[dependencies]
map-anything = "^0.4.4"

map-anything

Metric 3-D reconstruction from plain images, in pure NURL.

This is a port of Meta's MapAnything — reference implementation and paper at facebookresearch/map-anything — running the Apache-licensed facebook/map-anything-apache checkpoint. Given a set of photos or a video, it predicts for every view: ray directions, depth along ray, camera pose (quaternion + translation) and a global metric scale, and fuses them into a metric world-space point cloud. The cloud is written as PLY and can be orbited in the browser through the ply package's built-in WebGL viewer.

Usage

map-anything photos/            # folder of images -> cloud.ply
map-anything walk.mp4           # video (MJPEG AVI in pure NURL, rest via ffmpeg)
map-anything photos/ --view     # open the browser viewer when done
map-anything hill.mp4 --mask-sky   # outdoor: drop sky via skyseg.onnx
map-anything walk/ --max-views 400 # long capture: windowed + Sim3-stitched

Up to --window views (default 24) the whole set runs through the model's global attention at once. Longer captures are reconstructed in overlapping windows and stitched with a closed-form Sim(3) fitted on the overlap views' pixels — exact correspondences, no matching step.

The checkpoint (~4.5 GB, F32 safetensors) is fetched from Hugging Face on first use via the hub package and cached under ~/.nurl/models.

Architecture (from the reference)

• encoder — DINOv2-giant, 24 layers, dim 1536, patch 14
• info_sharing — 16-layer alternating attention (frame-wise / global) over all

views, 24 heads, SwiGLU MLP

• dense_head — DPT head predicting ray directions + depth + confidence + mask
• pose_head — per-view quaternion + translation
• scale_head + scale_token — global metric scale
• optional geometric-input encoders (known intrinsics / depth / poses) exist in

the checkpoint; image-only inference uses none of them

Every ported stage is verified against the reference PyTorch modules by importing and driving them directly, not by re-implementing the math in a second oracle.

License

Apache-2.0, matching the upstream reference implementation and checkpoint.