Not a screenshot gallery. Drag inside any sphere — this is the actual training data, rendered equirectangular and explorable. 12 curated worlds, 6 NASA public-domain 360 video clips, every caption engineered to teach a diffusion model the projection format itself.
Most 360 datasets are YouTube rips or gated robot rosbags with no license. This one is clean, captioned in a uniform grammar, and built so the format is learnable — not just the subject.
Every caption opens with the trigger token, then the identical scaffold:
w0rldw3aver360, a 360-degree equirectangular photograph of…. 56–72 words, 12/12 files. The repetition is the point.
Captions describe the wrap: horizon circling the frame, seamless seam, zenith and nadir. Rendered cold, the base model produced true 2:1 equirect from caption alone — scored 9/10.
Every still is CC0 from Poly Haven. Every video is US Government public domain (NASA/JPL/MSFC). Curation and captions released CC0. No gray area to litigate later.
A pairs/ folder ships in the repo — .jpg + identically-named .txt side by side. Drag the whole folder into AI-Toolkit and train. No renaming script needed.
15-second 2:1 clips at 30fps are the right shape for Gaussian splat training — enough parallax to reconstruct geometry, short enough to stay coherent.
The real pipeline: train the LoRA on these 12, then generate hundreds of new worlds with Wan I2V and feed those back in. The dataset bootstraps itself.
NASA public-domain equirectangular video, cut to 15-second windows and cropped to exact 2:1. Flat renders and partial mosaics were QC-rejected.
Proof that chained generation holds a story together. Five MiniMax H3 clips, each starting from the previous clip's final frame. One still panorama → a 26-second first-person short film with synchronized audio.
End to end, this is what the studio runs. Each stage is a real artifact that exists right now.
Poly Haven CC0 stills; NASA 360 video pulled from the public images API. Rejected: Curiosity 8K (partial Navcam mosaic with burned-in text) and Hurricane Maria (flat data-viz, not a panorama).
ffprobe every candidate for true 2:1 equirect. Crop letterboxed sources to exact ratio — Perseverance needed crop=1752:876:72:40 to kill the zenith black bar.
Trigger token first, uniform scaffold, 56–72 words describing subject, horizon, lighting, and wrap behavior. Captions are CC0 too.
Published to HuggingFace with pairs/ ready for drag-and-drop, README with YAML front-matter, and a montage preview.
AI-Toolkit config targeting Krea 2 Turbo: rank 32, 1024×512 for panos, 3000 steps, low_vram + cache_text_embeddings for 12GB cards. No trigger override — captions carry it.
Render new worlds with the trained LoRA, motionize with Wan I2V, feed the best back into the dataset. Synthetic data flywheel.
Datasets, LoRAs, sites, and releases — built by an agent swarm on a home GPU and a $20 runway. This page rendered itself from the same repo it documents.