The core difficulty that Influpaint addresses is uncertainty that refuses to behave neatly. Traditional mechanistic models, built on equations describing transmission between susceptible and infected individuals, and classical statistical models fitted to historical patterns, both tend to produce forecasts that average over possible futures. Yet real influenza seasons are notoriously multimodal: a season might plausibly peak early and severe, or late and mild, and a forecast that hedges between these scenarios can end up predicting something that matches neither. When an unexpected variant emerges or holiday travel reshapes contact patterns, models anchored to past behavior often miss the turn entirely. Capturing this branching, lumpy uncertainty—many distinct plausible trajectories rather than one smoothed expectation—has been a persistent weakness of existing approaches.

Denoising diffusion models offer a different path. These models, which underpin many of today’s leading image generators, learn to create data by reversing a gradual noising process. During training, the model is shown examples that are progressively corrupted with random noise, and it learns to strip that noise away step by step. Once trained, the model can start from pure noise and iteratively refine it into a realistic sample drawn from the distribution it learned. Crucially, this generation can be made conditional: the model can be guided to produce samples consistent with partial observations, filling in the missing pieces much like an image-editing tool completes a partially masked photograph. Epidemiologists recognized that this inpainting capability maps naturally onto forecasting, where the observed recent past is known and the future is the region to be filled in.

Influpaint exploits that mapping with an elegant representational trick. Instead of treating influenza data as a table of weekly case counts by region, the researchers encode each influenza season as a spatiotemporal image. In these images, one axis represents geography—individual states or regions of the United States—while the other represents time, and the intensity of each pixel corresponds to influenza incidence in that place at that moment. A whole season’s epidemic, unfolding across the entire country, becomes a single picture. Patterns that are hard for conventional models to capture, such as a wave sweeping from the South toward the Northeast or synchronized peaks across distant states, become visual textures that a convolutional, image-oriented model can learn directly. The forecasting task then becomes literally an act of inpainting: given the left portion of the image, which shows the season so far, generate the right portion that completes it.

Training such a model requires a rich dataset, and here the team made a deliberate and consequential design choice. Surveillance data alone, however valuable, cover a limited number of seasons and may not span the full range of dynamics the model should learn to represent. To expand and diversify the training material, the researchers built a hybrid dataset combining real surveillance observations with simulated epidemic trajectories generated from models of influenza transmission. The simulations act as a kind of synthetic widening of experience, exposing the network to plausible epidemic shapes that history has not yet delivered. The composition of this mixture turned out to matter greatly: the best performance was achieved with a training set containing roughly 30 percent real surveillance data and 70 percent simulated trajectories, a balance that provided enough grounding in reality while leveraging the breadth of the synthetic data.

The evaluation of Influpaint proceeded in two stages, beginning with retrospective tests against historical data. In these experiments, the model was asked to generate forecasts as if it stood at various points during past seasons, and its outputs were scored against what actually happened. The results showed that Influpaint produces realistic and diverse epidemic trajectories, generating ensembles of possible futures that span the plausible outcome space rather than collapsing onto a single average path. Its forecast accuracy proved competitive with leading ensemble methods, the established top performers in influenza forecasting competitions. This is a notable result for a fundamentally new approach: rather than incrementally refining existing mechanistic or statistical machinery, the diffusion model matched them by learning the structure of epidemics directly from data.

The more demanding test came in real time. Influpaint was entered into the U.S. Centers for Disease Control and Prevention’s FluSight challenges during the 2022–2023, 2023–2024, and 2024–2025 influenza seasons, providing weekly forecasts alongside other teams’ models while the season was still unfolding. Performance improved substantially across the three seasons, reflecting both refinements to the model and the accumulating benefits of training data. The 2024–2025 season stood out: the model’s projections were highly accurate, capturing the course of the epidemic with a precision that few competitors matched. Yet the same season revealed a characteristic weakness. The forecasts, while accurate on average, were somewhat overconfident, meaning the model expressed more certainty in its predictions than the actual spread of outcomes warranted. In probabilistic forecasting, where decision-makers rely on the width of prediction intervals to gauge risk, overconfidence can be as damaging as inaccuracy.

