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 SpatialReward, a verifiable reward model explicitly designed to evaluate spatial layouts in generated images. SpatialReward adopts a multi-stage pipeline: a 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 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.
Figures & tables
Figure 1: Performance comparison of SD3.5-M [ 16 ] optimized via RL using SpatialReward versus Baseline Rewards.
Figure 2: Overall framework of our approach. (a) Standard Flow-GRPO [ 40 ] reinforcement learning pipeline for text-to-image generation. (b) The proposed SpatialReward, which parses prompts into structured spatial and attribute constraints, verifies them on generated images via expert detection, and uses vision–language chain‑of‑thought reasoning to produce the final reward score.
Figure 3: Overview of SpatRelBench, depicting benchmark tasks and their data distribution (a), the construction pipeline (b), and the evaluation methodology (c) designed to assess spatial relation understanding in text‑to‑image models.
Reward Model
GenEval (80-Obj)
SpatRelBench (1k-Obj)
S-Obj.
T-Obj.
Cnt.
Color
Pos.
Attr-C.
Overall
P-Text.
C-Text.
Cpx.
Ori.
3DRel.
Overall
Proprietary T2I models
GPT Image 1
0.99
0.92
0.85
0.92
0.75
0.61
0.84
0.50
0.22
0.53
0.15
0.45
0.37
Seedream 3.0
0.99
0.96
0.91
0.93
0.47
0.80
0.84
0.25
0.20
0.11
0.07
0.41
0.21
Qwen-Image
0.99
0.92
0.90
0.88
0.76
0.77
0.87
0.22
0.21
0.32
0.07
0.32
0.23
Based on Stable Diffusion 3.5
Table 1: Quantitative comparison of T2I generation models aligned with different reward models. Results are on GenEval(80-Obj) and SpatRelBench(1k-Obj), where parentheses indicate the number of object categories. S-Obj: Single object, T-Obj: Two objects, Cnt: Counting, Pos: Positions, Attr-C: Attribute (Color), P-Text: Position-Text OCR, C-Text: Counting-Text OCR, Cpx: Complex spatial relations, Ori: Orientation, 3DRel: 3D spatial relations, Overall: average score over all metrics in each dataset. Bold denotes the best score, and underline denotes the second best.
Figure 4: Qualitative comparison of generated image quality across different methods.
Model
Wise
DPG
Aesthetic
PickScore
SD3.5-M
0.45
83.96
5.39
22.34
+ SpatialReward
0.46
84.08
5.23
22.52
Flux1-dev
0.50
83.84
6.13
22.45
+ SpatialReward
0.52
84.19
6.15
23.22
Table 2: General-Purpose comparison on Wise, DPG, Aesthetic, and PickScore metrics.
Reward Model
Spearman ρ
Pearson r
Accuracy
CLIPScore
0.42
0.40
0.68
ImageReward
0.48
0.45
0.70
UnifiedReward
0.51
0.49
0.72
VisionReward
0.55
0.53
0.74
SpatialReward
0.63
0.61
0.79
Table 3: Correlation and accuracy with human spatial-consistency judgments; accuracy measured at threshold τ=0.8 .
Removed Component
GenEval
SpatRel.
T2IComp.
Full SpatialReward
95.2
37.1
50.1
– Exclusion Constraints
90.5
25.9
45.9
– Expert Detection
70.3
21.6
39.2
– CoT Reasoning
94.2
27.9
47.5
Table 4: Ablation results for SpatialReward. Scores (accuracy) are reported for GenEval, SpatRel: SpatialRelBench, and T2IComp: T2I-CompBench. T2I-CompBench values are averaged over its 2D and 3D spatial-consistency tasks.
Figure 5: Effect of CoT reasoning in spatial relations. CoT combines bounding boxes, orientation, and scene semantics, yielding correct classifications where detected-only matching fails.