Thermal-to-RGB Video Translation for Wildlife Monitoring: Enhancing Low-Resolution Thermal Imagery with Large Diffusion Models

Wildlife monitoring frequently relies on thermal cameras to capture animal activity in low-light or nocturnal conditions, as many species are primarily active at night. However, high-resolution thermal cameras are significantly more expensive than RGB cameras of comparable resolution, while low-cost thermal sensors generally produce low-resolution, noisy imagery, limiting interpretability for human operators and downstream machine learning tasks. We introduce the first framework for spatially and temporally consistent super-resolution RGB video generation from low-resolution thermal inputs, leveraging large diffusion models. Our pipeline operates in a fully tuning-free manner and is qualitatively evaluated across diverse animal species, datasets, thermal cameras, and resolutions, demonstrating strong generalization. The generated RGB outputs enhance scene interpretability while preserving temporal coherence, offering a practical tool for wildlife conservation, anti-poaching efforts, and behavioral analysis.