DA-Lion: Efficient Neural Video Representation via Direction-Aware Optimization
TL;DR
Direction-Aware Lion (DA-Lion), a task-driven optimizer tailored for NVR, introduces a direction-consistency criterion that switches between sign update and momentum update based on gradient--momentum alignment, and a learning-rate-aware magnitude modulation that stabilizes effective step sizes across training phases.
Abstract
Implicit neural video representation (NVR) encodes a video as the parameters of an overfitted neural network, where training and testing data are identical and the objective is instance-specific signal fitting. In this deterministic regime, optimization dynamics directly determine reconstruction quality under a fixed budget, yet existing methods universally adopt general-purpose optimizers designed for stochastic gradient training. Adam-family optimizers descend steadily but converge slowly, while Lion achieves faster early progress but oscillates markedly in later stages when gradient directions become unstable. We trace this instability to the insensitivity of the sign-based update to gradient--momentum alignment. To address this, we propose Direction-Aware Lion (DA-Lion), a task-driven optimizer tailored for NVR. DA-Lion introduces (i) a direction-consistency criterion that switches between sign update and momentum update based on gradient--momentum alignment, and (ii) a learning-rate-aware magnitude modulation that stabilizes effective step sizes across training phases. DA-Lion modifies only the parameter update rule, preserving model architecture, parameter count, FLOPs, and decoding throughput. Experiments on three datasets with four NeRV-family backbones show that DA-Lion improves convergence stability and boosts PSNR/MS-SSIM across all backbones. The source code will be made available at https://github.com/maoqingyu1996/DA-Lion.