Conditional generation of building bubble diagrams based on stochastic differential equations

This study introduces a novel conditional generative model based on stochastic differential equations (SDEs) for synthesizing architectural bubble diagrams that meet specific customer requirements. The forward SDE progressively injects noise into the data, transforming it into a tractable prior distribution, while the reverse SDE removes the noise to reconstruct the original data distribution. Since the reverse SDE relies on the gradient of the data distribution (i.e., the score function), we employ a neural network to approximate these gradients. The trained score-based model enables conditional sampling from pure noise to generate new diagrams. To evaluate the quality of the generated outputs, we propose an effective metric tailored to conditionally generated graphs. Experimental results demonstrate that the proposed framework produces high-quality diagrams that adhere to specified structural constraints.