Enable always returning depth for Gaussian Splatting (#2873)

enable always returning depth, if specified in the config

nerfstudio/models/splatfacto.py

@@ -155,6 +155,8 @@ class SplatfactoModelConfig(ModelConfig):
     """threshold of ratio of gaussian max to min scale before applying regularization
     loss from the PhysGaussian paper
     """
+    output_depth_during_training: bool = False
+    """If True, output depth during training. Otherwise, only output depth during evaluation."""
 
 
 class SplatfactoModel(Model):
@@ -766,7 +768,7 @@ class SplatfactoModel(Model):
         alpha = alpha[..., None]
         rgb = torch.clamp(rgb, max=1.0)  # type: ignore
         depth_im = None
-        if not self.training:
+        if self.config.output_depth_during_training or not self.training:
             depth_im = rasterize_gaussians(  # type: ignore
                 self.xys,
                 depths,