Text-to-image generation on MJHQ-30K (Category FID/CLIP Subset)
26.58FID (Plants)HACK++
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| HACK++Backbone=HART, ηa=30%, ηc=20%, KV Size=0.61 GB, Attn. TFLOPs=7.392026.06 | 26.58 | 27.34 | 27.75 | 27.46 | 20.08 | 26.4 | 25.13 | 28.03 | 7.3 | 27.83 | |
| HACKBackbone=HART, ηa=30%, ηc=30%, KV Size=0.84 GB, Attn. TFLOPs=7.402026.06 | 27.59 | 27.25 | 28.02 | 27.4 | 21.7 | 26.32 | 27.36 | 27.99 | 8.62 | 27.76 | |
| HACKBackbone=Infinity-8B, ηa=30%, ηc=30%, KV Size=5.06 GB, Attn. TFLOPs=33.542026.06 | 28.72 | 28.01 | 28.71 | 28.42 | 23.28 | 27.43 | 29.94 | 29.42 | 8.62 | 29.08 | |
| FullBackbone=Infinity-8B, ηa=100%, ηc=100%, KV Size=16.86 GB, Attn. TFLOPs=78.542026.06 | 28.81 | 28.12 | 28.61 | 28.45 | 24.58 | 27.32 | 30.26 | 29.42 | 8.75 | 29.1 | |
| HACK++Backbone=Infinity-8B, ηa=30%, ηc=10%, KV Size=1.69 GB, Attn. TFLOPs=31.532026.06 | 28.82 | 28.09 | 28.57 | 28.47 | 24 | 27.39 | 30.1 | 29.44 | 8.56 | 29.11 | |
| ScaleKVBackbone=Infinity-8B, ηc=10%, KV Size=1.69 GB, Attn. TFLOPs=41.982026.06 | 28.97 | 28.1 | 29.12 | 28.39 | 24.51 | 27.3 | 30.01 | 29.35 | 8.78 | 29.06 | |
| H2OBackbone=Infinity-8B, ηa=30%, ηc=30%, KV Size=5.06 GB, Attn. TFLOPs=33.552026.06 | 29.55 | 28.01 | 29.19 | 28.41 | 23.78 | 27.45 | 30.43 | 29.36 | 9.04 | 29.02 | |
| StreamingBackbone=Infinity-8B, ηa=30%, ηc=30%, KV Size=5.06 GB, Attn. TFLOPs=33.552026.06 | 29.7 | 28.02 | 28.48 | 28.47 | 23.88 | 27.34 | 30.49 | 29.4 | 8.98 | 29.05 | |
| HACK++Backbone=Infinity-8B, ηa=30%, ηc=1%, KV Size=0.17 GB, Attn. TFLOPs=26.502026.06 | 30.09 | 28.08 | 29.66 | 28.47 | 24.39 | 27.42 | 30.72 | 29.41 | 9.11 | 29.09 | |
| SnapKVBackbone=Infinity-8B, ηa=30%, ηc=30%, KV Size=5.06 GB, Attn. TFLOPs=33.552026.06 | 30.23 | 28.01 | 29.64 | 28.45 | 23.23 | 27.58 | 30.65 | 29.46 | 9.45 | 29.08 | |
| FullBackbone=HART, ηa=100%, ηc=100%, KV Size=2.68 GB, Attn. TFLOPs=17.662026.06 | 30.47 | 27.25 | 30.67 | 27.26 | 24.53 | 26.14 | 30.15 | 27.9 | 10.7 | 27.69 | |
| ScaleKVBackbone=Infinity-8B, ηc=1%, KV Size=0.17 GB, Attn. TFLOPs=31.402026.06 | 30.57 | 28.08 | 28.49 | 28.74 | 23.92 | 27.42 | 31.33 | 29.61 | 9.49 | 29.13 | |
| CAKEBackbone=Infinity-8B, ηa=30%, ηc=30%, KV Size=5.06 GB, Attn. TFLOPs=33.552026.06 | 30.74 | 27.86 | 30.06 | 28.44 | 23.38 | 27.54 | 31.19 | 29.44 | 9.8 | 29.05 | |
| HACK++Backbone=Infinity-2B, ηa=30%, ηc=10%, KV Size=0.77 GB, Attn. TFLOPs=14.322026.06 | 30.78 | 26.27 | 31.95 | 26.69 | 26.01 | 26.11 | 31.27 | 28.06 | 10.49 | 27.54 | |
| SnapKVBackbone=Infinity-2B, ηa=30%, ηc=30%, KV Size=2.31 GB, Attn. TFLOPs=15.342026.06 | 30.83 | 26.3 | 32.47 | 26.79 | 25.46 | 26.3 | 30.62 | 28.13 | 10.6 | 27.56 | |
