DA3_DownloadModel

This node exists to do one thing: fetch a Depth Anything 3 model file to disk without loading it into memory. It's a utility, a convenience - you point it at a model variant, run it, and the weights get downloaded. That's it. If you just want to grab a model ahead of time, or make sure it's on disk before a big run, this is the node.

Why it's separate from the loader

The main loader, DownloadAndLoadDepthAnythingV3Model, already downloads weights if they're missing - that's the "Download" half of its name. So why a standalone downloader? Two honest reasons. First, pre-caching: on a fresh machine or a cloud instance, you might want to pull all the variants you plan to use up front, in a throwaway graph, so your real workflows never stall mid-run on a cold download. Second, clarity: a run of this node either succeeds or reports why it didn't, which makes it a clean way to confirm a download works before you build anything around it.

The interface

Dead simple. One required input:

• model - a dropdown of the DA3 variants, defaulting to da3_large.safetensors. Same list as the loader: Small, Base, Large, Giant, plus the Mono, Metric, and Nested specialists.

One output:

• status - a string telling you how it went. This is an output node, so running the graph triggers the download; the status string is your confirmation.

No precision, no attention, no model object - because it isn't loading anything, just fetching the file.

A licensing note worth repeating

Depth Anything 3 is a ByteDance release, and ByteDance's open components are generally permissive - but not uniformly. The Giant and Nested variants carry a CC BY-NC 4.0 license: non-commercial use only. Small, Base, and Large are the ones to download if your work is commercial. Downloading a model doesn't change its license, so pick the variant that matches how you'll use the output, not just the one with the best benchmark.

When to use it, when to skip it

Use it when you're setting up a machine and want to warm the cache, when you're scripting a reproducible environment, or when you're debugging whether a download even works. Skip it for a normal workflow - the loader downloads on demand anyway, so a plain depth or point-cloud graph never needs this node in front of it. It's infrastructure, not part of the creative chain.

Common issues

If the status comes back as a failure, it's almost always network or disk: a flaky connection mid-download, or no space left for a multi-gigabyte model (Giant and Nested are the big ones). Retry on a stable connection and check free space. If a later loader still says the model is missing after this reported success, make sure both nodes are looking at the same models directory - a mismatched path means the downloaded file is sitting somewhere the loader isn't checking. And the first download of a large variant just takes a while; a node that seems to hang is usually mid-transfer, not stuck.