HyperFLINT's claims, and the comparison is against external benchmarks as well. The parameter-space exploration (Sec. 7) is qualitative and does not report numerical PSNR/EPE for unseen parameters, which limits the strength of generalization claims, but this is a completeness and validation concern rather than circular reasoning. The paper itself acknowledges that novel unseen features cannot be generated and that errors grow for distant parameters, which is an honest limitation statement. The triplet correlation in Sec. 7.1 is an empirical post-hoc analysis; the hypernetwork was not trained with a triplet objective, so the correlation is not forced by construction. No equation reduces to its own input, and no self-citation chain is load-bearing. Hence no circularity is exhibited, and the score is 0.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the sufficiency of simulation parameters as conditioning variables, on the representativeness of the data splits, and on the warping model. These are domain assumptions rather than derived facts. The model's hyperparameters are fitted to the data via search, and no new physical entities are introduced.

free parameters (6)
  • lambda_flow (loss balance weight) = 0.2
    Balance between reconstruction and flow loss; determined experimentally in Sec 6.3.
  • gamma (flow loss exponential weight) = 0.8
    Exponential weighting across network blocks, adopted from RAFT and validated via hyperparameter search (Sec 6.3).
  • N (number of convolutional blocks) = 3
    Chosen via hyperparameter search; 3 blocks outperformed 2, 4, and 5 in Table 4.
  • Channel widths (128, 96, 64) = 128, 96, 64
    Architecture capacity selected through hyperparameter search (Sec 6.3).
  • Maximum training time window size = 12
    Maximum gap between sampled timesteps s and u; confirmed through hyperparameter search (Sec 4.1).
  • Learning rate schedule = 1e-4 to 1e-5 (cosine)
    Set experimentally, as stated in Sec 4.1.
assumptions (3)
  • domain assumption The simulation parameters provided (Omega_m, Omega_b, h for Nyx; M_P, M_S for Castro) are sufficient to characterize the variation between ensemble members.
    The hypernetwork conditions only on these parameters; if other factors such as initial conditions or stochasticity dominate, generalization across the ensemble would fail. Stated implicitly in Sec 4.2 and used in Sec 3.2 and Sec 7.2.
  • domain assumption Training, validation, and test splits by ensemble member are representative, so performance on test members reflects generalization to unseen parameter combinations.
    The paper uses different ensemble members for each split (Sec 4.1 and 4.2) but does not list the parameter values of each member, so it is not verified that test parameters lie within or at the boundary of the training parameter range.
  • domain assumption Backward warping with trilinear interpolation, guided by the estimated flow fields, can represent the true temporal correspondence between the input scalar fields.
    The warping operator in Sec 3.3 assumes apparent motion between D_s and D_u is captured by the flow; this is the standard optical-flow assumption and is not independently validated on these datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HyperFLINT: Hypernetwork-based Flow Estimation and Temporal Interpolation for Scientific Ensemble Visualization." pith.science (2026). https://pith.science/paper/VR3KATQB

@misc{pith2026241204095,
  author       = {Pith},
  title        = {Pith review of: HyperFLINT: Hypernetwork-based Flow Estimation and Temporal Interpolation for Scientific Ensemble Visualization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VR3KATQB}},
  note         = {Machine review of arXiv:2412.04095}
}
read the original abstract

We present HyperFLINT (Hypernetwork-based FLow estimation and temporal INTerpolation), a novel deep learning-based approach for estimating flow fields, temporally interpolating scalar fields, and facilitating parameter space exploration in spatio-temporal scientific ensemble data. This work addresses the critical need to explicitly incorporate ensemble parameters into the learning process, as traditional methods often neglect these, limiting their ability to adapt to diverse simulation settings and provide meaningful insights into the data dynamics. HyperFLINT introduces a hypernetwork to account for simulation parameters, enabling it to generate accurate interpolations and flow fields for each timestep by dynamically adapting to varying conditions, thereby outperforming existing parameter-agnostic approaches. The architecture features modular neural blocks with convolutional and deconvolutional layers, supported by a hypernetwork that generates weights for the main network, allowing the model to better capture intricate simulation dynamics. A series of experiments demonstrates HyperFLINT's significantly improved performance in flow field estimation and temporal interpolation, as well as its potential in enabling parameter space exploration, offering valuable insights into complex scientific ensembles.

