Remote photoplethysmography estimates pulse signals from subtle color variations in facial videos, but remains sensitive to motion, illumination, and partial occlusion. We propose GridPulseFormer, a video Transformer that combines regional temporal consensus with robust waveform reconstruction. A three-dimensional convolutional stem converts video into a spatial grid that retains the arrangement of facial regions. A divided spatiotemporal Transformer adjusts spatial attention according to the agreement between latent regional trajectories. The output stage combines regional amplitude and first-order difference estimates through differentiable iterative reweighting to reconstruct the pulse waveform. We evaluate the model on PURE, UBFC-rPPG, and MMPD using within-dataset and cross-dataset comparisons, module ablations, and visualizations. In the within-dataset PURE experiment, the full model achieves a mean absolute error of 0.2250 bpm and a root mean square error of 0.3486 bpm. The ablations favor the joint use of the two modules over either module alone. Spectral and spatial visualizations reveal differences in harmonic peak selection and regional weight distributions. GridPulseFormer retains spatial information throughout feature modeling and imposes explicit waveform constraints for remote heart rate estimation.