LoadSALADModel

This is a helper-loader node with exactly one job: load a SALAD model and hand it to the streaming depth node to make its frame alignment more robust. It has no inputs and one output. You'll only ever see it as an optional attachment to DepthAnythingV3_Streaming - it doesn't do anything on its own, and most simple workflows never need it.

What SALAD is doing here

When DepthAnythingV3_Streaming chunks a long video and stitches the pieces together, it has to figure out how the chunks relate - which frames overlap, how to line them up so the depth scale is consistent across the seam. That alignment is the hard part of long-video depth, and it's where things go wrong (a visible "pop" in depth at a chunk boundary is an alignment failure). SALAD is an image-matching / place-recognition helper: it produces descriptors that let the pipeline recognize when two frames are looking at the same thing, which makes the chunk-to-chunk alignment sturdier than relying on the predicted poses alone.

You can see the hook on the other end: DepthAnythingV3_Streaming has an optional salad_model input, and its align_lib / align_method settings govern how chunks get registered. Feeding a SALAD model in gives that alignment step better matches to work from. That's the entire reason this node exists.

The interface

There's almost nothing to it, which is appropriate for a loader:

• No inputs. You just drop the node in.
• One output: salad_model (a SALAD_MODEL). It wires straight into the salad_model input on DepthAnythingV3_Streaming.

Like the depth-model loader, the first run will fetch the SALAD weights if they aren't already cached, so expect a one-time download the first time you use it.

When to bother

Use it when you're processing real video through the streaming node and your chunk seams aren't clean - depth jumping or drifting where windows meet is the symptom that says "your alignment needs help." Adding a SALAD model is one of the two levers for that (raising overlap on the streaming node is the other). For a single image, a short multi-view batch, or any workflow that doesn't touch DepthAnythingV3_Streaming, skip it entirely - there's nowhere for its output to go.

Common issues

The most likely snag is a first-run download: if the node seems to hang, it's probably fetching weights - let it finish, then it's cached. If loading fails outright, it's the same dependency story as the rest of the pack: a half-built environment where a required library didn't install cleanly, so reinstall the pack's requirements with --upgrade. And if plugging it in doesn't visibly improve your streaming result, that's not necessarily a bug - SALAD helps most when alignment was actually struggling. If your chunks were already lining up well on the predicted poses, you won't see much difference, and that's fine.