Figures

Figures reproduced from arXiv: 2412.04095 by the authors.

Figure 1
Figure 1. Overview of HyperFLINT pipeline during inference. The FLINT* deep neural network, whose weights are generated by the HyperNet, performs flow field estimation Fˆ t and temporal (scalar) field interpolation Dˆ t , where s < t < u, by utilizing the available densities Ds and Du from the previous and following timesteps, and their simulation parameters. each parameter set. This capacity is particularly valuable for unde… view at source ↗
Figure 2
Figure 2. HyperFLINT network architecture and pipeline during training: Given the input fields Ds and Du, and their simulation parameters, HyperFLINT predicts the Dˆ t scalar field and Fˆi t flow fields used in the loss function for optimizing network parameters. The HyperFLINT model consists of two key components: the HyperNet and the main network, FLINT*. The HyperNet, depicted within the red box, generates weights for the … view at source ↗
Figure 3
Figure 3. Illustration of the 3D backward warping W ← : (scalar) fields Ds and Du are reversely mapped according to the flow fields Fˆ t→s and Fˆ t→u. The fields Dˆ t←s and Dˆ t←u are then reconstructed using trilinear interpolation considering the values at the coordinates shown with green dots (for the visible front surface of the cube). the network operates solely with its learned parameters. The hyper￾network plays a cruc… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Nyx and Castro: HyperFLINT flow field estimation and temporal density interpolation, 5×. From top to bottom, the rows show GT density, HyperFLINT interpolated density, FLINT interpolation, STSR-INR interpolation, GT flow, HyperFLINT flow estimation, and FLINT flow esti…
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Nyx simulation parameter space exploration and transfer functions for density and flow field components. From top to bottom, the rows show GT density, HyperFLINT interpolated density, difference between GT and HyperFLINT density, GT flow, HyperFLINT flow estimation, an…
Figure 8
Figure 8. Figure 8: Nyx ensemble: transfer function for density and x, y, z components of the flow field. (a) TF for density field. (b) TF for flow field in x direction. (c) TF for flow field in y direction. (d) TF for flow field in z direction [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Castro ensemble: transfer function for density and x, y, z components of the flow field. 3D Flow Estimation and Density Interpolation Results This subsection presents the complete results for 3D flow estimation and temporal interpolation tasks on the Nyx ( [PITH_FULL_…
Figure 10
Figure 10. Figure 10: Nyx: HyperFLINT flow field estimation and temporal density interpolation, 5×. From top to bottom, the rows show GT density, HyperFLINT interpolated density, FLINT interpolation, STSR-INR interpolation, GT flow, HyperFLINT flow estimation, and FLINT flow estimation. 3D…
Figure 11
Figure 11. Figure 11: Castro: HyperFLINT flow field estimation and temporal density interpolation, 5×. From top to bottom, the rows show GT density, HyperFLINT interpolated density, FLINT interpolation, STSR-INR interpolation, GT flow, HyperFLINT flow estimation, and FLINT flow estimation.…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 37 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry.original add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 i...

  3. [3]

    S., Beckner V

    Almgren A. S., Beckner V. E., Bell J. B., Day M. S., Howell L. H., Joggerst C. C., Lijewski M. J., Nonaka A., Singer M., Zingale M. : CASTRO: A New Compressible Astrophysical Solver. I. Hydrodynamics and Self-gravity . The Astrophysical Journal 715, 2 (June 2010), 1221--1238. http://arxiv.org/abs/1005.0114 arXiv:1005.0114 , https://doi.org/10.1088/0004-63...

    arXiv 2010

  4. [4]

