arxiv: 2512.22317 · v3 · submitted 2025-12-26 · 💻 cs.LG · cs.AI· cs.CV

Recognition: 1 theorem link

· Lean Theorem

LangPrecip: Language-Aware Multimodal Precipitation Nowcasting

Xudong Ling , Chaorong Li , Tianxi Huang , Qian Dong , Guiduo Duan

Authors on Pith no claims yet

Pith reviewed 2026-05-16 19:09 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CV

keywords precipitation nowcastingmultimodal learninglanguage-aware modelsrectified flowradar nowcastingweather forecastingheavy rainfall

0 comments

The pith

Meteorological text descriptions constrain radar-based precipitation forecasts to achieve higher accuracy in heavy rain events.

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

Existing nowcasting methods rely on radar images but leave future motion ambiguous, especially for extreme events. LangPrecip treats meteorological text as a semantic constraint on precipitation trajectories and solves the problem as a rectified flow generation task in latent space. This multimodal approach is supported by a new 160k paired dataset of radar sequences and motion descriptions. Tests show consistent gains over prior methods, with over 60 percent improvement in heavy-rainfall critical success index at 80-minute lead times on Swedish data and 19 percent on MRMS data.

Core claim

LangPrecip formulates short-term precipitation nowcasting as a semantically constrained trajectory generation problem under the Rectified Flow paradigm, enabling efficient integration of textual motion descriptions and radar information in latent space for physically consistent forecasts.

What carries the argument

Semantically constrained trajectory generation under the Rectified Flow paradigm that uses meteorological text as motion constraints on precipitation evolution.

If this is right

Consistent improvements over state-of-the-art methods on Swedish and MRMS datasets.
Over 60% gains in heavy-rainfall CSI at an 80-minute lead time on Swedish data.
19% gains in heavy-rainfall CSI at an 80-minute lead time on MRMS data.
Introduction of the LangPrecip-160k dataset with paired radar sequences and text descriptions.

Where Pith is reading between the lines

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

This approach could extend to other spatiotemporal prediction tasks like storm tracking where descriptive text is available.
Textual constraints might allow effective forecasting even with lower-resolution radar inputs in resource-limited settings.
Integration with real-time weather reports could enable dynamic updates to forecasts based on new textual observations.

Load-bearing premise

That the meteorological text descriptions accurately and sufficiently constrain future precipitation motion without introducing new ambiguities or biases that the model cannot resolve.

What would settle it

A controlled experiment showing that ablating the language input eliminates the reported CSI gains, or a dataset analysis revealing frequent mismatches between provided text and actual precipitation evolution, would falsify the claim.

Figures

Figures reproduced from arXiv: 2512.22317 by Chaorong Li, Guiduo Duan, Qian Dong, Tianxi Huang, Xudong Ling.

**Figure 1.** Figure 1: Overview of the proposed LangPrecip framework. Radar echo sequences are first encoded into a latent space, where spatiotemporal self-attention captures visual dynamics while semantic motion descriptions provide high-level constraints on precipitation evolution. The two streams are integrated in the latent space to guide future trajectory generation. Leveraging the LangPrecip160K text–radar paired dataset… view at source ↗

**Figure 2.** Figure 2: LangPrecip-160K dataset construction pipeline. tation motion patterns. These motion descriptions capture coarse-grained and temporally aggregated dynamics (e.g., dominant propagation direction and structural evolution) and are treated as high-level semantic motion priors that complement radar-based visual inputs rather than direct observational measurements. 3.2. Semantically Constrained Rectified Flow We … view at source ↗

**Figure 3.** Figure 3: The decoder uses a dual-path design that combines UpBlocks with Temporal Shift Modules for cross-frame alignment and wavelet-based unfolding blocks for data-consistent refinement, followed by feature fusion to produce the final output. Downsample (bilinear to 𝟑𝟐×𝟑𝟐 ) Upsample (𝑫𝒔 𝑻) 𝑿𝒉𝒓 (𝒕) (Input 𝒀𝒍𝒓 (Observ ation) ×𝟐𝜼 (Gradient Update) 𝜼:Learnable Parameter DWT (Forward) DWT (Forward) LL Soft Shrinkage S… view at source ↗

