Pith. sign in

REVIEW 4 major objections 5 minor 20 references

Tropical Cyclone Forecasting via Latent Rectified Flow using Satellite Imagery and Atmospheric Fields

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a single latent rectified-flow model can jointly forecast cyclone satellite imagery and four atmospheric fields for nine hours, and that this one-pass model beats a reproduced fifty-step cascaded-diffusion baseline…

desk verdict Genuine modeling contribution, but the headline margins over the reproduced baseline are unverifiable as written because the baseline's five-channel metrics and track error are undefined. read the letter →

arxiv 2608.08354 v1 pith:6JH7UTQK submitted 2026-08-08 cs.CV

Meheru Zannat , Sk. Md. Masudul Ahsan This is my paper · ORCID
classification cs.CV
keywords tropicalcycloneforecastingrectifiedflowmatchinglatentgenerativemodelsERA5reanalysisGRIDSAT-B1steeringrewardfine-tuning
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

This paper tries to establish that a single-pass latent rectified-flow model can replace the slow, modality-specific diffusion pipelines used for tropical cyclone forecasting. It jointly predicts five channels—infrared brightness temperature plus four ERA5 fields (U-wind, V-wind, temperature, surface pressure)—for three future frames at three-hour cadence from three past frames, in one forward pass. On held-out 2022 storms, the authors report higher PSNR than a reproduced cascaded-diffusion baseline at every lead time (+0.84 dB at +9 h), a track error of 62.4 km at +9 h (15% lower), and roughly 30x faster sampling. They also show that the storm track can be read out from the predicted winds through a steering-flow formula, and that fine-tuning against a differentiable track error cuts track error by another 8-11%. A sympathetic reader would care because the speed and the physical coupling of track to atmosphere are exactly what current deep-learning cyclone systems lack.

What carries the argument

The load-bearing object is the rectified-flow velocity field $v_\theta(Z_\tau, \tau, c)$ trained to match the constant velocity $Z_1-Z_0$ of the straight-line interpolant between noise and the target latent block. Because trajectories are nearly straight, Euler sampling with $N_{\text{steps}}=1$ to 4 suffices, which is what produces the 27-56 ms forecast times. Around this sit a five-channel VAE (unbounded decoder to preserve pressure and wind z-score range), a factorized temporal-attention module that lets the +6 h and +9 h frames attend to earlier output frames at the bottleneck, and the steering-flow readout that converts area-mean 850 hPa winds into a storm displacement.

What would settle it

Run the original Nath et al. cascaded-diffusion code or checkpoint on the identical 2022 SETCD test windows; if the original model's PSNR and track error match or beat the reproduced baseline numbers, the claimed +0.84 dB and 15% track-error margins do not reflect the actual state of the art.

Watch

Extended reading notes

Core claim

The central claim is that a latent rectified-flow model trained with a straight-line interpolant $Z_\tau=(1-\tau)Z_0+\tau Z_1$ can generate the next three $5\times256\times256$ cyclone frames from three past frames in a single Euler pass, and that this is enough to beat a fifty-step cascaded-diffusion baseline on the same benchmark. The authors report 16.35 dB PSNR and 0.759 SSIM on held-out 2022 storms across all five channels; at +9 h the model reaches 15.48 dB against the baseline's 14.64 dB, and the steering-flow track error falls from 73.4 km to 62.4 km. The trajectory is not produced by a regression head; it is computed from the area-mean generated winds via $\Delta\phi_k = \bar{v}_k\Delta t / 111{,}000$, so the track is a direct physical probe of the generated atmosphere. A VAE-ceiling diagnostic puts the generative stage, not the autoencoder, as the bottleneck, and a reward fine-tuning study reports a further 8-11% track-error reduction. The authors also document two negative results: policy-gradient fine-tuning (GRPO) is flat, and a CRPS reward cannot fix the under-dispersed ensembles that a near-deterministic flow sampler produces.

