ThinkBLOX: 3D Indoor Scene Generation with Progressive Reasoning

2026-07-15

AI Summary

Existing vision-language model–based approaches to 3D indoor scene generation typically rely on one-shot layout planning, which hinders efficient interactive editing and often yields layouts that are physically implausible or semantically inconsistent. This work proposes a progressive reasoning framework that formulates layout generation as a state-conditioned sequential decision-making process. It introduces a novel generative paradigm integrating explicit chain-of-thought reasoning with structured layout representations and devises a hierarchical, decoupled GDPO reinforcement learning algorithm to reconcile conflicting multi-objective rewards. Trained on a newly curated ThinkBLOX-Data-200K dataset—comprising 220,000 multi-view contextual examples paired with chain-of-thought rationales and JSON-encoded layouts—and leveraging supervised fine-tuning alongside a tier-decoupled GDPO strategy, the proposed method significantly outperforms existing baselines in physical plausibility, semantic alignment, and interactive editability, enabling flexible global and local scene generation and rearrangement.

Abstract

While traditional graphics methods often synthesize 3D indoor scenes autoregressively or hierarchically, recent vision-language model (VLM)-based generators predominantly adopt a one-shot paradigm where the full layout is planned at once. This one-shot approach often requires global re-optimization or complete reconstruction during interactive editing (e.g., inserting or moving objects) and can lead to physically or semantically poorly organized arrangements. To address these challenges, we propose ThinkBLOX, a VLM-based progressive reasoning framework that iteratively designs and refines 3D scenes. ThinkBLOX treats layout generation as a state-conditioned, step-by-step reasoningand-action process. To power this, we construct the ThinkBLOX-Data-200K dataset, containing 224,757 procedural placement pairs annotated with multi-view scene context, explicit Chain-of-Thought (CoT) rationales, and structured JSON layouts. Through supervised fine-tuning (SFT) on this dataset, the VLM learns to bridge the reasoning-action gap under incremental updates. Furthermore, recognizing that scene synthesis is inherently a multisolution task where SFT suffers from reward conflict, we introduce Tier-Decoupled GDPO. This reinforcement learning scheme organizes heterogeneous rewards into distinct tiers, stabilizing policy optimization across physical validity, semantic plausibility, and reasoning-action consistency. Extensive experiments show that ThinkBLOX significantly outperforms recent one-shot and iterative baselines in physical plausibility, semantic alignment, and interactive editability. Additionally, we show that it supports diverse applications, including both global and local generation and rearrangement of 3D scenes.

Problem

• 3D indoor scene generation
• one-shot layout planning
• interactive editing
• physical plausibility
• semantic organization

Innovation

• progressive reasoning
• vision-language model
• Chain-of-Thought
• Tier-Decoupled GDPO
• interactive 3D scene generation