Lightweight thermal infrared to visible unpaired video translation

In the conversion of night-time infrared videos to visible light, high model complexity and difficult real-time edge deployment pose significant challenges. The existing ROMA model demonstrates excellent performance in generating high fidelity visible-light videos while preserving infrared scene structural information through cross-domain constraints and multi scale discrimination. However, its 16×16 fine-grained patch di vision and 9-layer ResNet lead to substantial computational resource demands during training and inference. To address this, we propose a lightweight framework that collaboratively optimizes large Patch ViT and shallow ResNet. While retaining global structural, local detail, and temporal consistency constraints, the framework significantly reduces self-attention computation and network depth. Experiments on the Infrared City-Lite dataset show that the model achieves competitive FID scores for style authenticity and notably improves computational efficiency. This work provides a lightweight solution for cross-modal conversion in resource constrained environments.