    M., Weston N

    Brock A., Lim T., Ritchie J. M., Weston N. : Smash: one-shot model architecture search through hypernetworks. arXiv preprint arXiv:1708.05344 (2017)

    arXiv 2017

  5. [5]

    D., Ahrens J., Bauer A

    Childs H., Ahern S. D., Ahrens J., Bauer A. C., Bennett J., Bethel E. W., Bremer P.-T., Brugger E., Cottam J., Dorier M., et al. : A terminology for in situ visualization and analysis systems. The International Journal of High Performance Computing Applications 34, 6 (2020), 676--691

    work page 2020

  6. [6]

    : In situ visualization for computational science

    Childs H., Bennett J., Garth C., Hentschel B. : In situ visualization for computational science. IEEE Comput. Graph. Appl. 39, 6 (2019), 76--85

    work page 2019

  7. [7]

    C., He K., Tang X

    Dong C., Loy C. C., He K., Tang X. : Image super-resolution using deep convolutional networks. IEEE Trans. Pattern Anal. Mach. Intell. 38, 2 (2015), 295--307

    work page 2015

  8. [8]

    : Flow-based temporal selection for interactive volume visualization

    Frey S., Ertl T. : Flow-based temporal selection for interactive volume visualization. Comput. Graph. Forum 36, 8 (2017), 153--165. https://doi.org/https://doi.org/10.1111/cgf.13070 doi:https://doi.org/10.1111/cgf.13070

    work page doi:10.1111/cgf.13070 2017

Show all 45 references
  1. [9]

    : FLINT : Learning-based flow estimation and temporal interpolation for scientific ensemble visualization

    Gadirov H., Roerdink J., Frey S. : FLINT : Learning-based flow estimation and temporal interpolation for scientific ensemble visualization. arXiv preprint arXiv:2409.19178 (2024)

    2024 arXiv

  2. [10]

    Ha D., Dai A., Le Q. V. : Hypernetworks. arXiv preprint arXiv:1609.09106 (2016)

    2016 arXiv

  3. [11]

    : TSR-TVD : Temporal super-resolution for time-varying data analysis and visualization

    Han J., Wang C. : TSR-TVD : Temporal super-resolution for time-varying data analysis and visualization. IEEE Trans. Vis. Comput. Graph. 26, 1 (2019), 205--215

    2019

  4. [12]

    : SSR-TVD : Spatial super-resolution for time-varying data analysis and visualization

    Han J., Wang C. : SSR-TVD : Spatial super-resolution for time-varying data analysis and visualization. IEEE Transactions on Visualization and Computer Graphics 28, 6 (2020), 2445--2456

    2020

  5. [13]

    : CoordNet: Data generation and visualization generation for time-varying volumes via a coordinate-based neural network

    Han J., Wang C. : CoordNet: Data generation and visualization generation for time-varying volumes via a coordinate-based neural network. IEEE Trans. Vis. Comput. Graph. (2022)

    2022

  6. [14]

    S., Peterka T

    He W., Wang J., Guo H., Wang K.-C., Shen H.-W., Raj M., Nashed Y. S., Peterka T. : InSituNet : Deep image synthesis for parameter space exploration of ensemble simulations. IEEE Trans. Vis. Comput. Graph. 26, 1 (2019), 23--33

    2019

  7. [15]

    Z., Wang C

    Han J., Zheng H., Chen D. Z., Wang C. : STNet : An end-to-end generative framework for synthesizing spatiotemporal super-resolution volumes. IEEE Trans. Vis. Comput. Graph. 28, 1 (2021), 270--280

    2021

  8. [16]

    : Delving deep into rectifiers: Surpassing human-level performance on imagenet classification

    He K., Zhang X., Ren S., Sun J. : Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proc. IEEE Int. Conf. Comput. Vis. (ICCV) (2015), pp. 1026--1034

    2015

  9. [17]

    : Ffeinr: Flow feature-enhanced implicit neural representation for spatio-temporal super-resolution

    Jiao C., Bi C., Yang L. : Ffeinr: Flow feature-enhanced implicit neural representation for spatio-temporal super-resolution. arXiv preprint arXiv:2308.12508 (2023)