**Figure 5.** Figure 5: Spatial precipitation forecasts at multiple lead times (5-80 minutes) for different models with corresponding skill scores (CSI ≥ 0.06 mm/h, CSI ≥ 6.3 mm/h) during a Swedish weather event in Central Sweden (centered approximately at 63.17°N, 14.63°E, in Ostersund) on October 3, 2021, at 16:45. ¨ We solve this problem via an iterative scheme, where each iteration performs a data-consistency update by backp… view at source ↗

**Figure 6.** Figure 6: Temporal performance evolution on the Swedish dataset across lead times (5–80 minutes). All methods exhibit performance degradation as the prediction horizon increases [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗

**Figure 7.** Figure 7: Temporal performance evolution on the MRMS dataset across lead times (5–80 minutes). All methods exhibit performance degradation as the prediction horizon increases [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗

**Figure 8.** Figure 8: Dominant motion visualization of text-guided precipitation generation. Motion directions align with textual descriptions; 0 ◦ and 90◦ denote rightward and downward motion, respectively. alone may not align with meteorological fidelity. Figures 6 and 7 show all methods degrade with increasing lead time, but LangPreci maintains larger margins in CSI and FSS at longer horizons. This indicates that text-encod… view at source ↗

**Figure 9.** Figure 9: Effect of CFG scale on precipitation nowcasting performance. Top row: Swedish dataset. Bottom row: MRMS dataset. Event-based metrics (CSI) and spatial consistency (FSS) consistently improve with moderate CFG, while CRPS exhibits limited variation. 5. Ablation Experiment 5.1. Text scheduling degree [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗

read the original abstract

Short-term precipitation nowcasting is an inherently uncertain and under-constrained spatiotemporal forecasting problem, especially for rapidly evolving and extreme weather events. Existing generative approaches rely primarily on visual conditioning, leaving future motion weakly constrained and ambiguous. We propose a language-aware multimodal nowcasting framework(LangPrecip) that treats meteorological text as a semantic motion constraint on precipitation evolution. By formulating nowcasting as a semantically constrained trajectory generation problem under the Rectified Flow paradigm, our method enables efficient and physically consistent integration of textual and radar information in latent space.We further introduce LangPrecip-160k, a large-scale multimodal dataset with 160k paired radar sequences and motion descriptions. Experiments on Swedish and MRMS datasets show consistent improvements over state-of-the-art methods, achieving over 60 \% and 19\% gains in heavy-rainfall CSI at an 80-minute lead time.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

LangPrecip adds a new paired text-radar dataset and frames nowcasting as text-constrained rectified flow, but the abstract leaves the text source and ablation evidence too thin to confirm the claimed CSI gains.

read the letter

The main thing here is a new 160k paired radar-plus-text dataset and the idea of using meteorological descriptions as semantic constraints inside a rectified-flow trajectory model. That combination is not in the prior work they cite, and the reported CSI lifts at 80 minutes on Swedish and MRMS data (over 60% and 19% for heavy rain) would matter if they hold up. The practical angle for flood warning is clear enough that the work is worth a look from people who already run radar nowcasters. What the paper does cleanly is treat the text as an external motion prior rather than just another visual channel, which avoids some of the usual ambiguity in pure generative nowcasting. The dataset itself is the most concrete contribution so far. The soft spots are exactly where the stress-test note flags them. The abstract gives no information on how the motion descriptions were produced, whether they were written by humans or derived from the radar, or how much inter-annotator agreement exists. Without that, it is hard to know whether the text is adding real physical constraint or simply extra capacity that a larger visual model might also capture. There are also no ablations shown that isolate the language term from the flow rectification schedule or the alignment weight. The free parameters listed suggest the method still needs tuning, and the lack of error analysis on extreme cases leaves open whether the gains are uniform or concentrated in easier events. If the text turns out to be coarse or partially leaked from the radar sequences, the central claim weakens. This is the kind of paper that belongs in a weather-forecasting or multimodal time-series venue rather than a general ML conference. A serious referee would ask for the text-generation protocol, the full ablation table, and a breakdown of performance by event type. I would bring it to a reading group as a maybe, mainly to see the dataset and the actual conditioning mechanism. I would not cite it yet because the evidence is still abstract-level. It deserves peer review because the dataset is new and the framing is distinct, even if the current write-up needs more verification work.

Referee Report

3 major / 2 minor

Summary. The paper proposes LangPrecip, a multimodal nowcasting framework that formulates short-term precipitation forecasting as semantically constrained trajectory generation under the Rectified Flow paradigm, using meteorological text descriptions as motion constraints on radar data. It introduces the LangPrecip-160k dataset of paired radar sequences and text descriptions, and reports consistent CSI improvements over SOTA baselines on Swedish and MRMS datasets, including >60% and >19% gains for heavy-rainfall CSI at 80-minute lead times.