That tension—sharp accuracy paired with miscalibrated confidence—illustrates both the promise and the open questions of generative epidemic forecasting. Diffusion models excel at learning rich distributions and producing samples that respect complex spatiotemporal structure, but ensuring that the spread of generated samples faithfully reflects true uncertainty requires careful calibration. The authors’ findings indicate that the framework is flexible and improvable: because forecasting is cast as conditional generation, the same machinery could in principle incorporate additional data streams, such as hospitalizations, virological reports, or behavioral indicators, simply by conditioning on more observed channels of the spatiotemporal image. The approach is not tied to influenza specifically, and the researchers present it as a general framework for probabilistic infectious disease forecasting rather than a single-purpose tool.

The broader significance of the work lies in what it says about the migration of generative AI into the sciences. Image-generation models succeeded because they learned the statistics of natural pictures without being told explicit rules of composition. Influpaint suggests that epidemic dynamics, when framed appropriately, have learnable statistical structure of a similar kind—waves that propagate, peaks that cluster, seasons that rhyme without repeating. For public health, the practical stakes are considerable: better-calibrated, spatially resolved forecasts could sharpen decisions about vaccine distribution and healthcare capacity during the critical weeks when an influenza season’s character is still being decided. For the modeling community, the study establishes a benchmark showing that generative approaches can stand shoulder to shoulder with the best ensemble methods, and points toward a research agenda in which simulated and observed data are deliberately blended to teach models about futures that have not yet happened. As diffusion models continue to mature, the picture of next winter’s flu season may arrive not as a single predicted curve, but as a gallery of plausible epidemics—each one painted by an algorithm that has learned, from data and simulation alike, how influenza seasons take shape.

Subject of Research: Generative diffusion models applied to spatiotemporal influenza forecasting

Article Title: Generative diffusion models for spatiotemporal influenza forecasting

Article References: Lemaitre, J., & Lessler, J. (2026). Generative diffusion models for spatiotemporal influenza forecasting. PLOS Computational Biology, 22(9), e1014846. https://doi.org/10.1371/journal.pcbi.1014846

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014846

Keywords: influenza, epidemic forecasting, diffusion models, machine learning, Influpaint, FluSight, CDC, spatiotemporal modeling, probabilistic forecasting, PLOS Computational Biology, public health, inpainting

The research, published in the journal Water Resources Research, describes a diffusion-based probabilistic modeling framework for hourly streamflow prediction and data assimilation. The team trained their model on rainfall and streamflow data collected from 516 river basins across the continental United States between 1990 and 2003, and then tested it on a window of data spanning 2009 to 2014 — a period the model had never encountered during training. According to the researchers, the diffusion approach generally outperformed the best existing hourly deep-learning models, particularly in capturing how much water flows through rivers during high-flow periods, which are notoriously difficult to anticipate. The work was led by Chaopeng Shen, professor of civil and environmental engineering at Penn State and corresponding author on the paper, whose group has spent years applying machine learning to national-scale flood forecasting.

The connection between image generation and flood prediction may sound whimsical, but the underlying mathematics is deeply relevant to hydrology. Generative AI models must be primed with existing information called training data, and they make predictions based on patterns observed in previous data sets. Shen describes the process as giving a model an image or a sentence with a portion missing and asking it to predict the content in that unknown region. Before the age of AI, scientists had to manually assimilate data, relying on numerous assumptions and what Shen calls mathematical gymnastics to make predictions. The explosion of AI science over the last decade has fundamentally changed how hydrologists — the scientists who study how water moves through the environment — approach forecasting, allowing them to assimilate information far faster and more easily using generative models trained on existing data.

