Map2World — Segment Map Conditioned Text-to-3D World Generation

Source: HuggingFace Daily Papers Raw: raw/huggingface/2026-05-04-map2world-segment-map-conditioned-text-to-3d-world-generation.md arXiv: https://arxiv.org/abs/2605.00781 Date: 2026-05-04 Tier: 4 — 3D / spatial reconstruction

TL;DR

Generates 3D worlds from user-defined segment maps of arbitrary shape and scale, contrasting with prior grid-layout-constrained methods. A detail-enhancer network adds fine-grained details without breaking global coherence. Improves user-controllability, scale consistency, and content coherence over baselines.

Quick note for cere-bro

Tier 4 (3D generation, not directly relevant to routing/efficiency). Worth noting only as a data point in the spatial-intelligence-camp tracking (Ken Huang World Models, 05-03) — World Labs / Marble line of work; segment-map conditioning is a useful primitive if the camp's bet pays off.