👁 Website
👁 GitHub
👁 Hugging Face
👁 Follow on X

 _ _ _ _______ ___________ _ _ ___________ _ ______ 
| | | | \ | | _ \ ___| ___ \ | | || _ | ___ \ | | _ \
| | | | \| | | | | |__ | |_/ / | | || | | | |_/ / | | | | |
| | | | . ` | | | | __|| /| |/\| || | | | /| | | | | |
| |_| | |\ | |/ /| |___| |\ \\ /\ /\ \_/ / |\ \| |___| |/ / 
 \___/\_| \_/___/ \____/\_| \_|\/ \/ \___/\_| \_\_____/___/ 
 

UNDERWORLD Dataset v1

This dataset was generated by the UNDERWORLD app by webXOS

Download the UNDERWORLD app in the /underworld/ folder of the repo to create your own similar datasets.

Generated: 2026-01-26T19:15:51.223Z Frames: 226 FPS: 30 Resolution: 1280x562 Objects: 61 Triangles: 1497.3333333333333

Usage for Overworld-style training:

1. Convert PNG sequence to video: ffmpeg -i %05d.png -c:v libx264 clip.mp4
2. Use metadata.json for camera pose conditioning
3. Use controls.json for temporal conditioning
4. Train with diffusion models like Waypoint-1-Small

This dataset is optimized for low-end device simulation and micro-scale spatial studies.

Short dataset description (one-liner for card):

Synthetic 226-frame 1280×562 PNG sequence + metadata for training diffusion models in OVERWORLD-style environments with camera pose and temporal controls; optimized for low-end / micro-scale 3D simulation.

Use cases:

Fine-tuning small diffusion models (e.g. Waypoint-1-Small) on pose-conditioned video generation. Testing temporal + spatial consistency in low-resource 3D diffusion pipelines. Generating controllable micro-scale synthetic scenes for robotics / sim2real prototyping. Benchmarking lightweight video diffusion on constrained hardware.

Integration with OVERWORLD (huggingface.co/overworld):

UNDERWORLD_Dataset_v1 is the low-end / micro counterpart to OVERWORLD-style datasets on Hugging Face. It supplies short, lightweight, pose-controlled PNG sequences explicitly designed for training the same class of diffusion models (OVERWORLD-style = controllable 3D world generation via camera waypoints + temporal signals). Use it to pre-train or distill models before scaling to larger OVERWORLD-style data → enables faster iteration on small hardware while preserving compatible conditioning format.

Source

webXOS webxos.netlify.app github.com/webxos huggingface.co/webxos

License:

MIT