Objective: Building fire safety design commonly relies on computational fluid dynamics simulations to obtain fire temperature fields. However, simulating the complete fire process in complex scenarios is time-consuming and cannot readily meet rapid assessment requirements. Existing methods for predicting fire temperature fields with conditional generative adversarial networks mostly rely on static mappings from building conditions to single-frame temperature fields, without fully accounting for the temporal dependencies between temperature fields at adjacent time steps as a fire develops. To address this issue, this study develops a temporal conditional generative adversarial network (TC-GAN), incorporating historical-state constraints into the pix2pix model. This study further optimizes the model to mitigate error accumulation in autoregressive prediction, thereby enabling rapid dynamic prediction of temperature-field evolution in single-story building fires. Methods: A fire dataset was constructed using Fire Dynamics Simulator. It comprised 35 single-story building layouts, each configured with two fire-source locations, resulting in a total of 70 fire scenarios. Given the building layout, fire-source location, and prediction time, TC-GAN introduces the temperature field at the previous time step as a historical-state input and, during inference, continuously generates temperature-field sequences in an autoregressive manner. To mitigate error accumulation, three optimization strategies, namely, multi-step loss, scheduled sampling, and a video discriminator, were compared. Model performance was comprehensively evaluated using the normalized root mean square error (NRMSE), the structural similarity index measure (SSIM), temperature-evolution curves at key monitoring points, and the intersection over union (IoU) and area error of high-temperature regions above 70 ℃. Results: The results showed that TC-GAN with only historical-state constraints was affected by autoregressive error accumulation and exhibited lower predictive performance than pix2pix, whereas the model with scheduled sampling achieved the best performance. On the test set, the model with scheduled sampling reduced average NRMSE by 39.3% and increased average SSIM by 24.6% compared with pix2pix. Temperature-field visualizations showed that this variant captured the temperature diffusion process in complex building spaces more consistently, reducing the local high-temperature accumulation, spatial-pattern fragmentation, and abnormal diffusion observed in the pix2pix predictions. The predicted curves at key monitoring points likewise showed closer agreement with the reference temperature trends. For regions with temperatures above 70 ℃, the model with scheduled sampling increased the average IoU from 0.357 4 to 0.561 1 and decreased the average absolute area error from 0.276 5 to 0.212 2, indicating improved predictive capability for capturing the location and extent of high-temperature regions. Conclusions: Historical-state constraints supply temporal evolution information for building-fire temperature-field prediction, but effective error-suppression strategies are required to mitigate autoregressive error accumulation. Under the conditions investigated in this study, TC-GAN combining scheduled sampling with historical-state constraints achieved the best overall predictive performance, outperforming the conventional pix2pix model in temperature-field accuracy, spatial structural consistency, temporal evolution at key monitoring points, and prediction of high-temperature regions. This method enables rapid prediction of temperature-field evolution in single-story building fires, reduces the time and computational costs of conventional numerical simulation, and provides a potential technical pathway for building fire safety design and rapid fire-risk assessment.