pix2gestalt is a framework for zero-shot segmentation that learns to estimate the overall shape and appearance of partially visible objects. It leverages large-scale diffusion models and transfers their representations to this task, learning a conditional diffusion model for reconstructing the whole object in challenging zero-shot scenarios, including examples that break natural and physical priors like art. We utilize a synthetically curated dataset for training, containing occluded objects and their complete counterparts. Experiments demonstrate that our method outperforms supervised baselines on established benchmark tests. Furthermore, our model can be used to significantly enhance the performance of existing object recognition and 3D reconstruction methods in scenarios with occlusions.