LongE2V unifies three event-camera tasks under video diffusion

Event cameras record pixel-level brightness changes rather than full frames, and the stream they produce is sparse and asynchronous by design. That is useful for latency but awkward for anyone who wants to actually watch what happened, and the workarounds have historically been task-specific. A paper on arXiv from a team at National Yang Ming Chiao Tung University, accepted to SIGGRAPH 2026, argues you can unify three of the standard tasks, reconstruction, prediction and frame interpolation, under a single pre-trained video diffusion prior.

The framing the authors offer for why this problem was open is worth quoting. "Regression methods often blur textures, while existing generative models struggle with long-term stability" is how they describe the two failure modes their approach is aimed at. The proposed system, LongE2V, folds in three named components: Autoregressive Unrolling with Adaptive Context Switching to reduce temporal drift over long sequences, Reencoding Alignment with Cross Residual Correction for bidirectional consistency during interpolation, and Event Voxel Density Augmentation for robustness across varying sensor resolutions. The authors report outperforming state-of-the-art baselines across all three tasks; the reconstruction comparison names E2VID, FireNet, ET-Net, SPADE-E2VID, SSL-E2VID and HyperE2VID, with VDM-EVFI as the prediction baseline and CBMNet-Large and TLXNet+ on interpolation.

The reason the diffusion prior angle is the interesting part, and not the benchmark table, is what it implies about direction of travel. If a foundational video model can be fine-tuned into a competitive event-stream decoder without task-specific pipelines, the ceiling on sparse-sensor video recovery starts to look more like the ceiling on video generative models generally, which is a much faster-moving line. The authors' own claims are for "high data efficiency and superior perceptual quality" with "exceptional temporal coherence and zero-shot generalization", which is the language of a shared prior doing a lot of heavy lifting.

The honest caveat is that specific numerical metrics are not provided in the abstract, and the surfaced material does not name the diffusion backbone, the fine-tuning compute footprint, or the inference latency, which is the load-bearing property for anything downstream that actually cares about event cameras. What the reporting does not give you is how the model behaves on the long-tail cases event cameras were designed for. If the numbers hold up when the code drops, the more useful outcome for the field is not a new leaderboard entry but a template other event-vision groups can follow.