Paper 2602.24233
Enhancing Spatial Understanding in Image Generation via Reward Modeling
- Published
- Feb 2026
- Research lab
- Independent
- Citations
- 0
- GitHub
- 86 stars
01 In brief
Summary
The paper introduces a method to improve spatial understanding in text-to-image generation using reward modeling.
The authors construct the SpatialReward-Dataset, containing over 80,000 adversarial preference pairs, where each pair consists of an image correctly depicting complex spatial relationships and a perturbed image violating some relationships, generated by models like Qwen-Image, HunyuanImage-2.1, and Seedream 4.0, and verified by human experts.
They then train SpatialScore, a reward model based on Qwen2.5-VL-7B with LoRA, which evaluates spatial relationship accuracy and surpasses proprietary models like GPT-5 and Gemini-2.5 Pro on a 365-pair benchmark, achieving 95.8% pairwise accuracy.
SpatialScore is used as the reward signal for online reinforcement learning (GRPO) to fine-tune FLUX.1-dev, with a top-k filtering strategy to mitigate advantage bias from varying prompt difficulty.
Experiments show significant improvements on in-domain and out-of-domain benchmarks (e.g., DPG-Bench, TIIF-Bench, UniGenBench++), outperforming a GenEval-based baseline, which fails on long prompts.
The method also improves performance on the GenEval benchmark and generalizes to Qwen-Image, demonstrating consistent gains in spatial understanding.
02 From the paper
Abstract
Recent progress in text-to-image generation has greatly advanced visual fidelity and creativity, but it has also imposed higher demands on prompt complexity-particularly in encoding intricate spatial relationships. In such cases, achieving satisfactory results often requires multiple sampling attempts. To address this challenge, we introduce a novel method that strengthens the spatial understanding of current image generation models. We first construct the SpatialReward-Dataset with over 80k preference pairs. Building on this dataset, we build SpatialScore, a reward model designed to evaluate the accuracy of spatial relationships in text-to-image generation, achieving performance that even surpasses leading proprietary models on spatial evaluation. We further demonstrate that this reward model effectively enables online reinforcement learning for the complex spatial generation. Extensive experiments across multiple benchmarks show that our specialized reward model yields significant and consistent gains in spatial understanding for image generation.