**Abstract:** This paper introduces pix2gestalt, 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. By leveraging large-scale diffusion models and transferring their representations to this task, the framework learns a conditional diffusion model for reconstructing whole objects in challenging zero-shot cases, including examples that break natural and physical priors, such as art. The training data consists of a synthetically curated dataset containing occluded objects paired with their whole counterparts. Experiments show that pix2gestalt outperforms supervised baselines on established benchmarks and can significantly improve the performance of existing object recognition and 3D reconstruction methods in the presence of occlusions. **Introduction:** The ability to visualize and recognize whole objects from only partial visibility is crucial for various applications in vision, graphics, and robotics. Pix2gestalt addresses the challenge of amodal completion by learning to synthesize whole objects from partially visible ones. The approach capitalizes on denoising diffusion models, which are excellent representations of the natural image manifold and capture all types of whole objects and their occlusions. By fine-tuning a pre-trained diffusion model on a synthetic dataset of varied occlusions, the framework can generate accurate and diverse completions. **Related Work:** The paper reviews related work in amodal completion, analysis by synthesis, and denoising diffusion models. It highlights the limitations of prior work, which often operate in closed-world settings or rely on synthetic data. Pix2gestalt, by contrast, generalizes to diverse zero-shot settings and outperforms state-of-the-art methods in both closed-world and open-world scenarios. **Amodal Completion via Generation:** The method uses a conditional diffusion model to predict the whole object from an input image and a point or mask prompt. The model is trained to maintain the zero-shot capabilities of the pre-trained diffusion model while learning to group and complete the object. The approach is evaluated on amodal segmentation, occluded object recognition, and amodal 3D reconstruction tasks, demonstrating strong performance and generalization capabilities. **Experiments:** The paper evaluates pix2gestalt on several benchmarks, including Amodal COCO and Amodal Berkeley Segmentation datasets for amodal segmentation, Occluded and Separated COCO for occluded object recognition, and Google Scanned Objects for amodal 3D reconstruction. The results show that pix2gestalt outperforms existing methods, providing accurate and complete reconstructions of occluded objects and improving the performance of downstream tasks. **Conclusion:** Pix2gestalt is a novel approach for zero-shot amodal segmentation that leverages whole object priors learned by large-scale diffusion models. The method demonstrates strong performance and generalization capabilities, making it a valuable tool for various computer vision tasks involving occlusions.**Abstract:** This paper introduces pix2gestalt, 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. By leveraging large-scale diffusion models and transferring their representations to this task, the framework learns a conditional diffusion model for reconstructing whole objects in challenging zero-shot cases, including examples that break natural and physical priors, such as art. The training data consists of a synthetically curated dataset containing occluded objects paired with their whole counterparts. Experiments show that pix2gestalt outperforms supervised baselines on established benchmarks and can significantly improve the performance of existing object recognition and 3D reconstruction methods in the presence of occlusions. **Introduction:** The ability to visualize and recognize whole objects from only partial visibility is crucial for various applications in vision, graphics, and robotics. Pix2gestalt addresses the challenge of amodal completion by learning to synthesize whole objects from partially visible ones. The approach capitalizes on denoising diffusion models, which are excellent representations of the natural image manifold and capture all types of whole objects and their occlusions. By fine-tuning a pre-trained diffusion model on a synthetic dataset of varied occlusions, the framework can generate accurate and diverse completions. **Related Work:** The paper reviews related work in amodal completion, analysis by synthesis, and denoising diffusion models. It highlights the limitations of prior work, which often operate in closed-world settings or rely on synthetic data. Pix2gestalt, by contrast, generalizes to diverse zero-shot settings and outperforms state-of-the-art methods in both closed-world and open-world scenarios. **Amodal Completion via Generation:** The method uses a conditional diffusion model to predict the whole object from an input image and a point or mask prompt. The model is trained to maintain the zero-shot capabilities of the pre-trained diffusion model while learning to group and complete the object. The approach is evaluated on amodal segmentation, occluded object recognition, and amodal 3D reconstruction tasks, demonstrating strong performance and generalization capabilities. **Experiments:** The paper evaluates pix2gestalt on several benchmarks, including Amodal COCO and Amodal Berkeley Segmentation datasets for amodal segmentation, Occluded and Separated COCO for occluded object recognition, and Google Scanned Objects for amodal 3D reconstruction. The results show that pix2gestalt outperforms existing methods, providing accurate and complete reconstructions of occluded objects and improving the performance of downstream tasks. **Conclusion:** Pix2gestalt is a novel approach for zero-shot amodal segmentation that leverages whole object priors learned by large-scale diffusion models. The method demonstrates strong performance and generalization capabilities, making it a valuable tool for various computer vision tasks involving occlusions.