Respiratory motion can induce anatomic motion of greater than 1.5 cm, which poses a significant challenge for targeting radiotherapy to thoracic solid malignancies. Traditional 4DCT models do not capture the wide variation in respiratory dynamics including the amplitude and frequency of respiration. Here, we use free-breathing, low-dose, fast helical scans (n = 25 scans per patient) to build a generative motion model. A conditional variational autoencoder model represents the deformation vector fields within the latent space, while the moving image is concatenated within the decoder. Current work is underway to additionally incorporate an external surrogate tracer (bellows) into the model to permit the generation of deformed volumes from a single external signal. This approach will enable on-the-fly real-time generation of 3D volumes while the patient is lying on the treatment table undergoing radiotherapy. Future experiments are planned to initiate the model with a cone beam CT scan obtained just prior to treatment delivery.