Text-to-Image on SD3.5-Medium synthetic data
84.6OCR AccuracyDanceGRPO
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| DanceGRPObase_model=SD3.5-Medium, post-training=RL2025.12 | 84.6 | -53 | 0.309 | 0.837 | 0.8 | |
| FlowGRPObase_model=SD3.5-Medium, post-training=RL2025.12 | 84.5 | -22.1 | 0.313 | 0.855 | 1.03 | |
| DDRLbase_model=SD3.5-Medium, post-training=RL with data regularization, fine-tuning=LoRA (alpha=64, r=32), regularization_data=synthetic2025.12 | 82.3 | 0 | 0.32 | 0.865 | 1.14 | |
| w/o RLbase_model=SD3.5-Medium2025.12 | 56.6 | — | 0.321 | 0.865 | 1.13 |