Combining DA3-Metric Depth with DA3-Nested Pose for Multi-view Reconstruction?

Hi, and thanks for the great work on DA3, I'm new to the reconstruction field and this is truly impressive.

I’m trying to reconstruct a small indoor scene from 3 views with low overlap. Visual correctness matters more than strict metric accuracy at this stage.

**My setup is:** Single moving camera + Known intrinsics, unknown extrinsics + 3 views, low overlap (for max information).
I tried to compare reconstruction results on DA3-METRIC-LARGE and DA3NESTED-GIANT-LARGE.

**My observation:** DA3-METRIC-LARGE produces very crisp depth maps, often better along edges, especially on objects farther away from the camera, compared to DA3NESTED-GIANT-LARGE-1.1.
I am converting to metric depth using `metric_depth = net_output * fx / 300`.

**For 3D point cloud reconstruction I use:**
Metric:
```
depth_all = prediction_metric.depth # from metric model
K_all = intrinsics # from ground truth, scaled to our image
Extrinsics_all = prediction.extrinsics # from nested model
images_all = prediction_metric.processed_images # Already numpy uint8 array
```
Nested:
```
depth_all = prediction.depth
K_all = prediction.intrinsics
Extrinsics_all = prediction.extrinsics
images_all = prediction.processed_images
```

**Result:**
- Per-view point clouds from the metric model look much better and accurate to the real-world depths
- Multi-view alignment is more consistent with the nested model, whereas merging multi-views of the metric's point clouds results in much less accurate alignment (poses are coarsely correct, but do not fit together: this makes sense as I'm using the nested models pose estimation, and that pose estimation fits with the nested predictions since the model is trained for relative depth / pose. )
- Trying to use the 3 views from metric and run ICP / registration fails due to low overlap

Is there a recommended way to combine: metric model depth quality with the nested model pose consistency to get the best of both worlds?
Or do I need to have a fixed rig with 3 calibrated cameras and use both known intrinsic and extrinsic information to acquire the desired result?
Is this a known limitation given low overlap and missing pose estimation in the metric model?
Any pointers would be greatly appreciated.