Text-to-Image Generation on T2I-CompBench (VQA/Detection/CoT Evaluation)
75B-VQA ScorePPCL(14B)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| PPCL(14B)Model Context=Qwen-Image, P.(B)↓=14, M.(%)↓=79.4, L.(ms)↓=17922025.11 | 75 | 52.4 | 82.15 | 0.42 | |
| PPCL(12B)Model Context=Qwen-Image, P.(B)↓=12, M.(%)↓=71.4, L.(ms)↓=15142025.11 | 73.3 | 52.9 | 81.94 | 3.03 | |
| PPCL(Fine-tuning)Model Context=Qwen-Image, P.(B)↓=10, M.(%)↓=66.9, L.(ms)↓=14622025.11 | 71.5 | 51.5 | 81.33 | 3.29 | |
| Base modelModel Context=Qwen-Image, P.(B)↓=20, M.(%)↓=100, L.(ms)↓=26252025.11 | 70.9 | 53.2 | 82.47 | 0 | |
| HierarchicalPruneModel Context=Qwen-Image, P.(B)↓=14, M.(%)↓=79.4, L.(ms)↓=17862025.11 | 70.6 | 48.7 | 79.94 | 6.49 | |
| PPCL(10B)Model Context=Qwen-Image, P.(B)↓=10, M.(%)↓=66.9, L.(ms)↓=14622025.11 | 70.1 | 49.9 | 79.24 | 6.88 | |
| TinyFusionModel Context=Qwen-Image, P.(B)↓=14, M.(%)↓=79.4, L.(ms)↓=17892025.11 | 68.9 | 46.4 | 78.99 | 8.75 | |
| Base modelModel Context=FLUX.1-dev, P.(B)↓=12, M.(%)↓=100, L.(ms)↓=7152025.11 | 64 | 42.6 | 78.57 | 0 | |
| Chroma1-HDModel Context=FLUX.1-dev, P.(B)↓=8.9, M.(%)↓=82.5, L.(ms)↓=17142025.11 | 62.1 | 33.9 | 76.43 | 1.02 | |
| PPCL(8B)Model Context=FLUX.1-dev, P.(B)↓=8, M.(%)↓=74.4, L.(ms)↓=5352025.11 | 61.5 | 39.1 | 78.15 | 4.03 | |
| TinyFusionModel Context=FLUX.1-dev, P.(B)↓=8, M.(%)↓=74.4, L.(ms)↓=5342025.11 | 58.4 | 36.9 | 74.17 | 13.8 | |
| PPCL(6.5B)Model Context=FLUX.1 Lite, P.(B)↓=6.5, M.(%)↓=69.2, L.(ms)↓=4282025.11 | 58.1 | 39.8 | 76.91 | 0.07 | |
| HierarchicalPruneModel Context=FLUX.1-dev, P.(B)↓=8, M.(%)↓=74.4, L.(ms)↓=5432025.11 | 57.9 | 37.1 | 74.99 | 13.38 | |
| FLUX.1 LiteModel Context=FLUX.1-dev, P.(B)↓=8, M.(%)↓=78.8, L.(ms)↓=5722025.11 | 54.7 | 37.9 | 77.21 | 6.09 | |
| Base modelModel Context=FLUX.1 Lite, P.(B)↓=8, M.(%)↓=78.8, L.(ms)↓=5722025.11 | 54.7 | 37.9 | 77.21 | 0 | |
| Dense2MoEModel Context=FLUX.1-dev, P.(B)↓=12, M.(%)↓=100, L.(ms)↓=3122025.11 | 47.3 | 31.1 | 76.25 | 21.52 |