Inconsistent Inference Results with Large Batch Sizes

I am using DA3NESTED-GIANT-LARGE for inference on the Things dataset. I noticed a significant degradation in the quality of the estimated depth maps when increasing the batch size, despite all input images having the same resolution.

I tested with an image of aircraft carrier below from Things dataset.


- Batch Size = 1 or 2: The results are clean and accurate.

- Batch Size = 10, 64, or 128: The depth maps exhibit strange noise, particularly in the sky regions. Also the global color distribution changes drastically between batch sizes..

(batch size = 10)

Is this expected behavior due to numerical precision issues with large batches, or is there a specific configuration needed to maintain consistency across batch sizes?

(batch size = 64)

Environment:

Model: DA3NESTED-GIANT-LARGE

Hardware: RTX 4090 48G

No files, tests, or entry points are identified in the report. First reproduce the depth-map differences on the Things image with batch sizes 1, 2, 10, 64, and 128, then trace the inference configuration to determine whether the results should remain consistent. Done means identifying the cause and documenting or fixing the configuration needed for consistent outputs.

Written by the indexing model from the issue text.