In this presentation, we show some drawbacks of the well-known Denoising Diffusion Probabilistic Models (DDPMs) – one of the most well-known synthesis approaches – that emerge when applied to thermal images in automotive scenarios for synthesis of novel objects.

We then show that, based on our internal research effort, we are able to improve these results and to generate novel images of objects better than the baseline. Our synthesized images exhibit high domain fidelity and realistic appearance. Finally, to further demonstrate the potential of this image synthesis method, we show that using synthesized images of a rare category we can improve its result on an object classification task.