ThinkBLOX: 3D Indoor Scene Generation with Progressive Reasoning

ThinkBLOX tackles a familiar problem in procedural scene generation: once a room layout is planned in a single pass, even small edits can force a full rebuild or leave the scene in a bad state. The new framework treats indoor scene creation as a step-by-step reasoning-and-action loop, where the model conditions each move on the current scene state rather than committing everything up front.

To train that behavior, the team built ThinkBLOX-Data-200K, a dataset with 224,757 procedural placement pairs. Each pair includes multi-view scene context, explicit chain-of-thought rationales, and structured JSON layouts, giving the model both visual grounding and a trace of why a placement was chosen. That matters for developers because the hard part here is not just generating plausible rooms, but making incremental edits behave sensibly when objects are inserted, moved, or rearranged.

The training stack also includes Tier-Decoupled GDPO, a reinforcement learning approach that separates heterogeneous rewards into tiers. In practice, that means physical validity, semantic plausibility, and reasoning-action consistency are optimized more stably instead of fighting each other in a single reward bucket. For teams building tools, editors, or content pipelines, that kind of separation is especially relevant because scene synthesis is rarely a single-objective problem.

The claimed payoff is better physical plausibility, stronger semantic alignment, and more editable scenes than recent one-shot and iterative baselines. The system is positioned for both global generation and local...