Load-bearing premise

The comparison to the baseline assumes the authors' re-implementation of the cascaded-diffusion model is faithful and well-tuned; if that reproduction underperforms the original, the reported PSNR and track-error advantages are overstated.

Editorial extensions

If this is right

  • If the claim holds, real-time ensemble forecasting of cyclone structure becomes feasible on a single commodity GPU: four Euler steps cost 56 ms per member, versus 1673 ms for the 50-step baseline.
  • Because the track is read from generated winds rather than a separate head, improvements to the atmospheric fields translate directly into better tracks; the paper's reward fine-tuning demonstrates this coupling by cutting +9 h track error 8-11%.
  • Joint generation of imagery and atmospheric fields means cross-channel consistency is learned: cloud structure, winds, temperature, and pressure are produced by one latent, not stitched from separate models.
  • The ablation results imply that multi-frame conditioning, not just a bigger generator, is what preserves long-lead dynamics: autoregressive rollout loses 1.38 dB at +9 h, mostly in the temperature channel.
  • The sampler budget sweep implies that a single Euler step (27 ms) is within 0.25 dB of the four-step optimum, so the practical operating point can be chosen by available compute.

Reading between the lines

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

  • Inference: replacing the surface-pressure channel with mean-sea-level pressure would likely remove the cross-dataset blow-up on terrain-crossing storms; the paper identifies the failure mode but does not test this fix.
  • Inference: the structural CRPS under-dispersion implies that for probabilistic applications the model would need stochastic inference sampling or a CRPS-trained distillation; the speed advantage is bought at the cost of ensemble spread.
  • Inference: the steering-flow readout could be made robust to rapidly recurring storms by using a deep-layer mean wind (for example 200-850 hPa) instead of the central 64x64 area mean; the paper notes the failure mode but leaves the fix open.
  • Inference: the same joint image-plus-atmosphere formulation could be applied to other advected hazards where a physics-derived trajectory readout exists, such as wildfire smoke plumes or atmospheric rivers.
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 / 5 minor

Summary. The paper proposes a two-stage latent rectified-flow model for tropical cyclone forecasting. A five-channel variational autoencoder compresses GRIDSAT-B1 infrared imagery plus ERA5 U-wind, V-wind, temperature, and surface pressure into a shared latent space; a rectified-flow UNet with factorized temporal attention generates the next three latent frames from three past frames plus best-track coordinates and timestamps; and a steering-flow readout computes storm track positions from the predicted winds, with optional DRaFT reward fine-tuning. The authors report that on held-out 2022 storms the model outperforms a reproduced cascaded-diffusion baseline on PSNR at every lead time, reduces track error by 15% at +9 h, and samples roughly 30x faster.

Significance. If the comparative claims hold, the paper would make a useful contribution: it is one of the first cyclone-forecasting models to jointly generate satellite imagery and atmospheric fields, it demonstrates that a near-one-step rectified-flow sampler can approach the quality of a fifty-step diffusion baseline, and it ties the track readout to the generated winds through a physically motivated steering-flow calculation rather than a separate regression head. The paper also reports several strengths worth credit: a VAE-ceiling diagnostic, a zero-shot cross-sensor evaluation, honest negative results (GRPO and CRPS-reward fine-tuning did not help), and a hardware budget small enough for single-GPU reproduction. The significance is currently conditional, however, because the headline margins rest on a baseline whose output modality and metrics are not clearly defined, and because the reported per-lead differences are not accompanied by uncertainty estimates.

