Load Depth Model (AnotherUtils)

Most "model loader" nodes expect you to have already downloaded weights, figured out which folder they go in, and guessed the right filename. AnotherLoadDepth breaks that pattern: it auto-downloads Depth Anything V3 on first run and hands you a ready-to-use model object. It's the front door to the pack's native depth inference, and for anyone who's ever stared at an empty models/ folder wondering where the weights were supposed to come from, it's a small mercy.

It's part of AnotherUtils (marcoc2/ComfyUI-AnotherUtils), whose whole pitch is "self-contained inference." Where the YOLO and SAM2 loaders in this pack manage their own weight downloads, this one leans on Hugging Face's own auto-download for Depth Anything V3 - you pick a size, it fetches the right checkpoint, done.

How it works

The loader doesn't load a file you point it at; it maps a name to a Hugging Face model ID and instantiates the Depth Anything V3 pipeline (DepthAnything3.from_pretrained(...)). Your device choice of auto/cuda/cpu decides where the model lands. The returned object is the pack's internal ANOTHER_MODEL dict - type tag "DepthAnything", plus the service wrapper and device info - which you then feed into AnotherDepthInference along with your images.

The inputs

Only two, both dropdowns:

• model_name - v3-tiny, v3-small (default), v3-medium, or v3-large. Size vs. speed, same as everywhere in this space: tiny is snappy but coarse, large is slower and sharper. For ControlNet conditioning, medium is a sweet spot.
• device - auto (picks CUDA if present, else CPU), cuda, or cpu. On a 12GB card the small models run fine; go CPU and you'll be waiting.

What comes out

A single ANOTHER_MODEL output. That's a custom type - it won't plug into standard nodes, only into the pack's own inference nodes, specifically AnotherDepthInference. Wire them together and you're live.

Installing it

Standard AnotherUtils install:

cd ComfyUI/custom_nodes
git clone https://github.com/marcoc2/ComfyUI-AnotherUtils.git

Restart ComfyUI, or search "AnotherUtils" in ComfyUI Manager. Then the one extra pip package the depth path needs:

pip install depth-anything-3

The weights themselves download automatically on first load. If you'd rather pre-fetch them outside a run, download the matching checkpoint from Hugging Face and let the library cache handle it.

Where people get burned

Two things trip people up. First, the missing depth-anything-3 package: the loader raises a clear ImportError when it's absent, so install it before you wire the graph. Second, the ANOTHER_MODEL output is pack-private - you can't feed it to a generic "Load Checkpoint" or a ControlNet apply node; it only talks to AnotherDepthInference. That's not a bug, it's the design: the pack keeps its inference self-contained. And one small real-world note: first load can take a while as it downloads several hundred MB, so don't panic if the node sits there on run one.