Atlas Split Equirect 🌐

AtlasSplitEquirect takes a 360° equirectangular panorama (normally 2:1, like 4096×2048) and cuts it into N perspective crops, each of which is already a valid Atlas pinhole camera. That sentence is doing more work than it looks like - Atlas is pinhole end to end: every solver, relief mesh and projection path assumes fx/fy/cx/cy intrinsics. An equirect is not a pinhole camera, so instead of modeling it as one, the pack just cuts it into perspective views that are. Twelve crops at 90° FOV is ComfyUI's default ring, and the tooltip explains why: overlap is what lets neighbouring depths agree where they meet.

Why bother? A single perspective plate has no data for what the camera never saw - which is the whole reason the pack's occlusion-graph, move-budget and guided-hole-repair machinery exists. A 360° capture supplies that coverage as real measured geometry instead of inventing it. This is the node that turns a panorama from "nice background" into "a full surrounding scene with real parallax."

Inputs that matter

• equirect - the 2:1 panorama, one image per run.
• n_views - crops around the ring (1–64, default 12).
• fov_deg - horizontal/vertical FOV of each square crop (default 90). Wider = fewer views to cover the circle but more lens distortion for the solver to fit.
• size - crop side length in pixels (default 1024).
• view_index - which view leaves the single view/exact_view outputs (clamped to n_views−1); all_views always carries the batch.
• pitch_deg - tilt of the whole ring. 0 keeps the horizon level, which is what the ground-plane scale fit wants.
• yaw_offset_deg - rotate the ring, so you can move a seam off your subject.

Outputs: view (IMAGE), exact_view (STRING - the exact view descriptor), focal_mm (FLOAT), all_views (IMAGE batch), and a report. The README's wiring hint: view → AtlasAddPatchView.patch_image, exact_view → its exact_view_override.

The depth gotcha (read this one)

The node's own source carries a warning worth taking seriously: use MoGe for the depth, not V2-Metric-Outdoor. A 90° square crop is framing the DA2 V2-Metric-Outdoor model wasn't trained on, and it mis-scales badly - measured on two 8K panoramas, V2-Outdoor returned camera heights 4.3–4.6× off where MoGe-2 returned plausible tripod heights. FOV stays exact either way; it's purely a metric-scale failure. So if your pano solves with believable focal but absurd height, the depth backend is the suspect. The pack supports [moge] (MoGe-2) and [neural-da3] as alternatives to the default Depth Anything V2 specifically because of this.

Install

The split itself is core and dependency-free:

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

Restart, or install via ComfyUI Manager ("atlas-camera"). It's under Atlas/01 · Input & Camera. If you're taking panoramas seriously, also install the [moge] backend (GitHub-only package - use the --no-deps route into ComfyUI's venv that INSTALL.md documents, since its own dependency list pulls gradio/matplotlib).