3D-Layout-R1: Structured Reasoning for Language-Guided Spatial Layout Editing

Forum topic · 小凯 · 2026-03-25

Summary

3D-Layout-R1 is a structured reasoning framework for text-conditioned spatial layout editing via scene-graph reasoning, proposed by researchers including Haoyu Zhen and Chuang Gan (arXiv 2603.22279). While LLMs and VLMs show strong reasoning abilities, they struggle with spatial understanding and layout consistency in fine-grained visual editing. Given an input scene graph and a natural-language instruction, the model reasons over the graph to produce an updated scene graph that satisfies the text condition while preserving spatial coherence. Explicit structured relational representations improve interpretability and control over spatial relationships. On a new text-guided layout editing benchmark covering ordering, spatial alignment, and room editing tasks, the training paradigm outperforms chain-of-thought SFT (CoT-SFT) and vanilla GRPO baselines, achieving an average 15% IoU improvement and 25% reduction in center distance error.

Paper Overview

Field: Computer Vision (CV) Authors: Haoyu Zhen, Xiaolong Li, Yilin Zhao, Han Zhang, Sifei Liu, Kaichun Mo, Chuang Gan, Subhashree Radhakrishnan Published: 2026-03-23 arXiv: 2603.22279

Abstract

Large Language Models (LLMs) and Vision Language Models (VLMs) have shown impressive reasoning abilities, yet they struggle with spatial understanding and layout consistency when performing fine-grained visual editing.

The authors introduce a Structured Reasoning framework that performs text-conditioned spatial layout editing via scene-graph reasoning. Given an input scene graph and a natural-language instruction, the model reasons over the graph to generate an updated scene graph that satisfies the text condition while maintaining spatial coherence. By explicitly guiding the reasoning process through structured relational representations, the approach improves both interpretability and control over spatial relationships.

Evaluation

The method is evaluated on a new text-guided layout editing benchmark covering:

Ordering tasks
Spatial alignment tasks
Room editing tasks

Compared with Chain-of-Thought supervised fine-tuning (CoT-SFT) and vanilla GRPO baselines, the proposed training paradigm achieves:

~15% average improvement in IoU
~25% reduction in center distance error