The new model differs in important ways from the large language models that power conversational chatbots. Instead of text, it incorporates diffusion, a training method most commonly used in image generation. In diffusion training, a model first receives a complete dataset or picture. Once that clear data has been processed, noise — randomized, unpredictable information — is deliberately introduced into the dataset. The AI model is then tasked with removing the excess noise and reproducing an interpretable dataset or photo. Through countless iterations of this corruption-and-reconstruction cycle, the model learns the statistical structure of the data so thoroughly that it can generate realistic new examples or fill in missing information. For hydrologists, that reconstruction skill translates directly into the ability to reason about messy, incomplete, and uncertain environmental measurements.

This training approach turns out to be particularly effective for flood prediction because weather data such as rainfall can be extraordinarily unpredictable on an hourly basis. The diffusion framework allows the model to quantify uncertainty in its predictions, and it reduces doubt by assimilating recent gauge observations into the dataset through a process known as inpainting. Inpainting, in the image-generation world, means filling in a missing or corrupted region of a picture; in the flood-forecasting world, it means injecting fresh, near-real-time measurements into the model without the need for retraining. That distinction matters enormously for operational forecasting, where retraining a model on new data is often too slow and too expensive to be practical during an unfolding emergency.

Shen explains the mechanics with a concrete example. The basic training inputs include hourly rainfall, projected streamflow — the rate at which water flows in a river — and all the data collected from previous patterns. But the model can also accept new, almost real-time information. Given the daily streamflow recorded yesterday and the hourly rainfall observed today, the model can be asked to predict what the hourly peak water level might be today. This ability to fold the latest gauge readings into an ongoing prediction, without retraining, gives forecasters a continuously updated picture of flood risk, exactly the kind of rapid situational awareness that flash flood response demands.

The motivation for moving from daily to hourly forecasting is rooted in hard experience. Shen’s team had previously delivered a daily flood forecasting model that made accurate predictions using historical rainfall and streamflow, but it struggled to capture the severity of rapid weather swings that produce flash floods. Many of the most devastating floods in recent memory, including the flooding that swept across Central Texas in the summer of 2025, swell and recede in a matter of hours. On a daily averaging scale, Shen notes, water levels might still seem elevated but not disaster-inducing, even as an extraordinary peak races through a river channel. By calibrating the model for an hourly scale, the team can more reliably capture those short-lived but catastrophic surges that wreak havoc on communities.

The versatility of the diffusion framework extends beyond forward prediction. Shen explains that the model can also run in reverse: given the streamflow recorded at a river gauge, it can estimate the hourly rainfall that most likely produced that flow. This inverse capability offers another layer of utility for hydrologists, potentially helping to reconstruct rainfall events in basins where precipitation measurements are sparse or unreliable, and to validate observations against what the rivers themselves recorded. To verify the accuracy of their predictive model across the country’s river basins, the team built a map visualizing each basin’s Nash-Sutcliffe Efficiency, a standard metric used to assess the performance of hydrological models, with darker dots indicating less accurate predictions and lighter dots indicating more accurate ones.

The research team includes, in addition to Shen, several Penn State contributors: Wencong Yang, a postdoctoral researcher; Yalan Song, assistant research professor of civil and environmental engineering; Kathryn Lawson, a research associate; and civil and environmental engineering doctoral candidates Haoyu Ji and Leo Lonzarich. Ming Pan of the Scripps Institution of Oceanography at the University of California, San Diego, also contributed to the work. The effort was supported by the Cooperative Institute for Research to Operations in Hydrology and the National Oceanic and Atmospheric Administration Cooperative Agreement, with additional funding from the California Department of Water Resources.