    2023 arXiv

  10. [18]

    : Super slomo: High quality estimation of multiple intermediate frames for video interpolation

    Jiang H., Sun D., Jampani V., Yang M.-H., Learned-Miller E., Kautz J. : Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) (2018), pp. 9000--9008

    2018

  11. [19]

    S., Townsend P., Chen b

    Jänicke H., Weidner T., Chung D., Laramee R. S., Townsend P., Chen b. M. : Visual reconstructability as a quality metric for flow visualization. Comput. Graph. Forum 30, 3 (2011), 781--790. URL: http://onlinelibrary.wiley.com/doi/abs/10.1111/j.1467-8659.2011.01927.x, https://d...

    2011

  12. [20]

    : Visual analysis of the temporal evolution of ensemble forecast sensitivities

    Kumpf A., Rautenhaus M., Riemer M., Westermann R. : Visual analysis of the temporal evolution of ensemble forecast sensitivities. IEEE Trans. Vis. Comput. Graph. 25, 1 (2018), 98--108

    2018

  13. [21]

    P., Sch \"u tz L., Gunther S., Hufnagel L., Lemke S., Leitte H

    Kappe C. P., Sch \"u tz L., Gunther S., Hufnagel L., Lemke S., Leitte H. : Reconstruction and visualization of coordinated 3d cell migration based on optical flow. IEEE Trans. Vis. Comput. Graph. 22, 1 (2015), 995--1004

    2015

  14. [22]

    : SwinIR : Image restoration using swin transformer

    Liang J., Cao J., Sun G., Zhang K., Van Gool L., Timofte R. : SwinIR : Image restoration using swin transformer. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) (2021), pp. 1833--1844

    2021

  15. [23]

    : Fixing weight decay regularization in adam

    Loshchilov I., Hutter F., et al. : Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101 5 (2017)

    2017 arXiv

  16. [24]

    A., Berger M

    Lu Y., Jiang K., Levine J. A., Berger M. : Compressive neural representations of volumetric scalar fields. In Computer Graphics Forum (2021), vol. 40, Wiley Online Library, pp. 135--146

    2021

  17. [25]

    : Photo-realistic single image super-resolution using a generative adversarial network

    Ledig C., Theis L., Husz \'a r F., Caballero J., Cunningham A., Acosta A., Aitken A., Tejani A., Totz J., Wang Z., et al. : Photo-realistic single image super-resolution using a generative adversarial network. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) (2017), p...

    2017

  18. [26]

    : Feedback network for image super-resolution

    Li Z., Yang J., Liu Z., Yang X., Jeon G., Wu W. : Feedback network for image super-resolution. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) (2019), pp. 3867--3876

    2019

  19. [27]

    : A sparse-to-dense method for 3D optical flow estimation in 3D light-microscopy image sequences

    Manandhar S., Bouthemy P., Welf E., Roudot P., Kervrann C. : A sparse-to-dense method for 3D optical flow estimation in 3D light-microscopy image sequences. In 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI) (2018), IEEE, pp. 952--956

    2018

  20. [28]

    : Filling the Void : Deep learning-based reconstruction of sampled spatiotemporal scientific simulation data

    Mishra A., Hazarika S., Biswas A., Bryan C. : Filling the Void : Deep learning-based reconstruction of sampled spatiotemporal scientific simulation data. arXiv preprint arXiv:2205.12868 (2022)

    2022 arXiv

  21. [29]

    P., Tancik M., Barron J

    Mildenhall B., Srinivasan P. P., Tancik M., Barron J. T., Ramamoorthi R., Ng R. : NeRF : Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65, 1 (2021), 99--106

    2021

  22. [30]

    : Phase-based frame interpolation for video

    Meyer S., Wang O., Zimmer H., Grosse M., Sorkine-Hornung A. : Phase-based frame interpolation for video. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) (2015), pp. 1410--1418

    2015

  23. [31]

