Atlas Input 🎬

AtlasInput is the front door to the whole atlas-camera pack, and it's the one node most people should start with and never leave. Wire a photograph into it and you get a metric pinhole camera solve, projection geometry, a depth map and a report - the entire "photo → 2.5D projection scene" chain collapsed into a single node with sensible defaults. If you've ever tried to fake a dolly move on a still and ended up with smeared texture instead of parallax, this is the tool that makes the difference explicit: it recovers a real camera first, then projects the photo onto derived geometry from that camera's viewpoint. From where the camera actually stood, the plate reassembles exactly; scale error only shows up as parallax on a move, never as stretched pixels.

How it works

One photograph in. The solve recovers focal length, orientation and horizon - either from a deterministic geometric solve (vanishing points, straight lines) or, for harder frames, from a learned single-image prior called GeoCalib. You've probably already met GeoCalib's results without knowing the name: it's the same estimator people use to pull camera intrinsics off a video frame for 360° panorama projection. Monocular depth (Depth Anything V2 Metric-Outdoor by default) then turns that camera into a relief mesh, and the plate gets cast back through it.

Under the hood the node doesn't actually do all that in one Python function - it expands into a real mini-graph of the pack's inner nodes at execution, so every step keeps its own cache and missing optional packs degrade to a note in the report output instead of killing your queue. That's worth knowing when something's slow: the pieces are all individually tunable if you later decide AtlasInput isn't enough.

The inputs that matter

The required input is just image - any tensor ComfyUI can produce. Almost everything else has a default that's right for a first run:

• layers (default 0) - the single biggest knob. 0 gives one full-range relief mesh, instant. Set it to 2–4 and you get depth-band clean-plate layers (the splits are proven: 2→0.55, 3→0.2/0.65, 4→0.3/0.6/0.8) that are watertight by construction. That's the 2.5D matte-painting stack.
• mesh_resolution (default 512) - relief grid cells along the long edge. 512 is a fine default for a 4K plate; drop it for quick previews.
• sky (default off) - SAM-segments the sky onto its own flat card and feeds the mask into every mesh. Big for outdoor plates.
• inpaint (default off) - builds each occluded band's clean plate (occlusion mask → expand → crop → LaMa → stitch). Needs comfyui-inpaint-nodes + big-lama.pt; if they're absent it's skipped and noted, not fatal.
• use_vlm (default off) - an advisory VLM pass that wires segmentation prompts and per-band geometry into the inner nodes. This is where the pack gets clever, but it's strictly optional and only if you have a local VLM server.

The outputs are solve (the recovered camera - feed this to export or viewport nodes), image (the processed plate), depth, sky_mask, and report (a plain-text run-down of what actually happened).

Install

Install the pack once and every Atlas node comes with it - search atlas-camera in ComfyUI Manager, or:

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

Restart ComfyUI. The core solve is pure Python with no dependencies; the learned depth backend needs transformers (ComfyUI already ships torch), and GeoCalib installs separately if you want the neural solve for hard frames. A GPU only matters for the neural parts - the deterministic solve runs on a potato.

Troubleshooting

• First queue is slow. It's loading the depth model for the first time; that's a one-off.
• Relief mesh looks "combed" or shredded - that's the world-space edge tear threshold (max_edge_factor, default 12) being tripped by grazing walls or receding floors. Raise it to 40–80; the tooltip is unusually honest about this.
• Interior plates punch big holes in walls/ceilings. sky_heuristic (default on) is eating far regions as sky. Turn it off indoors.
• A note in report about a skipped stage (LaMa, SAM3) means an optional extra isn't installed - check the README's dependency tiers before blaming the node.

AtlasInput is the pack's whole thesis in one node. Outgrow it and the rest of the Atlas menu takes over individual stages - but most people won't need to.