DownloadAndLoadDepthAnythingV3Model

This is the node that puts a Depth Anything 3 model into memory and hands it to everything else in the pack. Nothing in this node pack runs without it - the depth node, the point cloud nodes, the multi-view stuff, they all take a da3_model input, and this is where that comes from. If you're building a DA3 workflow, this is node one.

Depth Anything 3 is ByteDance's late-2025 depth model, and it's a real jump from the v2 that's been the community default since 2024. v2 gave you a depth map and that was it. v3 was trained jointly for depth and camera pose across any number of views, so it also spits out ray maps, camera intrinsics, and geometry you can fuse into point clouds. Same lineage that gave us the depth preprocessor everyone already uses - it's just grown a 3D half. This loader is the front door to all of it.

What it actually does

You pick a model variant, it makes sure the weights are on disk (downloading them if they aren't), builds the model object, and outputs it as a DA3MODEL. That's the whole job. It's the equivalent of a checkpoint loader, just for the depth backbone instead of a diffusion model.

The one input you care about is model - a dropdown of the seven DA3 variants. The default is da3_large.safetensors, which is the right call for most people: 350M params, high quality, doesn't melt your GPU. The ladder runs Small (80M, fast) → Base (220M) → Large (350M) → Giant (1.15B, best quality but slow), plus three specialist weights: Mono-Large (tuned for single-image depth), Metric-Large (real-world-scale depth), and Nested-Giant-Large (1.4B, the everything model). Note the licensing trap: Giant and Nested are CC BY-NC 4.0, non-commercial only. Small through Large are the safe commercial picks.

Which one to grab isn't just about size. The variants have genuinely different capabilities. If you want sky masks (which you do, for clean ControlNet depth), you need Mono, Metric, or Nested - the plain Small/Base/Large/Giant series doesn't produce them. If you want camera conditioning or multi-view attention for consistent video depth, you want the main series or Nested. For plain single images, honestly anything works, so start with Large and only reach for a specialist when a downstream node needs a feature it doesn't have.

The two optional inputs, precision and attention, both default to auto and you should almost always leave them there. auto picks fp16/bf16 and the best available attention backend for your card. Only touch them if you're chasing a specific out-of-memory or compatibility problem.

Output is a single da3_model - wire it straight into DepthAnything_V3, DepthAnythingV3_MultiView, DepthAnythingV3_Streaming, or through DA3_EnableTiledProcessing first if you're doing huge images.

Installing the pack

Three routes, easiest first. ComfyUI Manager is the one to try: search Depth Anything V3, install the highest version, and if that acts up, try the nightly. Manager also has an "Install via Git URL" option that takes https://github.com/PozzettiAndrea/ComfyUI-DepthAnythingV3.git. The manual route is the most reliable when the others fight you:

cd ComfyUI/custom_nodes
git clone https://github.com/PozzettiAndrea/ComfyUI-DepthAnythingV3.git
cd ComfyUI-DepthAnythingV3
pip install -r requirements.txt --upgrade
python install.py

Heads up: the pack leans on an experimental comfy-env one-click installer that pulls in the pixi package manager. The dependency stack (torch, xformers, and gsplat for the 3D bits) is genuinely picky about versions - the model's own release thread is full of people wrestling xformers into the right CUDA build. If install.py chokes, that's almost always the culprit. Match your xformers to your torch/CUDA and it settles down.

Common issues

The most frequent one isn't really an error: the first time you load a variant, it downloads the weights, so the node "hangs" for a few minutes on a cold run. That's normal - Large is a few hundred MB, Giant and Nested are gigabytes. Let it finish. After that it's cached.

If loading itself fails, it's the install, not the node - a half-built environment where xformers or gsplat didn't compile against your torch. Reinstall requirements with --upgrade and pin xformers to your CUDA version. And if you picked Giant or Nested and it's crawling or OOM-ing, that's expected; drop to Large. The quality gap between Large and Giant is real but small, and Large is the one you'll actually keep running.