This paper introduces a novel framework called Versatile Behavior Diffusion (VBD) for generating scene-consistent traffic scenarios. The framework leverages diffusion models, which are capable of generating realistic and controllable agent behaviors in traffic simulations. VBD integrates a scene encoder, a denoiser, and a behavior predictor to generate multi-agent interactions that are scene-consistent and controllable. The scene encoder encodes the scene context, the denoiser generates joint behaviors of agents from noise, and the behavior predictor forecasts individual agents' intentions as behavior priors. VBD is capable of generating scenarios conditioning on priors, integrating with model-based optimization, sampling multi-modal scene-consistent scenarios by fusing marginal predictions, and generating safety-critical scenarios when combined with a game-theoretic solver. The model is evaluated on the Waymo Sim Agents benchmark and demonstrates state-of-the-art performance. Additionally, the model is shown to be versatile, capable of adapting to various applications and generating diverse, realistic, and specified traffic scenarios through structured guidance. The paper also discusses the connection between diffusion-based generative modeling and imitation learning, and proposes various guidance strategies for the diffusion policy, including cost function, goal priors, and game theory structure, to generate diverse, realistic, and specified traffic scenarios. The results show that VBD achieves high performance in generating realistic and interactive traffic scenarios, and is capable of generating safety-critical scenarios through game-theoretic guidance. The model is also shown to be effective in generating scene-consistent scenarios through multi-step sampling and guided sampling. The paper concludes that VBD is a promising approach for generating scene-consistent traffic scenarios and has the potential to be applied in various applications.This paper introduces a novel framework called Versatile Behavior Diffusion (VBD) for generating scene-consistent traffic scenarios. The framework leverages diffusion models, which are capable of generating realistic and controllable agent behaviors in traffic simulations. VBD integrates a scene encoder, a denoiser, and a behavior predictor to generate multi-agent interactions that are scene-consistent and controllable. The scene encoder encodes the scene context, the denoiser generates joint behaviors of agents from noise, and the behavior predictor forecasts individual agents' intentions as behavior priors. VBD is capable of generating scenarios conditioning on priors, integrating with model-based optimization, sampling multi-modal scene-consistent scenarios by fusing marginal predictions, and generating safety-critical scenarios when combined with a game-theoretic solver. The model is evaluated on the Waymo Sim Agents benchmark and demonstrates state-of-the-art performance. Additionally, the model is shown to be versatile, capable of adapting to various applications and generating diverse, realistic, and specified traffic scenarios through structured guidance. The paper also discusses the connection between diffusion-based generative modeling and imitation learning, and proposes various guidance strategies for the diffusion policy, including cost function, goal priors, and game theory structure, to generate diverse, realistic, and specified traffic scenarios. The results show that VBD achieves high performance in generating realistic and interactive traffic scenarios, and is capable of generating safety-critical scenarios through game-theoretic guidance. The model is also shown to be effective in generating scene-consistent scenarios through multi-step sampling and guided sampling. The paper concludes that VBD is a promising approach for generating scene-consistent traffic scenarios and has the potential to be applied in various applications.