Soot temperature is a relevant factor related to the efficiency of combustion processes. Artificial neural networks have started to be used to estimate soot temperature distributions in laminar flames by analyzing images captured with optical setup of varying complexity. These networks often achieve greater accuracy and precision than traditional methods that rely on explicit theoretical models and numerical approaches. However, most prior studies validate the neural networks on limited subsets of canonical flames, which may lead to overfitting. For these methods to be practically useful, a trained network should generalize across diverse flame conditions without needing retraining. This paper introduces the use of Attention U-Net models for soot pyrometry, utilizing only broadband flame emission images captured with a RGB camera. Simulation results demonstrate that the Attention U-Net achieves more accurate temperature estimations compared to previously reported learning-based methods. Additionally, we evaluate the model's generalization capabilities, showing that a network trained on simulated data maintains high accuracy when applied to images of laminar flames across various experimental conditions with errors below 30 K. Tests with experimental data further reveal that the proposed approach, using a single, produces temperature estimates comparable to those obtained through well-established techniques that require more complex equipment and processing. Moreover, the network exhibits strong robustness to measurement noise and remains effective inflames with low soot loading, where traditional reference techniques suffer from reduced signal-to-noise ratios and diminished accuracy.