major comments (4)
  1. [§4.3, Table 3, Table 1] The baseline comparison is not well defined. Table 1 and Section 2 classify Nath et al. [1] as producing satellite imagery only (all five design axes are marked ×), yet Table 3 reports for that baseline MAE/PSNR/SSIM/CRPS over 'the five output channels' and a track error in km, both of which require predicted U-wind, V-wind, temperature, and surface pressure. Section 4.3 describes only the cascade configuration (64→128→256, CFG 3.0, 50-step DDIM) and never states that the reimplementation was modified to predict five channels. If the baseline was extended to produce five channels, it is no longer the published Nath et al. model and the comparison needs a full description of the extension; if it was not extended, the non-IR metrics and the wind-derived track error in Table 3 are undefined. This is load-bearing because every headline superiority claim, including the +0.84 dB PSNR margin, the 15% track-error reduction, and the ~30× speedup, depends on the Table 3 numbers. The authors should clarify the baseline's architecture, retrain or reproduce the original imagery-only model where applicable, and make the reimplementation code or checkpoints available.
  2. [§3.2, §4.8] The missing-value handling is a potentially biasing preprocessing choice. Section 3.2 states that 'missing entries (NaN) in the GRIDSAT and ERA5 arrays are replaced with zeros,' after which all channels are z-score normalized with global statistics. A raw zero is not a neutral value: after z-scoring it becomes an arbitrary non-mean value whose effect depends on the per-channel mean and standard deviation. The paper itself shows in Section 4.8 that surface pressure over high terrain falls to roughly −29 standard deviations relative to the ocean-dominated training statistics, which indicates that extreme out-of-distribution values do occur. The manuscript should report the fraction of NaNs per channel, justify zero imputation relative to alternatives such as temporal interpolation or masking, and show a sensitivity analysis. Without this, the learned distribution and the reported metrics may be contaminated by an imputation artifact.
  3. [Table 3, Table 4] The headline per-lead metrics in Table 3 lack uncertainty estimates. Table 4 reports seed-to-seed standard deviations for the proposed model (e.g., 0.08–0.14 dB PSNR depending on step count), but Table 3 gives only point estimates for both the proposed model and the baseline. At +3 h the claimed PSNR margin over the baseline is +0.15 dB, which is comparable to the seed noise reported in Table 4; the track-error margins and the 'ahead at every lead' claim need confidence intervals or paired multi-seed results. The authors should report mean ± s.d. over at least three seeds for both methods, or otherwise establish that the margins are statistically significant.
  4. [§3.7, §4.6, Table 5] The steering-flow track readout is central to the claimed 15% track-error improvement, but the approximation used in Equations (6)–(7) — an area mean of the generated 850 hPa winds over the central 64×64 crop — is not validated as a proxy for the environmental steering flow. The paper itself notes in Section 4.6 that track error doubles on rapidly recurving storms, and the cross-dataset results in Table 5 show a much larger degradation (218 km at +9 h zero-shot), so the 17.5/41.0/62.4 km numbers may be sensitive to this choice. The authors should compare their steering estimate with a deep-layer-mean steering definition (e.g., 200–850 hPa) or with observed storm-motion residuals, and quantify how the track-error comparison changes under that alternative.
minor comments (5)
  1. [Abstract, §3.5, §4.7] The abstract and contribution list describe the model as 'single-pass' and state that a 'single Euler step' matches the baseline, while Section 4.4 says evaluation uses 50-step Euler sampling and Section 4.7 reports that quality peaks at four Euler steps. Please clarify which configuration produces the headline numbers in Table 3 and distinguish 'single forward pass' from 'single Euler step.'
  2. [Table 6] The caption of Table 6 says 'overall PSNR/SSIM aggregated across leads,' but the single-frame model is reported at 18.71 dB while the multi-frame models are at roughly 16.2–16.35 dB; the footnote explains that the single-frame problem is easier because it only forecasts +3 h. This is confusing in a table that purports to aggregate across leads. Please either restrict Table 6 to the same lead set or change the caption to state that the single-frame row is evaluated at +3 h only.
  3. [§3.3] The scale factor s = 1/σ_z estimated on the validation split is introduced, but the text does not specify where it is applied (to the latent codes before flow training, to the decoder output, or to the loss). Please define the exact use of s in the VAE and flow training equations.
  4. [Table 3] Table 3 states that MAE is in z-scored units but does not state the units of CRPS. Please add the corresponding normalization note for CRPS so readers can interpret the reported 1.5–1.8× ratio correctly.
  5. [§4.8] The claim that cross-sensor degradation is 'graceful' is undercut by the large PSNR drop in Table 5 (17.36 to 10.00 dB at +3 h and 15.48 to 7.93 dB at +9 h). Even if part of the gap is attributable to terrain, the presentation should be more cautious and ideally report the per-channel breakdown for the over-ocean subset.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the forecast model, steering-flow track readout, and reward fine-tuning evaluation are all anchored to external ERA5/GRIDSAT/IBTrACS data and held-out test windows.

