Pix2gestalt is a framework for zero-shot amodal segmentation, which learns to estimate the shape and appearance of whole objects that are only partially visible behind occlusions. The method leverages large-scale diffusion models to reconstruct whole objects in challenging zero-shot cases, including examples that break natural and physical priors. The approach uses a synthetically curated dataset containing occluded objects paired with their whole counterparts. Experiments show that the method outperforms supervised baselines on established benchmarks. The model can significantly improve the performance of existing object recognition and 3D reconstruction methods in the presence of occlusions. The diffusion framework allows sampling several variations of the reconstruction, naturally handling the inherent ambiguity of occlusions. The paper introduces pix2gestalt, a method for amodal segmentation and reconstruction by synthesizing whole objects from partially visible ones. The method uses a conditional diffusion model to generate whole objects behind occlusions and other obstructions. The approach is based on denoising diffusion models, which are excellent representations of the natural image manifold and capture all different types of whole objects and their occlusions. The method is trained on a synthetic dataset of occlusions and their whole counterparts, creating a conditional diffusion model that, given an RGB image and a point prompt, generates whole objects behind occlusions. The paper evaluates pix2gestalt's ability to perform zero-shot amodal completion for three tasks: amodal segmentation, occluded object recognition, and amodal 3D reconstruction. The results show that the method provides amodal completions that directly lead to strong results in all tasks. The method outperforms existing methods in amodal segmentation, occluded object recognition, and 3D reconstruction. The method is also shown to be effective in improving the performance of existing object recognition and 3D reconstruction methods in the presence of occlusions. The diffusion framework allows sampling several variations of the reconstruction, naturally handling the inherent ambiguity of occlusions. The method is able to synthesize several plausible completions of occluded objects, and the results show that the method generalizes well beyond the typical occlusion scenarios found in those benchmarks. The method is also shown to be effective in handling out-of-distribution images, including art pieces, illusions, and images taken by the authors. The method is able to generate diverse samples in shape and appearance when there is uncertainty in the final completion. The method is also shown to be effective in handling common-sense and physics failures. The method is able to synthesize whole objects in novel situations, including artistic pieces, images taken by an iPhone, and illusions. The method is able to generate accurate, complete reconstructions of occluded objects on both COCO-A and BSDS-A datasets. The method is able to synthesize several plausible completions of the occluded house in the painting. The method is able to generate diverse samples in shape and appearance when there is uncertainty in the final completion. The method is able to synthesize several plausible completions ofPix2gestalt is a framework for zero-shot amodal segmentation, which learns to estimate the shape and appearance of whole objects that are only partially visible behind occlusions. The method leverages large-scale diffusion models to reconstruct whole objects in challenging zero-shot cases, including examples that break natural and physical priors. The approach uses a synthetically curated dataset containing occluded objects paired with their whole counterparts. Experiments show that the method outperforms supervised baselines on established benchmarks. The model can significantly improve the performance of existing object recognition and 3D reconstruction methods in the presence of occlusions. The diffusion framework allows sampling several variations of the reconstruction, naturally handling the inherent ambiguity of occlusions. The paper introduces pix2gestalt, a method for amodal segmentation and reconstruction by synthesizing whole objects from partially visible ones. The method uses a conditional diffusion model to generate whole objects behind occlusions and other obstructions. The approach is based on denoising diffusion models, which are excellent representations of the natural image manifold and capture all different types of whole objects and their occlusions. The method is trained on a synthetic dataset of occlusions and their whole counterparts, creating a conditional diffusion model that, given an RGB image and a point prompt, generates whole objects behind occlusions. The paper evaluates pix2gestalt's ability to perform zero-shot amodal completion for three tasks: amodal segmentation, occluded object recognition, and amodal 3D reconstruction. The results show that the method provides amodal completions that directly lead to strong results in all tasks. The method outperforms existing methods in amodal segmentation, occluded object recognition, and 3D reconstruction. The method is also shown to be effective in improving the performance of existing object recognition and 3D reconstruction methods in the presence of occlusions. The diffusion framework allows sampling several variations of the reconstruction, naturally handling the inherent ambiguity of occlusions. The method is able to synthesize several plausible completions of occluded objects, and the results show that the method generalizes well beyond the typical occlusion scenarios found in those benchmarks. The method is also shown to be effective in handling out-of-distribution images, including art pieces, illusions, and images taken by the authors. The method is able to generate diverse samples in shape and appearance when there is uncertainty in the final completion. The method is also shown to be effective in handling common-sense and physics failures. The method is able to synthesize whole objects in novel situations, including artistic pieces, images taken by an iPhone, and illusions. The method is able to generate accurate, complete reconstructions of occluded objects on both COCO-A and BSDS-A datasets. The method is able to synthesize several plausible completions of the occluded house in the painting. The method is able to generate diverse samples in shape and appearance when there is uncertainty in the final completion. The method is able to synthesize several plausible completions of