Atlas Depth Anything V2 🧠

If you've used Depth Anything before - and in 2026 that's most ComfyUI users - this node is the familiar face in an unfamiliar pack. Image in, normalized depth image out, a model picker in the middle. What catches people is the name: it says "Depth Anything V2," but the dropdown actually offers eleven backends, and this node's real job is preview, not geometry.

Here's the distinction that matters, and the pack's own catalog states it bluntly: AtlasDepthAnything's IMAGE output is a lossy normalized preview that cannot be used for metric geometry. The sibling node, AtlasDepthMap, emits the ATLAS_DEPTH_MAP type the relief-mesh and layer nodes actually consume, carrying the raw DepthResult by reference. Different wire types, different purposes. If your graph ends at "look at my depth," this is the node. If anything downstream builds geometry, you want AtlasDepthMap.

How it works

AtlasDepthAnything runs a monocular depth model through the pack's inference layer and returns a viewable grayscale image. The KB's depth-estimation essay has the backstory: Depth Anything V2 became the community's default depth model for its quality-per-second balance, and its metric variants (outdoor/indoor) output real-world scale rather than just relative ordering - which is what makes them usable in a scene pipeline. This node is the shallow, display-facing slice of that.

The inputs that matter

Only one is required - image. Everything else is optional:

• depth_model - the picker, default depth-anything/Depth-Anything-V2-Metric-Outdoor-Large-hf. Choices include V2 Metric Outdoor/Indoor, V2 Small, Depth Anything 3 variants (DA3METRIC-LARGE, DA3MONO-LARGE, DA3NESTED-GIANT-LARGE-1.1), MoGe-2 (Ruicheng/moge-2-*), Apple's DepthPro-hf, and jingheya/Lotus-2. Pick by scene: outdoor/indoor V2 for the default, MoGe when you're headed toward geometry, DA3 if you want its monocular accuracy.
• device - auto / cuda / mps / cpu. Leave on auto unless you're forcing CPU/MPS.
• solve - an optional ATLAS_SOLVE. The tooltip's the whole story: it feeds the solved focal (GeoCalib or vanishing-point) into DA3-Metric's canonical→metric conversion, instead of assuming a normal-lens fallback. Ignored by plain V2 models. Wire it when you have a solve and want DA3-Metric to scale correctly.

Output: depth_image - an IMAGE, ready for a preview node.

The licensing catch

The pack README flags this clearly: Depth Anything V2's large weights are CC BY-NC 4.0 (non-commercial), while the small/base weights are Apache 2.0. If you're doing client work, either pick a permissively-licensed variant (V2 Small, or the Apache-licensed DA3-Metric/Mono backends, or MoGe) or clear the license. The KB's depth essay confirms this is the same trap as the V2 Giant story - the best weights are the ones you can't sell against.

Install

The atlas-camera pack, one clone:

cd <ComfyUI>/custom_nodes
git clone https://github.com/mikejamesvfx/atlas-camera.git

Restart (or Manager → "atlas-camera"). This node needs the [neural] tier - torch + GeoCalib + transformers, per INSTALL.md, with GeoCalib installed from GitHub. Good news for a change: the default Depth Anything V2 backend is transformers-only, so the [neural] extra needs no extra install beyond that. Models download from HuggingFace on first run. If you're on portable ComfyUI (python_embeded), the install docs insist on installing GeoCalib with --no-deps so pip can't clobber your torch/numpy ABI, then adding opencv-python, kornia<0.8.3, and transformers separately.

It's a brand-new pack (registered mid-2026, zero community war stories yet), so the docs are the ground truth. The short version: use this node to look, use AtlasDepthMap to build.