Atlas Depth Map 🌊

In the Atlas pack there are two nodes that run a depth model, and confusing them costs you real time. AtlasDepthAnything returns a normalized IMAGE you can look at. AtlasDepthMap returns an ATLAS_DEPTH_MAP - a shared, metric depth object that every geometry-derivation, layer, and look node in the pack consumes. Same family, opposite purpose. This one is the workhorse: estimate once, feed several.

The pack's design philosophy is visible right here. Depth is estimated once and passed around by reference (the catalog notes it's a zero-serialization custom type, same pattern as ATLAS_SOLVE), so every downstream node - AtlasDeriveReliefMesh, AtlasDepthLayerMask, AtlasDefocus, the clean-plate layers - agrees on the exact same metric scale. Wire one AtlasDepthMap into three consumers and they can't drift apart.

How it works

It runs your chosen monocular model and wraps the raw result - depth plus validity - into the ATLAS_DEPTH_MAP. Which model you pick changes what the map can do. The default, Depth Anything V2 Metric-Outdoor, is metric and transformer-only. The big difference-maker is MoGe (Ruicheng/moge-2-*), which additionally carries a native metric pointmap (per-pixel 3D coordinates) and predicted normals plus its own focal estimate. The KB's depth essay backs this up: MoGe is the geometry-oriented pick, outputting a point map a plain depth value can't express. Feed the solve's focal into MoGe and it's echoed in as fov_x, which is why those moge_* knobs exist.

Inputs that matter

Required: image. Then a stack of optional dials, and most of them are MoGe-specific - ignore them until you're on MoGe:

• depth_model - the same eleven-choice roster as its preview sibling (V2 Metric Outdoor/Indoor, V2 Small, DA3 variants, MoGe-2, DepthPro, Lotus-2). Default is V2 Metric-Outdoor-Large.
• solve - optional. Supplies the solved focal for DA3-Metric's canonical→metric conversion. Ignored by V2.
• moge_resolution_level - MoGe's token-budget dial; 9 is full detail, lower is faster/coarser.
• moge_max_side - cap the long edge before inference to save VRAM/time. The tooltip's example is telling: an 8K plate is ~415 MB of float32 on-device before MoGe even runs. Depth still comes back at source size.
• moge_tile_side - run inference on overlapping tiles at source resolution instead of downscaling, so a 36MP plate keeps fine structure. Costs one pass per tile plus a global pass (a 4×4 tiling is ~17× the time). Try 1024.
• moge_report_free_focal - runs a second fov-free pass just to cross-check MoGe's independent focal against the solve's. ~+30% time; useful when you suspect a focal mismatch.

One output: depth, the ATLAS_DEPTH_MAP. That's it. Everything downstream pulls from it.

Where people get tripped up

• Wiring the preview node into geometry. AtlasDepthAnything's IMAGE won't connect to a geometry node's depth socket - different type. Use this node.
• V2 on panoramas. The AtlasEquirectMultiView node deliberately doesn't offer V2 because it mis-scales panorama crops ~4.4×. Same principle applies to any very wide plate: MoGe is the safer interior/wide pick.
• Non-commercial weights. V2-Large is CC BY-NC 4.0; V2 small/base and the DA3-Metric/Mono backends are Apache 2.0. Pick per your use case.

Install

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

Restart (or Manager → "atlas-camera"). Requires the [neural] tier: torch + GeoCalib + transformers (GeoCalib is GitHub-only), and optionally the [moge] extra for MoGe - on portable ComfyUI install GitHub-only packages with --no-deps to protect your CUDA stack. The kornia <0.8.3 cap exists so these extras don't break ComfyUI-LTXVideo; if you see an import error about kornia.geometry.transform.pyramid, pip install "kornia==0.8.2" fixes it.

Brand-new pack, so no community folklore - the report strings and install docs are the truth. The habit to build: share one AtlasDepthMap across your whole graph.