Video Depth Anything Metric S vs Metric B

[leastfamus]

I’m currently using a laptop with an RTX5060 and 8gb of VRAM. I have a good cooling pad so I’m not worried about heating issues.

I’m wondering if it is better to use VDA Metric S instead of VDA Metric B or not. I couldn’t find much info on VRAM requirements but according to deep wiki it looks as the the S model is better for less VRAM such as mine while Metric B requires 10-12 gb of VRAM to run efficiently. Maybe Nagadomi can answer this?

I can run Metric B but my fps is around 4-6 vs Metric S at 10-11. This is with torch compile on.

[nagadomi]

VRAM usage for VDA has been significantly reduced by my changes. The official VRAM usage for the Small model (518 resolution) is 6.8 GB, but in practice it runs at 2.4 GB.
DepthAnything/Video-Depth-Anything#24
The Base model should also run within 8 GB depending on the settings. It is expected to be slower than the Small model.

As a potentially confusing point regarding VRAM, on Linux an out of memory error simply occurs when VRAM runs out, while on Windows VRAM is swapped to system memory and continues running extremely slowly (Shared GPU Memory).
Also, the Windows Task Manager may not correctly report VRAM usage for NVIDIA GPUs. If performance becomes extremely slow, insufficient VRAM may be the cause.

You may feel more confident if you find a tool that can accurately reports VRAM usage.

[leastfamus]

Thank so much for getting back to me and answering my question. I really like iw3 by the way! This is the best conversion software I have found.