The paper introduces the Simformer, a novel method for simulation-based amortized Bayesian inference. Traditional methods often require fixed parametric priors, simulators, and inference tasks, making them simulation-intensive and inflexible. The Simformer overcomes these limitations by training a probabilistic diffusion model using transformer architectures. This approach allows the Simformer to handle function-valued parameters, missing or unstructured data, and arbitrary conditionals of the joint distribution of parameters and data, including posterior and likelihood. The method is demonstrated to outperform existing state-of-the-art approaches on benchmark tasks and is shown to be more flexible and efficient, particularly in handling complex and irregular data. The Simformer's ability to sample arbitrary conditionals and handle unstructured observations makes it a powerful tool for scientific and engineering applications.The paper introduces the Simformer, a novel method for simulation-based amortized Bayesian inference. Traditional methods often require fixed parametric priors, simulators, and inference tasks, making them simulation-intensive and inflexible. The Simformer overcomes these limitations by training a probabilistic diffusion model using transformer architectures. This approach allows the Simformer to handle function-valued parameters, missing or unstructured data, and arbitrary conditionals of the joint distribution of parameters and data, including posterior and likelihood. The method is demonstrated to outperform existing state-of-the-art approaches on benchmark tasks and is shown to be more flexible and efficient, particularly in handling complex and irregular data. The Simformer's ability to sample arbitrary conditionals and handle unstructured observations makes it a powerful tool for scientific and engineering applications.