w0rldw3aver360-motionSynthetic 360° First-Person Walkthroughs
Licensed first-person 360 walkthrough footage is nearly impossible to ship. Ours is built the other way: CC0 panoramas in, first-person walkthrough videos out — every clip generated, captioned, seam-measured, and released with a clean license for Gaussian splatting and camera-motion research.
Why this dataset exists
Training Gaussian splats or camera-control video models needs first-person walkthrough footage with a clean license. We surveyed the field and found the shelf nearly empty:
| Source | What | License | Verdict |
|---|---|---|---|
| 360-1M | 1M+ crowd 360 videos, link index | unclear at source | — not shippable |
| GGPS (Insta360) | 3 outdoor scenes, ERP + SfM + splats (reference study) | CC BY-NC 4.0 | △ research reference only — excluded from our shipped data |
| Walking Tours | 10 urban 4K60 egocentric walks | CC-BY · attribution | △ urban, attribution |
| NASA 360 | rocket stands, Mars, hurricanes | US Gov public domain | ✓ static camera |
| CC0 nature POV walks | — | — | — open gap |
So we generate the missing shelf ourselves. The source panoramas are CC0 (Poly Haven). The generation engine is our own pipeline. Every output is released CC0. The gap becomes the product.
The method
One panorama becomes a walk in three steps. No 3D scene, no game engine, no camera rig.
Perspective extract
v360 projects the 2:1 equirectangular pano into a natural 100° FOV head-height frame — the view a chest-mounted action-cam would see standing in that spot.
The video engine imagines the walk
A distilled video diffusion engine (8-step turbo) animates the frame: forward dolly, footstep bob, parallax at the edges — plus synchronized footsteps and ambience in the same pass.
Chain & measure
Each clip's last frame seeds the next clip — 5.17s at a time, indefinitely. Every seam is PSNR-measured; only >30 dB ships.
Every clip is provenance
Each sample ships with its source pano, exact caption, motion prompt, generation params, and seam scores. A walk_prompts.csv ties the whole catalog together.
Geometry-aware starts
Start frames aren't crops — they're true spherical projections from the pano, so the walk begins in the world the panorama describes, at the yaw we choose.
CC0 in, CC0 out
Poly Haven stills are CC0. Our generations and captions are CC0. The only encumbrance anywhere in the chain is the GGPS study set, which stays in research dirs.
Built for splats & motion LoRAs
Chained walks give multi-view continuity for splat training; the caption+prompt pairs are shaped for camera-motion LoRA fine-tunes on modern video models.
The catalog
Walks generated so far. Every clip below is real and streams from this page. Green = chained multi-clip walk. The queue refills as the night worker renders.
Format & access
Ships like every Sonic Forage dataset — drag-and-drop ready, no scripts needed.
Layout
walks/<slug>/walk_NN.mp4 per world · starts/<slug>_start.jpg perspective seeds · walk_prompts.csv (slug, clip, caption, motion_prompt) · per-clip walk_meta.json with seam scores.
Companion sets
Pairs with w0rldw3aver360 (12 still panos + caption pairs) and w0rldw3aver360-motion (NASA public-domain 360 video, 12 clips) — stills, real motion, synthetic motion, one grammar.
Splat path
Chained walks → frame extraction → COLMAP spherical SfM → Gaussian splat. The GGPS study (CC BY-NC) documents the method; our CC0 walks make it shippable.
The walk is the dataset
12 worlds · endless walks · one caption grammar · zero license landmines. Generated by the Sonic Forage agent swarm on Comfy Cloud.