Physics-based diffusion model for thermal image generation in space orbital rendez-vous scenarios

This paper presents an innovative physics-informed diffusion framework for generating thermal infrared imagery from visible-band observations in spacecraft rendezvous for on-orbit servicing (OOS) scenarios. To the best of our knowledge, this is the first learning-based approach to address visible-to-thermal translation tailored to space proximity operations, where strong cross-spectral appearance shifts and space-specific environmental effects limit the applicability of conventional image-to-image methods. The proposed model is trained conditionally using a composite, multi-feature representation designed to bridge the modality gap between visible and thermal bands and to improve robustness across illumination and viewing geometries. Training follows a multistage strategy that couples nominal generation with a refinement phase applied to sampled outputs, yielding improved structural fidelity and reduced artifacts. Finally, we introduce a model-driven physical constraint that embeds the radiometric temperature into the learning objective, guiding the refinement stage toward physically consistent radiance patterns and enhancing realism. Experiments in representative rendezvous configurations demonstrate that the method produces sharper, more plausible thermal renderings while preserving target geometry and its key thermally salient features, supporting downstream perception tasks in rendezvous operations.