Text-to-Image Generation on T2I-CompBench (ImageReward, HPSv2, PickScore)
0.9377ImageRewardTPM (2 × r*)
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| TPM (2 × r*)Base Model=SD3.5-Medium, Resolution=512×512, Precision=FP16, Inference Steps=15, Reward Scaling=2 × r*2025.12 | 0.9377 | 0.2729 | 0.2757 | 22.1 | 0.106 | |
| TPM (0.5 + r*)Base Model=SD3.5-Medium, Resolution=512×512, Precision=FP16, Inference Steps=15, Reward Scaling=0.5 + r*2025.12 | 0.8969 | 0.2722 | 0.2754 | 22.11 | 0.231 | |
| TPM (0.5 + r*)Base Model=Flux.1-dev, Resolution=512×512, Precision=FP16, Inference Steps=10, Reward Scaling=0.5 + r*2025.12 | 0.8725 | 0.2854 | 0.2673 | 22.57 | 0.499 | |
| SD BaseBase Model=SD3.5-Medium, Resolution=512×512, Precision=FP16, Inference Steps=152025.12 | 0.8642 | 0.2698 | 0.2754 | 22.09 | 0 | |
| TPM (2 × r*)Base Model=Flux.1-dev, Resolution=512×512, Precision=FP16, Inference Steps=10, Reward Scaling=2 × r*2025.12 | 0.8584 | 0.2866 | 0.2667 | 22.52 | 0.64 | |
| Flux BaseBase Model=Flux.1-dev, Resolution=512×512, Precision=FP16, Inference Steps=102025.12 | 0.7197 | 0.2747 | 0.2655 | 22.46 | 0 |