SpatialReward: Verifiable Spatial Reward Modeling for Fine-Grained Spatial Consistency in Text-to-Image Generation

2026-03-23

AI Summary

This work addresses the challenge that existing reward models for text-to-image generation struggle to accurately evaluate fine-grained spatial relationships, often resulting in distorted object layouts. To overcome this limitation, the authors propose SpatialReward, a verifiable spatial reward model that leverages a chain-of-reasoning mechanism composed of a Prompt Decomposer, expert detectors, and a vision-language model to precisely capture complex spatial relations—including directional cues, object attributes, and textual layout configurations. Additionally, they introduce SpatRelBench, the first benchmark specifically designed for evaluating fine-grained spatial layout fidelity. Experiments on Stable Diffusion and FLUX demonstrate that integrating SpatialReward significantly enhances both spatial consistency and overall image quality, with results aligning more closely with human judgments.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Recent advances in text-to-image (T2I) generation via reinforcement learning (RL) have benefited from reward models that assess semantic alignment and visual quality. However, most existing reward models pay limited attention to fine-grained spatial relationships, often producing images that appear plausible overall yet contain inaccuracies in object positioning. In this work, we present \textbf{SpatialReward}, a verifiable reward model explicitly designed to evaluate spatial layouts in generated images. SpatialReward adopts a multi-stage pipeline: a \emph{Prompt Decomposer} extracts entities, attributes, and spatial metadata from free-form prompts; expert detectors provide accurate visual grounding of object positions and attributes; and a vision-language model applies chain-of-thought reasoning over grounded observations to assess complex spatial relations that are challenging for rule-based methods. To more comprehensively evaluate spatial relationships in generated images, we introduce \textbf{SpatRelBench}, a benchmark covering object attributes, orientation, inter-object relations, and rendered text placement. Experiments on Stable Diffusion and FLUX show that incorporating SpatialReward into RL training consistently improves spatial consistency and overall generation quality, with results aligned more closely to human judgments. These findings indicate that verifiable reward models hold considerable potential for enabling more accurate and controllable optimization in text-to-image generation models.

Problem

• spatial consistency
• text-to-image generation
• reward modeling
• spatial relationships
• fine-grained alignment

Innovation

• SpatialReward
• spatial consistency
• reward modeling
• text-to-image generation
• verifiable reasoning