full rationale

The paper's central derivation is not circular. The latent rectified-flow model is trained with a standard flow-matching objective (Eq. 4) on external observations (GRIDSAT-B1 and ERA5), and its forecasts are evaluated against held-out 2022 storm data with standard pixel, probabilistic, and trajectory metrics. The steering-flow track readout (Eqs. 6-7) is a classical forward physical formula from Chan (2005), not a fitted parameter or a renamed version of the target. Reward fine-tuning directly optimizes a differentiable track-error reward, but the reported gains are measured on a held-out test set and at a sampler budget that was held out from fine-tuning, so the improvement is not a training-set artifact. The only notable comparison concern is that the Nath et al. baseline is described in Table 1 as imagery-only while Table 3 reports five-channel and track metrics for it; that is a fairness and reproducibility issue about the re-implementation, not a circularity of the paper's own derivation chain. No self-citations are load-bearing, and no claimed prediction reduces by construction to an input or to a fitted quantity.

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

The central claim rests on standard generative-model machinery and a classical steering-flow formula. The only hand-fitted constants are the latent scale factor and the steering-crop size, and the model introduces no new physical entities.

free parameters (3)
  • VAE latent scale factor s = 1/σ_z, estimated on the validation split
    Section 3.3: a scale factor rescales the latent representation to approximately unit variance for flow training; it is fitted on validation data.
  • Steering-flow crop size = 64x64
    Section 3.7: the area-mean winds are computed over the central 64x64 crop; this hand-chosen size determines the track readout for every storm.
  • Euler step count for evaluation = 50 steps (peak quality at 4 steps)
    Section 4.4 and 4.7: the default evaluation uses 50 Euler steps, while the sampler-budget sweep shows quality peaks at 4 steps; the chosen budget affects all timing and quality numbers.
assumptions (4)
  • domain assumption ERA5 reanalysis and GRIDSAT-B1 imagery are accurate enough to serve as ground truth for training and evaluation.
    The entire training signal and all evaluation metrics treat these archives as truth; systematic biases in the archives would propagate into the model. (Section 4.1)
  • domain assumption Tropical cyclone translation is well approximated by the area-mean of the generated winds over the central 64x64 crop.
    This classical steering approximation is used to compute every track error number; it is crude for small or rapidly recurving storms, and the paper reports 2x track errors on such storms. (Section 3.7)
  • ad hoc to paper Replacing missing values with zeros does not materially bias the learned distribution.
    Section 3.2 states missing entries are replaced with zeros; this is a strong assumption for pressure and wind fields where missing data may be systematic.
  • domain assumption The re-implementation of the Nath et al. cascaded diffusion baseline is faithful and competently tuned.
    All comparative claims rest on this reproduction, but no code or checkpoints are provided to verify parity with the original. (Section 4.3)

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tropical Cyclone Forecasting via Latent Rectified Flow using Satellite Imagery and Atmospheric Fields." pith.science (2026). https://pith.science/paper/6JH7UTQK

@misc{pith2026260808354,
  author       = {Pith},
  title        = {Pith review of: Tropical Cyclone Forecasting via Latent Rectified Flow using Satellite Imagery and Atmospheric Fields},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6JH7UTQK}},
  note         = {Machine review of arXiv:2608.08354}
}
read the original abstract

