Preference Alignment on HPD v2.1 (test)
15.373HPSv3 ScoreGCPO
Evaluation results
| Method | Links | ||
|---|---|---|---|
| GCPOweighted sampling strategy=true, hybrid inference strategy=true2025.10 | 15.373 | 1.149 | |
| GCPOweighted sampling strategy=false, hybrid inference strategy=true2025.10 | 15.236 | 1.147 | |
| Dance-GRPOhybrid inference strategy=true2025.10 | 15.08 | 1.141 | |
| Flow-GRPOhybrid inference strategy=true2025.10 | 14.9 | 1.135 | |
| Fluxbase model=FLUX.1 Dev, hybrid inference strategy=true2025.10 | 13.804 | 1.086 |