Open-source SplatKit World turns one photo or sentence into a walkable Gaussian-splat 3D world in ComfyUI

mickmumpitz · x · 2026-11-03

Creator mickmumpitz released ComfyUI-SplatKit-World (MIT license) for the #OpenSourceHiggsfield contest: one photo or one sentence becomes an explorable Gaussian-splat 3D world in your browser, running fully locally in ComfyUI with open weights only — no cloud APIs. Pipeline: photo/prompt → 360° pano → MoGe depth → hand-drawn drone paths → WAN fly-through videos → SphereSfM → splat training. Since the model can't see behind the camera, the drone-path fly-throughs are the trick that fills in unseen geometry. Ships with example workflows, install docs, and full training code for reproduction.

Original post →

More from Multimodal

Multimodal channel →