Significance. If the central claim holds, the work would represent a meaningful advance in generative nowcasting by demonstrating that external language can supply physically consistent constraints that reduce ambiguity in visual-only models. The release of a large-scale multimodal dataset would also provide a useful resource for future research on text-conditioned spatiotemporal forecasting.

major comments (3)

[Abstract / Experiments] Abstract and Experiments section: The headline CSI gains (over 60% on Swedish data and 19% on MRMS at 80 min lead time for heavy rainfall) are presented without ablation tables that isolate the language modality from added model capacity, without error bars or statistical significance tests, and without error analysis on failure cases. This prevents verification that the reported improvements stem from semantic motion constraints rather than other factors.
[Dataset / Methods] Dataset and Methods sections: No description is given of text provenance, generation process, inter-annotator agreement, or quality metrics for the LangPrecip-160k motion descriptions. Without this, it is impossible to rule out data leakage from radar sequences or to confirm that the text is sufficiently specific to disambiguate velocity fields, growth/decay, or orographic effects as claimed.
[Methods] Methods section: The text-radar alignment weight and flow rectification schedule are listed as free parameters, yet the paper asserts 'physically consistent' integration without showing how these parameters are chosen or whether the fusion remains robust when text descriptions contain ambiguities.

minor comments (2)

[Methods] Notation for the Rectified Flow latent-space integration could be clarified with an explicit equation showing how text embeddings are injected into the velocity field.
[Figures] Figure captions should explicitly label which panels show text-conditioned versus baseline predictions to aid comparison.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and will incorporate revisions to improve the clarity and verifiability of our claims.

read point-by-point responses

Referee: [Abstract / Experiments] Abstract and Experiments section: The headline CSI gains (over 60% on Swedish data and 19% on MRMS at 80 min lead time for heavy rainfall) are presented without ablation tables that isolate the language modality from added model capacity, without error bars or statistical significance tests, and without error analysis on failure cases. This prevents verification that the reported improvements stem from semantic motion constraints rather than other factors.

Authors: We agree that the current presentation does not sufficiently isolate the language contribution. In the revised manuscript we will add ablation tables comparing the full model against a capacity-matched vision-only baseline. We will also report standard deviations across multiple random seeds and include statistical significance tests (paired t-tests) on the CSI metrics at each lead time. Additionally, we will insert a dedicated error analysis subsection that examines representative failure cases, including instances of ambiguous text or weak radar signals, to clarify when semantic constraints provide the largest benefit. revision: yes
Referee: [Dataset / Methods] Dataset and Methods sections: No description is given of text provenance, generation process, inter-annotator agreement, or quality metrics for the LangPrecip-160k motion descriptions. Without this, it is impossible to rule out data leakage from radar sequences or to confirm that the text is sufficiently specific to disambiguate velocity fields, growth/decay, or orographic effects as claimed.

Authors: We will substantially expand the Dataset section. The motion descriptions were sourced from official meteorological reports and radar metadata provided by the Swedish Meteorological and Hydrological Institute and MRMS archives. Generation combined automated event parsing with expert human review to ensure descriptions capture motion, growth/decay, and orographic effects. In revision we will report the full provenance, the generation pipeline, inter-annotator agreement (Fleiss’ kappa on a sampled subset), and quality metrics such as keyword coverage for physical processes and a leakage audit confirming temporal but not content overlap with radar imagery. revision: yes
Referee: [Methods] Methods section: The text-radar alignment weight and flow rectification schedule are listed as free parameters, yet the paper asserts 'physically consistent' integration without showing how these parameters are chosen or whether the fusion remains robust when text descriptions contain ambiguities.

