Severe spiky artifacts along object boundaries with DA3MONO-LARGE

Hello,

thank you for releasing Depth Anything 3.

I am using the monocular model:

depth-anything/DA3MONO-LARGE

on single portrait photographs.

I am seeing severe and repeated spiky artifacts along object/background boundaries in the predicted depth. These artifacts become especially problematic when the depth map is converted into a 3D relief or height-field surface.

The problem is most visible around silhouettes and other strong object boundaries. Instead of producing a smooth transition at the edge, the depth contains many narrow, high-frequency spikes protruding from the surface.

For example, in a portrait image, the boundary around the subject can produce dozens of thin spikes. When the depth is converted into a 3D relief, these become very pronounced geometric teeth/spikes along the contour.

Important observations:

- The artifacts are already present in the DA3 depth prediction.
- They are not caused by the 3D rendering or lighting.
- They occur repeatedly along strong object boundaries.
- Manual post-processing can remove them, but this is currently necessary to obtain a usable relief surface.
- The effect is particularly problematic for CNC/relief manufacturing because these spikes become real unwanted geometry.

I would like to ask:

1. Is this a known limitation or bug of DA3MONO-LARGE?
2. Is there a recommended preprocessing, inference, or postprocessing method to reduce these boundary spikes?
3. Is there an updated/fixed checkpoint planned for the monocular model?
4. The repository mentions refreshed "-1.1" checkpoints after fixing a training bug. Does this bug fix also apply to DA3MONO-LARGE, or is an updated monocular checkpoint planned?

I can provide the original input image and the resulting depth map if that would help reproduce the issue.

I have attached an example showing the resulting spiky geometry along the subject boundary.

Thank you for your work and for any guidance.

Start by reproducing the boundary artifacts with the attached portrait and DA3MONO-LARGE depth prediction. Compare the result with the repository’s refreshed “-1.1” checkpoints, if applicable, and review whether the issue is addressed by preprocessing or postprocessing. Done means documenting the cause, mitigation, or checkpoint status.

Written by the indexing model from the issue text.