Text-to-Image Generation on DrawBench (UnifiedReward)
3.06UnifiedRewardB2: Uniform
Evaluation results
| Method | Links | |
|---|---|---|
| B2: UniformFusion Strategy=Uniform normalized averaging, Evaluation Model=UnifiedReward-2.0, Sampling steps=50, CFG scale=6.0, Scheduler=FlowMatch Euler2026.02 | 3.06 | |
| S2: DepthFusion Strategy=Depth-wise fusion, Evaluation Model=UnifiedReward-2.0, Sampling steps=50, CFG scale=6.0, Scheduler=FlowMatch Euler2026.02 | 3.06 | |
| S3: JointFusion Strategy=Joint fusion, Evaluation Model=UnifiedReward-2.0, Sampling steps=50, CFG scale=6.0, Scheduler=FlowMatch Euler2026.02 | 3.06 | |
| B3: StaticFusion Strategy=Learnable static fusion, Evaluation Model=UnifiedReward-2.0, Sampling steps=50, CFG scale=6.0, Scheduler=FlowMatch Euler2026.02 | 3.05 | |
| FuseDiTFusion Strategy=Deep-fusion baseline, Evaluation Model=UnifiedReward-2.0, Sampling steps=50, CFG scale=6.0, Scheduler=FlowMatch Euler2026.02 | 3.05 | |
| B1: Penult.Fusion Strategy=Penultimate-layer, Evaluation Model=UnifiedReward-2.0, Sampling steps=50, CFG scale=6.0, Scheduler=FlowMatch Euler2026.02 | 3.02 | |
| S1: TimeFusion Strategy=Time-wise fusion, Evaluation Model=UnifiedReward-2.0, Sampling steps=50, CFG scale=6.0, Scheduler=FlowMatch Euler2026.02 | 2.97 |