Looking ahead, the team plans to continue development, investigating the benefits of different hybrid approaches to training AI models and building models that incorporate more physical variables into their predictions. For Shen, however, the ultimate objective is not about crowning a single best model. He observes that no single approach stands out as superior — each model has its own aspect it handles most comfortably. Instead, he frames the mission in human terms: the goal is to deliver a forecasting system that is robust, reliable, and capable of helping save people’s lives, and he believes the field is making real progress in that direction. As generative AI techniques mature and merge with the physical sciences, the same machinery that conjures fantastical images on demand may increasingly stand guard over riverside communities, translating noisy streams of rainfall and gauge data into the precious minutes that separate warning from tragedy.

Subject of Research: Diffusion-based generative AI for hourly flash flood and streamflow prediction

Article Title: The techniques AI uses to create images can help predict floods, researchers say

Article References: The techniques AI uses to create images can help predict floods, researchers say. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: flood forecasting, generative AI, diffusion models, hydrology, streamflow prediction, flash floods, machine learning, data assimilation, inpainting, Penn State, Water Resources Research, river basins

The central insight behind TamperNet is that video tampering almost always leaves two kinds of fingerprints at once. The first kind lives in the spatial domain: a spliced object may have inconsistent lighting, unnatural edges, or texture statistics that do not match its surroundings. The second kind lives in the temporal domain: duplicated or deleted frames disturb the smooth flow of motion, and inpainted regions often fail to reproduce the plausible dynamics that a real camera would capture. Most existing detection approaches emphasize one of these cues at the expense of the other, which makes them brittle when an editor deliberately suppresses the dominant artifact type. TamperNet instead fuses spatial and temporal evidence throughout the network, so that a manipulation must evade both channels simultaneously to slip past the detector.

Technically, the framework is built from several cooperating modules. A ResNet-50 backbone, a convolutional network originally developed for image recognition and widely reused as a feature extractor, serves as the spatio-temporal descriptor, converting raw frames into rich intermediate representations that encode edges, textures, and object structure. Because ResNet-50 processes frames individually, the authors pair it with a temporal representation module based on a long short-term memory network, or LSTM, a recurrent architecture specifically designed to capture dependencies over sequences. The LSTM reads the stream of frame-level features and learns what normal temporal evolution looks like, so that abrupt discontinuities, repeated segments, or missing intervals register as anomalies in the learned sequence representation.

Fusion is handled by concatenation, a straightforward but effective strategy in which the spatial feature vectors and the temporal feature vectors are joined end to end before being passed to the decision stages. This design choice matters because it preserves the full information content of both branches rather than compressing them prematurely into a single embedding. On top of the fused representation, TamperNet applies a spatio-temporal anomaly scoring strategy that combines two complementary error signals. The first is a temporal prediction error, which measures how badly the model’s learned dynamics fail to anticipate the actual next frames; manipulations that break continuity produce spikes in this error. The second is an autoencoder reconstruction error, in which a network trained to compress and reconstruct normal video content produces unusually large residuals when asked to reconstruct manipulated regions it has never effectively learned to model.

The combination of these two error signals gives the framework a degree of redundancy that single-cue detectors lack. A skilled forger might smooth out temporal discontinuities by re-encoding the video, degrading the usefulness of prediction error alone, but the altered content would still tend to reconstruct poorly, and vice versa. By scoring each candidate video against both criteria and fusing the results with the concatenated deep features, TamperNet can flag manipulations that span both the inter-frame domain, where the relationships between successive frames are disturbed, and the intra-frame domain, where the internal consistency of a single frame has been violated. This dual coverage is what allows one model to address such a heterogeneous list of forgery types, from crude frame duplication to sophisticated content-aware inpainting.

The team evaluated the framework on the Video Forgery Dataset, a benchmark abbreviated VFD that contains both authentic and manipulated video clips. The reported results are strong on the classification metrics that matter most for a binary detection task. TamperNet achieved an accuracy of 93.3 percent, meaning it correctly labeled nearly nineteen out of twenty videos. Precision, the fraction of flagged videos that were genuinely tampered, came in at 91.4 percent, while recall, the fraction of tampered videos that were successfully caught, reached 91.3 percent, indicating that the model’s errors are balanced rather than skewed toward false alarms or missed forgeries. The F1-score, the harmonic mean of precision and recall that penalizes imbalance between them, was 89.9 percent.