Authors: We will revise the Methods section to document the hyperparameter selection procedure: both the alignment weight and rectification schedule were chosen by grid search on a held-out validation set, optimizing for heavy-rain CSI while preserving trajectory smoothness. We will add a sensitivity analysis figure and accompanying text showing performance variation across reasonable ranges. To address robustness, we will include new experiments that deliberately introduce controlled ambiguities into the text (e.g., vague motion statements) and demonstrate that the rectified-flow formulation maintains physical consistency better than vision-only baselines. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on independent dataset and external Rectified Flow paradigm

full rationale

The paper's central claims rest on empirical CSI gains from a newly introduced LangPrecip-160k dataset paired with radar sequences and an external text encoder, formulated under the established Rectified Flow paradigm. No equations or steps reduce the reported predictions to fitted inputs by construction, nor do they depend on self-citations for uniqueness or ansatz smuggling. The multimodal fusion is presented as an additive constraint without definitional loops, and results are benchmarked against state-of-the-art methods on independent Swedish and MRMS datasets. This is a standard non-circular empirical contribution.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The framework rests on the assumption that text embeddings can be aligned with radar latent trajectories in a way that enforces physical consistency; several standard deep-learning hyperparameters and the choice of text encoder are free parameters.

free parameters (2)

text-radar alignment weight
Scaling factor balancing textual constraint against radar fidelity during rectified flow training.
flow rectification schedule
Hyperparameters controlling the noise-to-data trajectory in the latent space.

axioms (1)

domain assumption Rectified flow produces physically plausible trajectories when conditioned on semantic embeddings
Invoked when claiming efficient and physically consistent integration of text and radar.

pith-pipeline@v0.9.0 · 5456 in / 1307 out tokens · 33469 ms · 2026-05-16T19:09:37.164789+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

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

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

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

We formulate nowcasting as a semantically constrained trajectory generation problem under the Rectified Flow paradigm... minimizing the mean squared error L = E ||u(xt, cctx, t; θ) − vt||² where cctx = {X0:4, m} and m is a motion description

What do these tags mean?

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

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages · 6 internal anchors

[1]

Terra: A multimodal spatio-temporal dataset spanning the earth

Chen, W., Hao, X., Wu, Y., and Liang, Y. Terra: A multimodal spatio-temporal dataset spanning the earth. In Advances in Neural Information Processing Systems, 2024

work page 2024
[2]

W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al

Chung, H. W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al. Scaling instruction-finetuned language models. Journal of Machine Learning Research, 25 0 (70): 0 1--53, 2024

work page 2024
[3]

B., Li, M., and Yeung, D.-Y

Gao, Z., Shi, X., Wang, H., Zhu, Y., Wang, Y. B., Li, M., and Yeung, D.-Y. Earthformer: Exploring space-time transformers for earth system forecasting. Advances in Neural Information Processing Systems, 35: 0 25390--25403, 2022

work page 2022
[4]

Gao, Z., Shi, X., Han, B., Wang, H., Jin, X., Maddix, D., Zhu, Y., Li, M., and Wang, Y. B. Prediff: Precipitation nowcasting with latent diffusion models. Advances in Neural Information Processing Systems, 36: 0 78621--78656, 2023

work page 2023
[5]

Denoising diffusion probabilistic models

Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33: 0 6840--6851, 2020

work page 2020
[6]

Imagen Video: High Definition Video Generation with Diffusion Models

Ho, J., Chan, W., Saharia, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D. P., Poole, B., Norouzi, M., Fleet, D. J., et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022
[7]

Precipitation nowcasting using diffusion transformer with causal attention

Li, C., Ling, X., Xue, Y., Luo, W., Zhu, L., Qin, F., Zhou, Y., and Huang, Y. Precipitation nowcasting using diffusion transformer with causal attention. IEEE Transactions on Geoscience and Remote Sensing, 2024

work page 2024
[8]

Alphapre: Amplitude-phase disentanglement model for precipitation nowcasting

Lin, K., Zhang, B., Yu, D., Feng, W., Chen, S., Gao, F., Li, X., and Ye, Y. Alphapre: Amplitude-phase disentanglement model for precipitation nowcasting. In Proceedings of the Computer Vision and Pattern Recognition Conference, pp.\ 17841--17850, 2025

work page 2025
[9]

Two-stage rainfall-forecasting diffusion model

Ling, X., Li, C., Qin, F., Zhu, L., and Huang, Y. Two-stage rainfall-forecasting diffusion model. IEEE Geoscience and Remote Sensing Letters, 21: 0 1--5, 2024 a

work page 2024
[10]

Spacetime separable latent diffusion model with intensity structure information for precipitation nowcasting

Ling, X., Li, C., Zhu, L., Qin, F., Zhu, P., and Huang, Y. Spacetime separable latent diffusion model with intensity structure information for precipitation nowcasting. IEEE Transactions on Geoscience and Remote Sensing, 2024 b

work page 2024
[11]

Flow Matching for Generative Modeling

Lipman, Y., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022
[12]

Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

Liu, X., Gong, C., and Liu, Q. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022
[13]

Ssrf-net: A stagewise scheduled rainfall forecasting network with an asymmetric architecture

Luo, W., Li, C., Ling, X., Deng, C., and Wang, Z. Ssrf-net: A stagewise scheduled rainfall forecasting network with an asymmetric architecture. IEEE Transactions on Geoscience and Remote Sensing, 63: 0 1--18, 2025

work page 2025
[14]

Latte: Latent diffusion transformer for video generation

Ma, X., Wang, Y., Chen, X., Jia, G., Liu, Z., Li, Y.-F., Chen, C., and Qiao, Y. Latte: Latent diffusion transformer for video generation. Transactions on Machine Learning Research, 2025

work page 2025
[15]

Chaosbench: A multi-channel, physics-based benchmark for subseasonal-to-seasonal climate prediction

Nathaniel, J., Qu, Y., Nguyen, T., Yu, S., Busecke, J., Grover, A., and Gentine, P. Chaosbench: A multi-channel, physics-based benchmark for subseasonal-to-seasonal climate prediction. arXiv preprint arXiv:2402.00712, 2024

work page arXiv 2024
[16]

Open-Sora 2.0: Training a Commercial-Level Video Generation Model in $200k

Peng, X., Zheng, Z., Shen, C., Young, T., Guo, X., Wang, B., Xu, H., Liu, H., Jiang, M., Li, W., et al. Open-sora 2.0: Training a commercial-level video generation model in 200 k. arXiv preprint arXiv:2503.09642, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[17]

A., Velasco-Forero, C., Seed, A., Germann, U., and Foresti, L

Pulkkinen, S., Nerini, D., P \'e rez Hortal, A. A., Velasco-Forero, C., Seed, A., Germann, U., and Foresti, L. Pysteps: An open-source python library for probabilistic precipitation nowcasting (v1. 0). Geoscientific Model Development, 12 0 (10): 0 4185--4219, 2019

work page 2019
[18]

Skilful precipitation nowcasting using deep generative models of radar

Ravuri, S., Lenc, K., Willson, M., Kangin, D., Lam, R., Mirowski, P., Fitzsimons, M., Athanassiadou, M., Kashem, S., Madge, S., et al. Skilful precipitation nowcasting using deep generative models of radar. Nature, 597 0 (7878): 0 672--677, 2021 a

work page 2021
[19]

Skilful precipitation nowcasting using deep generative models of radar

Ravuri, S., Willson, M., and et al. Skilful precipitation nowcasting using deep generative models of radar. Nature, 597(7878): 0 672--677, 2021 b

work page 2021
[20]

Convolutional lstm network: A machine learning approach for precipitation nowcasting

Shi, X., Chen, Z., Wang, H., Yeung, D.-Y., and Wong, W.-K. Convolutional lstm network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems, 28, 2015

work page 2015
[21]

Wan: Open and Advanced Large-Scale Video Generative Models

Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.-W., Chen, D., Yu, F., Zhao, H., Yang, J., et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[22]

Rainhcnet: Hybrid high-low frequency and cross-scale network for precipitation nowcasting

Wang, L., Wang, Z., Hu, W., and Bai, C. Rainhcnet: Hybrid high-low frequency and cross-scale network for precipitation nowcasting. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2025 a

work page 2025
[23]

Nowcasting echo top for aviation operations using cnn-transformer

Wang, S., Sun, M., and Li, Y. Nowcasting echo top for aviation operations using cnn-transformer. IEEE Transactions on Intelligent Transportation Systems, 2025 b

work page 2025
[24]

Lavie: High-quality video generation with cascaded latent diffusion models

Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al. Lavie: High-quality video generation with cascaded latent diffusion models. International Journal of Computer Vision, 133 0 (5): 0 3059--3078, 2025 c

work page 2025
[25]

Diffcast: A unified framework via residual diffusion for precipitation nowcasting

Yu, D., Li, X., Ye, Y., Zhang, B., Luo, C., Dai, K., Wang, R., and Chen, X. Diffcast: A unified framework via residual diffusion for precipitation nowcasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 27758--27767, 2024

work page 2024
[26]

Open-Sora: Democratizing Efficient Video Production for All

Zheng, Z., Peng, X., Yang, T., Shen, C., Li, S., Liu, H., Zhou, Y., Li, T., and You, Y. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[27]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

work page