KL Divergence Evaluation on FLUX2-9B
170KL DivergenceFlow-DPPO
Evaluation results
| Method | Links | |
|---|---|---|
| Flow-DPPOSampling schedule=Flow-SDE, Evaluation setup=Single2026.06 | 170 | |
| Flow-CPSSampling schedule=CPS, Evaluation setup=Single2026.06 | 240 | |
| Flow-DPPOSampling schedule=Flow-SDE, Evaluation setup=Multi2026.06 | 490 | |
| Flow-DPPOSampling schedule=Flow-SDE, Evaluation setup=+CFG2026.06 | 510 | |
| Flow-DPPO + CPSSampling schedule=CPS, Evaluation setup=Single2026.06 | 680 | |
| Flow-DPPO + CPSSampling schedule=CPS, Evaluation setup=Multi2026.06 | 700 | |
| Flow-GRPOSampling schedule=Flow-SDE, Evaluation setup=Single2026.06 | 770 | |
| Flow-GRPOSampling schedule=Flow-SDE, Evaluation setup=Multi2026.06 | 790 | |
| Flow-DPPO + CPSSampling schedule=CPS, Evaluation setup=+CFG2026.06 | 830 | |
| GRPO-GuardSampling schedule=Flow-SDE, Evaluation setup=Multi2026.06 | 1,010 | |
| GRPO-GuardSampling schedule=Flow-SDE, Evaluation setup=Single2026.06 | 1,070 | |
| Flow-GRPOSampling schedule=Flow-SDE, Evaluation setup=+CFG2026.06 | 1,360 | |
| Flow-CPSSampling schedule=CPS, Evaluation setup=+CFG2026.06 | 1,510 | |
| GRPO-GuardSampling schedule=Flow-SDE, Evaluation setup=+CFG2026.06 | 1,630 | |
| Flow-CPSSampling schedule=CPS, Evaluation setup=Multi2026.06 | 1,660 |