One metric, however, deserves careful attention. The area under the receiver operating characteristic curve, or ROC-AUC, which summarizes detection performance across all possible decision thresholds, was approximately 0.75. An ROC-AUC of 1.0 represents perfect ranking of tampered above authentic videos, while 0.5 represents chance. A value of 0.75 indicates discrimination clearly better than random but well short of the near-perfect separation implied by the headline accuracy figure, and the authors are candid about this gap. They note that the moderate ROC-AUC, together with the relatively small size of the tested benchmark, means that TamperNet’s demonstrated performance should be understood as evidence of feasibility on the tested set rather than proof of readiness for the messy conditions of real-world forensic deployment.

That candor is one of the more notable features of the paper. The authors explicitly state that further cross-dataset testing and compression-specific testing are required to assess how well the framework generalizes beyond the VFD benchmark. This matters because videos in the wild arrive heavily compressed by platform-specific codecs, resized, re-encoded multiple times, and captured by sensors with wildly different noise characteristics, and each of these transformations can mimic or mask the very artifacts a detector depends on. A model that excels on a curated benchmark can lose substantial accuracy when confronted with aggressive compression, and the research team’s decision to flag this limitation rather than bury it reflects a maturing attitude within the video forensics community, where benchmark overfitting has repeatedly inflated published claims.

The research itself is a genuinely international collaboration, with authors affiliated with SZABIST University’s Department of Robotics and Artificial Intelligence in Islamabad, the Department of Computer Science at Aalto University in Espoo, Finland, Amity University Dubai in the United Arab Emirates, Western Caspian University in Baku, Azerbaijan, and Jadara University in Irbid, Jordan. Open access funding was provided by Aalto University, and the corresponding author is Usman Muhammad. The article was received on 26 March 2026, accepted on 2 September 2026, and published on 1 October 2026 under a Creative Commons Attribution 4.0 license, meaning the full technical details, architecture descriptions, and experimental protocols are freely available to any researcher who wishes to build on or stress-test the work.

The road ahead, as the authors sketch it, involves three main directions. First, validation on larger benchmark sets is needed to establish that the reported accuracy holds when training and evaluation data are more diverse and abundant. Second, the framework must be hardened against the full range of compression and acquisition scenarios that real forensic casework involves, from low-bitrate smartphone uploads to professionally encoded broadcast footage. Third, and perhaps most ambitiously, the team plans to extend TamperNet from detection to localization, so that instead of merely issuing a verdict that a video has been manipulated, the system would highlight precisely which frames and which pixel regions were altered. Such spatially and temporally grounded output would transform the tool from an alarm into an evidentiary instrument, giving investigators, courts, and platform trust-and-safety teams a concrete map of where a video’s fabrications begin and end. As synthetic and manipulated media continue to erode the default assumption that footage can be trusted, frameworks like TamperNet mark an incremental but meaningful step toward restoring that trust, provided their limitations are tested as rigorously as their strengths are celebrated.

Subject of Research: Spatio-temporal deep learning for detecting video manipulations such as frame duplication, splicing, and inpainting

Article Title: TamperNet: A Spatio-Temporal Deep Learning Framework for Video Manipulation Detection

Article References: Mujtaba, M. H., Imran, M., Irfan, S., Muhammad, U., & Dawood, H. (2026). TamperNet: A Spatio-Temporal Deep Learning Framework for Video Manipulation Detection. Neural Processing Letters. https://doi.org/10.1007/s11063-026-11885-8

Image Credits: AI Generated

DOI: 10.1007/s11063-026-11885-8

Keywords: video tampering detection, deep learning, spatio-temporal analysis, LSTM, autoencoder, ResNet-50, video forensics, frame splicing, inpainting, anomaly detection, Neural Processing Letters, digital forensics