The mask is currently calculated by looking at the first HMI image and selecting where pixels > 0:

hmi = torch.Tensor(self.hmi_data[2010][ALL_COMPONENTS[0]][0])
hmi_mask = (torch.abs(hmi) > 0.0).to(dtype=torch.uint8)
hmi_mask_ratio = hmi_mask.sum().item() / hmi_mask.numel()

When applied to AIA this looked to be too agressive when reprojected:

This will be worth revisiting if we wish to focus on limb activity.