One phone video, a whole orbiting camera crew

What it does

Feed it one portrait video of a person—shot on a phone, mild camera motion, at least 121 frames—and it outputs a battery of synthetic viewpoints arranged in orbits or frontal arcs. Those rendered views are meant to be dropped straight into a 4D Gaussian Splatting reconstruction workflow (the repo points to nerfstudio). Think of it as a virtual cinematography crew that retroactively surrounds your subject.

The interesting bit

Instead of training an end-to-end 4D model, the system decouples the problem: it recovers motion via an embedded GVHMR submodule, then synthesizes target views with configurable camera geometry—pitch layers, yaw span, view density. You decide whether you want a sparse six-view ring or a dense 48-view three-tier rig.

Key highlights

• Accepts a single 720p+ monocular video (9:16 portrait) and expands it into N synchronized virtual camera feeds.
• Camera layouts are fully parameterized: views_per_layer, layer_pitches, start_yaw, and yaw_span.
• Outputs include per-view MP4s, a cameras.json rig description, and reusable motion-capture intermediates.
• Optional speed-ups via FlashAttention-3 or SageAttention.
• Models and sample data download automatically on first run.

Caveats

• The todo list explicitly flags that low-memory inference below 32 GB is still pending, so current hardware requirements appear steep.
• The actual 4DGS reconstruction step is delegated to external tooling (the README links to a nerfstudio guide rather than bundling its own).
• Input constraints are picky: one person, mild camera motion, portrait aspect ratio, minimum frame count.

Verdict

Worth a look if you’re building free-viewpoint human avatars and already have a 4DGS pipeline. Skip it if you need a lightweight, end-to-end reconstruction app or lack the GPU headroom.