Tropical cyclones are growing more destructive in a changing climate, and efficient forecasting of their structure and track has become a necessity. Deep generative models promise an alternative to computationally expensive numerical weather prediction (NWP), yet current systems produce either satellite imagery or atmospheric fields, never both; they need many sampling steps, putting them out of reach of modest hardware; and their storm tracks come from regression heads with no physical link to the generated atmosphere. This work presents a single-pass model that jointly forecasts GRIDSAT-B1 infrared imagery and four ERA5 atmospheric fields (U-wind, V-wind, air temperature, and surface pressure) out to nine hours. A five-channel variational autoencoder compresses each 5 x 256 x 256 frame to a 4 x 64 x 64 latent, and a conditional rectified-flow UNet with a factorized temporal-attention module predicts the next three frames from three past frames, their best-track coordinates, and timestamps. The model is then reward-fine-tuned (DRaFT) against a differentiable track error derived from the predicted winds through a steering-flow calculation. On held-out 2022 storms the model reaches 16.35 dB PSNR and 0.759 SSIM, ahead of a reproduced cascaded-diffusion baseline at every lead time (+0.84 dB at +9 h) while sampling ~30x faster (56 ms vs. 1673 ms). Track error at +9 h is 62.4 km, 15% below the baseline, and a reward fine-tuning study demonstrates a further 8-11% track-error reduction across sampler budgets.

Figures

Figures reproduced from arXiv: 2608.08354 by the authors.

Figure 1
Figure 1. Two-stage framework: VAE compression (Stage 1), conditioned flow-matching UNet with temporal attention [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. All-channel forecast for a held-out 2022 test storm: the last input frame, then ground truth (GT) and prediction [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 7 canonical work pages

  1. [1]

    Forecasting tropical cyclones with cascaded diffusion models

    Pritthijit Nath, Pancham Shukla, Shuai Wang, and César Quilodrán-Casas. Forecasting tropical cyclones with cascaded diffusion models. InICLR Workshop on Tackling Climate Change with Machine Learning, Vienna, Austria, 2024. doi: 10.48550/arXiv.2310.01690

    arXiv doi:10.48550/arxiv.2310.01690 2024

  2. [2]

    Global tropical cyclone intensity forecasting with multi-modal multi-scale causal autoregressive model

    Xinyu Wang, Kang Chen, Lei Liu, Tao Han, and Bin Li. Global tropical cyclone intensity forecasting with multi-modal multi-scale causal autoregressive model. InAdvances in Neural Information Processing Systems, volume 37, Vancouver, Canada, 2024

    work page 2024

  3. [3]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InAdvances in Neural Information Processing Systems, volume 33, pages 6840–6851, 2020

    2020

  4. [4]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, New Orleans, LA, USA, 2022. doi: 10.1109/CVPR52688.2022.01042

    arXiv 2022

  5. [5]

    Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Remi Lam, and Matthew Willson

    Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R. Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Remi Lam, and Matthew Willson. Probabilistic weather forecasting with machine learning.Nature, 637:84–90, 2025. doi: 10.1038/s41586-024-08252-9

    doi:10.1038/s41586-024-08252-9 2025

  6. [6]

    Improving tropical cyclone forecasting with video diffusion models

    Zhibo Ren, Pritthijit Nath, Pancham Shukla, and César Quilodrán-Casas. Improving tropical cyclone forecasting with video diffusion models. InICLR Workshop on Tackling Climate Change with Machine Learning, Singapore,

    work page

  7. [7]

    TC-Diffuser: Bi-condition multi-modal diffusion for tropical cyclone forecasting

    Sijia Zhang, Pan Mu, Cheng Huang, Jinglin Zhang, and Cong Bai. TC-Diffuser: Bi-condition multi-modal diffusion for tropical cyclone forecasting. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 32099–32107, Philadelphia, PA, USA, 2025. doi: 10.1609/aaai.v39i1.32099

    work page doi:10.1609/aaai.v39i1.32099 2025

  8. [8]

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

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InInternational Conference on Learning Representations, Kigali, Rwanda, 2023

    work page 2023

Show all 20 references
  1. [9]