    : Video frame interpolation via adaptive separable convolution

    Niklaus S., Mai L., Liu F. : Video frame interpolation via adaptive separable convolution. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) (2017), pp. 261--270

    2017

  24. [32]

    J., Florence P., Straub J., Newcombe R., Lovegrove S

    Park J. J., Florence P., Straub J., Newcombe R., Lovegrove S. : DeepSDF : Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2019), pp. 165--174

    2019

  25. [33]

    : Hyper MAML : few-shot adaptation of deep models with hypernetworks

    Przewie \'z likowski M., Przybysz P., Tabor J., Zieba M., Spurek P. : Hyper MAML : few-shot adaptation of deep models with hypernetworks. arXiv preprint arXiv:2205.15745 (2022)

    2022 arXiv

  26. [34]

    : High-resolution image synthesis with latent diffusion models

    Rombach R., Blattmann A., Lorenz D., Esser P., Ommer B. : High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022), pp. 10684--10695

    2022

  27. [35]

    : Integration-free learning of flow maps

    Sahoo S., Berger M. : Integration-free learning of flow maps. arXiv preprint arXiv:2211.03192 (2022)

    2022 arXiv

  28. [36]

    P., Bishop R., Rueckert D., Wang Z

    Shi W., Caballero J., Husz \'a r F., Totz J., Aitken A. P., Bishop R., Rueckert D., Wang Z. : Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) (2016), pp. 1874--1883

    2016

  29. [37]

    : Adversarial generation of continuous images

    Skorokhodov I., Ignatyev S., Elhoseiny M. : Adversarial generation of continuous images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021), pp. 10753--10764

    2021

  30. [38]

    : Learning a distance metric from relative comparisons

    Schultz M., Joachims T. : Learning a distance metric from relative comparisons. Advances in neural information processing systems 16 (2003)

    2003

  31. [39]

    : Nyx: A massively parallel amr code for computational cosmology

    Sexton J., Lukic Z., Almgren A., Daley C., Friesen B., Myers A., Zhang W. : Nyx: A massively parallel amr code for computational cosmology. The Journal of Open Source Software 6, 63 (2021), 3068

    2021

  32. [40]

    : Scene representation networks: Continuous 3d-structure-aware neural scene representations

    Sitzmann V., Zollh \"o fer M., Wetzstein G. : Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems 32 (2019)

    2019

  33. [41]

    : RAFT : Recurrent all-pairs field transforms for optical flow

    Teed Z., Deng J. : RAFT : Recurrent all-pairs field transforms for optical flow. In Computer Vision--ECCV 2020: 16th European Conference, Proceedings, Part II 16 (2020), Springer, pp. 402--419

    2020

  34. [42]

    : Local prediction models for spatiotemporal volume visualization

    Tkachev G., Frey S., Ertl T. : Local prediction models for spatiotemporal volume visualization. IEEE Trans. Vis. Comput. Graph. 27, 7 (2021-07), 3091--3108. https://doi.org/10.1109/TVCG.2019.2961893 doi:10.1109/TVCG.2019.2961893

    2021

  35. [43]

    : STSR-INR : Spatiotemporal super-resolution for multivariate time-varying volumetric data via implicit neural representation

    Tang K., Wang C. : STSR-INR : Spatiotemporal super-resolution for multivariate time-varying volumetric data via implicit neural representation. Computers & Graphics 119 (2024), 103874

    2024

  36. [44]

    : HyperINR : A fast and predictive hypernetwork for implicit neural representations via knowledge distillation

    Wu Q., Bauer D., Chen Y., Ma K.-L. : HyperINR : A fast and predictive hypernetwork for implicit neural representations via knowledge distillation. arXiv preprint arXiv:2304.04188 (2023)

    2023 arXiv

  37. [45]

    : Fast neural representations for direct volume rendering

    Weiss S., Herm \"u ller P., Westermann R. : Fast neural representations for direct volume rendering. In Computer Graphics Forum (2022), vol. 41, Wiley Online Library, pp. 196--211

    2022

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.