AnyMod-LLVE: Modality-Agnostic Inference for Low-Light Video Enhancement (AMNet)

Forum topic · 小凯 · 2026-06-11

Summary

AnyMod-LLVE introduces AMNet, a unified multimodal framework for low-light video enhancement (LLVE) that supports flexible, modality-agnostic inference. Existing multimodal LLVE methods rely on auxiliary modalities such as event streams and infrared images being available at inference time, which is rarely feasible in real-world scenarios. To overcome modality absence, the authors propose a Spatial-Spectral Dual-Gated Translator that learns the correspondence between auxiliary modalities and RGB inputs, generating implicit auxiliary representations for robust enhancement. Trained on RGB-only datasets combined with large-scale multimodal pretraining, the model can handle arbitrary combinations of input modalities at inference and achieves strong performance even when auxiliary modalities are missing. Authors: Hangfeng Liang, Yutao Hu, Yanhan Hu, Xiaohan Wu, Wenqi Shao, Ying Fu. arXiv:2606.11186.

Paper Overview

Field: Computer Vision (CV)
Authors: Hangfeng Liang, Yutao Hu, Yanhan Hu, Xiaohan Wu, Wenqi Shao, Ying Fu
Published: 2026-06-09
arXiv: 2606.11186

Abstract

Low-light video enhancement (LLVE) remains a challenging task due to severe information degradation under low-illumination conditions. Recent multimodal approaches have significantly improved enhancement performance by incorporating auxiliary modalities, such as event streams and infrared images. However, these methods typically assume the availability of these modalities at inference, which is often not feasible in real-world scenarios.

To solve this problem, the authors propose AMNet, a unified multimodal framework for LLVE that supports flexible, modality-agnostic inference, where auxiliary modalities may be unavailable.

Key Points

Introduces a Spatial-Spectral Dual-Gated Translator that learns the correspondence between auxiliary modalities and RGB inputs.
Generates implicit auxiliary representations to support robust enhancement when real auxiliary modalities are missing.
Leverages an RGB-only dataset together with large-scale multimodal pretraining, enabling the model to handle arbitrary combinations of input modalities at inference time.
Achieves excellent performance under modality-absent conditions, addressing a key limitation of prior multimodal LLVE methods.