Spatial Consistency in Text-to-Image Generation on SpatRelBench 1k-Obj
63P-TextFLUX1-dev + SpatialReward
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| FLUX1-dev + SpatialRewardBackbone=FLUX1-dev, Reward Model=SpatialReward2026.03 | 63 | 40 | 52 | 32 | 45 | 46 | |
| SD3.5-M + SpatialRewardBackbone=SD3.5-M, Reward Model=SpatialReward2026.03 | 51 | 33 | 43 | 26 | 55 | 42 | |
| GPT Image 1Type=Proprietary T2I model2026.03 | 50 | 22 | 53 | 15 | 45 | 37 | |
| FLUX1-devBackbone=FLUX1-dev2026.03 | 49 | 25 | 23 | 4 | 38 | 28 | |
| SD3.5-M + TextOCRBackbone=SD3.5-M, Reward Model=TextOCR2026.03 | 48 | 24 | 23 | 9 | 38 | 28 | |
| SD3.5-M + UnifiedRewardBackbone=SD3.5-M, Reward Model=UnifiedReward2026.03 | 46 | 26 | 40 | 12 | 40 | 33 | |
| SD3.5-M + ImageRewardBackbone=SD3.5-M, Reward Model=ImageReward2026.03 | 42 | 23 | 32 | 11 | 42 | 30 | |
| SD3.5-MBackbone=SD3.5-M2026.03 | 40 | 13 | 22 | 7 | 36 | 23 | |
| SD3.5-M + PickScoreBackbone=SD3.5-M, Reward Model=PickScore2026.03 | 36 | 15 | 28 | 6 | 33 | 24 | |
| SD3.5-M + QwenVLBackbone=SD3.5-M, Reward Model=QwenVL2026.03 | 32 | 21 | 30 | 12 | 37 | 26 | |
| Seedream 3.0Type=Proprietary T2I model2026.03 | 25 | 20 | 11 | 7 | 41 | 21 | |
| Qwen-ImageType=Proprietary T2I model2026.03 | 22 | 21 | 32 | 7 | 32 | 23 |