Atlas Multi-View Burst Solve 📷🎞️

The sibling of AtlasMultiViewSolve - same registration engine, but instead of wiring images one at a time, you point this node at a folder of 2–16 frames and it recovers one deterministic camera rig from the whole burst. It's the 📷🎞️ node in Atlas/01 · Input & Camera, and it's the one you reach for after a handheld walk around the subject, phone in burst mode, or a drone orbit that gave you a dozen overlapping frames.

Where the two-frame node is for deliberate, curated pairs, this is for "I shot a pile of frames and I want the rig out of all of them." The registration is the same engine as the explicit multi-view node - every selected file goes through the same RAW/JPEG import and the same solver - so the outputs are comparable and the same rules about scale apply.

How it works

Beyond three frames the solver switches to an anchor-star pair topology: every frame must share overlap with the first file in the folder (sorted by name), which is the natural shape of a walking burst - frame 1 sees the scene, and each subsequent frame overlaps both the anchor and its neighbours. frame_stride thins a dense burst (every Nth frame), max_frames caps how many actually feed the solver, and half_size (default on) halves the demosaic so 36 MP bursts iterate fast.

The scale story is identical to the two-frame node: baseline_m, when set, is the measured distance between the first two selected frames - your one real metre stick that turns a normalized rig into a metric one. write_plates (default on) writes the developed frames to plates_dir as a side effect.

Inputs and outputs

Required: burst_dir - path to the folder of frames (default CameraRaw/burst).

Optional: frame_stride, max_frames (2–16), half_size, capture_mode (auto/translated/rotation_only), camera_height_m, match_quality (balanced/conservative/permissive/salvage), seed, learned_anchor_fallback, baseline_m, learned_scale_fallback, write_plates, plates_dir, pair_topology.

Outputs: solve, report, registration_json, match_overlays, and anchor_image - the anchor frame itself, handy for knowing which viewpoint the rig is anchored on.

Installing it

Core + vision, same as its sibling - a clone is enough:

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

Restart and it's in Atlas/01 · Input & Camera (ComfyUI Manager: "atlas-camera"). If your burst is RAW, add the [raw] extra so each frame gets the same developed decode the single-frame AtlasLoadRAW does.

Common issues

The anchor-star shape is the gotcha: every frame needs overlap with the first file in the folder, and "first" means sorted by filename. If your burst's first frame is a black frame or a shot of your feet, the whole topology is poisoned. Rename/sort so a good anchor frame sorts first, and keep the burst to frames that actually see the scene - sparse or disconnected shots just get rejected as unmatched.

And don't expect rotation-only bursts to measure scale. If you spun in place, baseline_m and camera_height_m are your only anchors, and the result is inferred, not measured. The whole reason to shoot a burst is translation; give it some.