    Kevin Clark, Paul Vicol, Kevin Swersky, and David J. Fleet. Directly fine-tuning diffusion models on differentiable rewards. InInternational Conference on Learning Representations, Vienna, Austria, 2024. doi: 10.48550/arXiv. 2309.17400

    2024 doi arXiv

  2. [10]

    Johnny C. L. Chan. The physics of tropical cyclone motion.Annual Review of Fluid Mechanics, 37:99–128, 2005. doi: 10.1146/annurev.fluid.37.061903.175702

    2005

  3. [11]

    Accurate medium-range global weather forecasting with 3D neural networks.Nature, 619:533–538, 2023

    Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3D neural networks.Nature, 619:533–538, 2023. doi: 10.1038/s41586-023-06185-3

    2023 doi

  4. [12]

    Learning skillful medium-range global weather forecasting.Science, 382(6677):1416–1421, 2023

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, et al. Learning skillful medium-range global weather forecasting.Science, 382(6677):1416–1421, 2023. doi: 10.1126/science.adi2336

    2023 doi

  5. [13]

    Cheng Huang, Pan Mu, Cong Bai, and Peter A. G. Watson. TCP-Diffusion: A multi-modal diffusion model for global tropical cyclone precipitation forecasting with change awareness. InProceedings of the International Conference on Machine Learning, Vienna, Austria, 2024. doi: 10.48...

    2024 doi arXiv

  6. [14]

    Estimating atmospheric variables from digital typhoon satellite images via conditional denoising diffusion models.arXiv preprint arXiv:2409.07961, 2024

    Zhangyue Ling, Pritthijit Nath, and César Quilodrán-Casas. Estimating atmospheric variables from digital typhoon satellite images via conditional denoising diffusion models.arXiv preprint arXiv:2409.07961, 2024. doi: 10.48550/arXiv.2409.07961

    2024 doi arXiv

  7. [15]

    Knapp, Steve Ansari, Caroline L

    Kenneth R. Knapp, Steve Ansari, Caroline L. Bain, Mark A. Bourassa, Michael J. Dickinson, Chris Funk, Chip N. Helms, Christopher C. Hennon, Christopher D. Holmes, George J. Huffman, James P. Kossin, Hai-Tien Lee, Alexander Loew, and Gudrun Magnusdottir. Globally gridded satell...

    2011 doi

  8. [16]

    The ERA5 global reanalysis.Quarterly Journal of the Royal Meteorological Society, 146 (730):1999–2049, 2020

    Hans Hersbach et al. The ERA5 global reanalysis.Quarterly Journal of the Royal Meteorological Society, 146 (730):1999–2049, 2020. doi: 10.1002/qj.3803

    1999 doi

  9. [17]

    Knapp, Michael C

    Kenneth R. Knapp, Michael C. Kruk, David H. Levinson, Howard J. Diamond, and Charles J. Neumann. The international best track archive for climate stewardship (IBTrACS): Unifying tropical cyclone data.Bulletin of the American Meteorological Society, 91(3):363–376, 2010. doi: 10...

    2010 doi

  10. [18]

    Flow-GRPO: Training flow matching models via online RL

    Jie Liu, Gongye Liu, Jiajun Liang, Yangguang Li, Jiaheng Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Wanli Ouyang. Flow-GRPO: Training flow matching models via online RL. InAdvances in Neural Information Processing Systems, San Diego, CA, USA, 2025. doi: 10.48550/arXiv.2505.05470

    2025 doi arXiv

  11. [19]

    Diffusion fine-tuning with rewarded moment matching distillation.arXiv preprint arXiv:2606.30414, 2026

    Alexis Jacq, Guillaume Couairon, Valentin De Bortoli, Quentin Berthet, Arnaud Doucet, and Romuald Elie. Diffusion fine-tuning with rewarded moment matching distillation.arXiv preprint arXiv:2606.30414, 2026. doi: 10.48550/arXiv.2606.30414. 9

    2026 doi arXiv

  12. [2025]

    doi: 10.48550/arXiv.2501.16003

    doi arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.