| FullBackbone=Infinity-2B, ηa=100%, ηc=100%, KV Size=7.71 GB, Attn. TFLOPs=35.912026.06 | 30.96 | 26.33 | 31.68 | 26.64 | 25.98 | 26.15 | 31.62 | 30.26 | 10.34 | 27.52 | |
| HACK++Backbone=Infinity-2B, ηa=30%, ηc=1%, KV Size=0.08 GB, Attn. TFLOPs=12.112026.06 | 31.39 | 26.16 | 32.17 | 26.66 | 26.67 | 26.1 | 31.04 | 28.03 | 10.5 | 27.5 | |
| H2OBackbone=HART, ηa=30%, ηc=30%, KV Size=0.80 GB, Attn. TFLOPs=7.402026.06 | 31.49 | 26.51 | 33.78 | 27.28 | 22.42 | 26.07 | 31.97 | 27.64 | 11.13 | 27.32 | |
| ScaleKVBackbone=Infinity-2B, ηc=10%, KV Size=0.77 GB, Attn. TFLOPs=19.192026.06 | 31.53 | 26.28 | 32.3 | 26.67 | 25.95 | 26.13 | 31.53 | 28.04 | 10.55 | 27.54 | |
| HACKBackbone=Infinity-2B, ηa=30%, ηc=30%, KV Size=2.31 GB, Attn. TFLOPs=15.342026.06 | 31.54 | 26.24 | 31.95 | 26.79 | 26.14 | 26.29 | 31.61 | 28.23 | 10.56 | 27.62 | |
| CAKEBackbone=Infinity-2B, ηa=30%, ηc=30%, KV Size=2.31 GB, Attn. TFLOPs=15.342026.06 | 31.67 | 26.25 | 32.49 | 26.77 | 25.94 | 26.3 | 31.39 | 28.16 | 10.59 | 27.56 | |
| SnapKVBackbone=HART, ηa=30%, ηc=30%, KV Size=0.80 GB, Attn. TFLOPs=7.402026.06 | 31.71 | 26.51 | 33.87 | 27.34 | 22.44 | 26.16 | 32.36 | 27.62 | 11.33 | 27.33 | |
| H2OBackbone=Infinity-2B, ηa=30%, ηc=30%, KV Size=2.31 GB, Attn. TFLOPs=15.342026.06 | 31.72 | 26.29 | 32.59 | 26.71 | 26.62 | 26.26 | 30.99 | 28.13 | 10.68 | 27.57 | |
| StreamingBackbone=HART, ηa=30%, ηc=30%, KV Size=0.80 GB, Attn. TFLOPs=7.402026.06 | 32.13 | 26.96 | 31.84 | 27.41 | 23.22 | 26.12 | 30.37 | 27.43 | 11.16 | 27.38 | |
| LOOK-MBackbone=Infinity-8B, ηa=30%, ηc=30%, KV Size=5.06 GB, Attn. TFLOPs=33.552026.06 | 32.29 | 27.89 | 30.79 | 28.5 | 24.55 | 27.55 | 32.81 | 29.45 | 10.84 | 29.04 | |
| MEDABackbone=Infinity-2B, ηa=30%, ηc=30%, KV Size=2.31 GB, Attn. TFLOPs=15.342026.06 | 32.47 | 26.26 | 33.02 | 26.91 | 26.41 | 26.35 | 31.44 | 28.2 | 11.14 | 27.59 | |
| LOOK-MBackbone=Infinity-2B, ηa=30%, ηc=30%, KV Size=2.31 GB, Attn. TFLOPs=15.342026.06 | 32.53 | 26.26 | 33.2 | 26.91 | 26.49 | 26.34 | 31.36 | 28.21 | 11.14 | 27.59 | |
| StreamingBackbone=Infinity-2B, ηa=30%, ηc=30%, KV Size=2.31 GB, Attn. TFLOPs=15.342026.06 | 32.59 | 26.16 | 33.67 | 26.63 | 27.04 | 26.17 | 31.89 | 28.17 | 11.2 | 27.54 | |
| CAKEBackbone=HART, ηa=30%, ηc=30%, KV Size=0.80 GB, Attn. TFLOPs=7.402026.06 | 32.75 | 26.4 | 36.27 | 27.24 | 23.78 | 26.06 | 34.55 | 27.5 | 12.88 | 27.2 | |
| ScaleKVBackbone=Infinity-2B, ηc=1%, KV Size=0.08 GB, Attn. TFLOPs=14.352026.06 | 32.87 | 26.14 | 33.36 | 26.51 | 27.97 | 26.09 | 31.41 | 28.01 | 10.99 | 27.47 | |
| MEDABackbone=Infinity-8B, ηa=30%, ηc=30%, KV Size=5.06 GB, Attn. TFLOPs=33.552026.06 | 33.48 | 28.01 | 32.14 | 28.42 | 25.11 | 27.43 | 33.2 | 29.41 | 11.56 | 29 | |
| LOOK-MBackbone=HART, ηa=30%, ηc=30%, KV Size=0.80 GB, Attn. TFLOPs=7.402026.06 | 44.69 | 24.87 | 48.21 | 26.04 | 41.37 | 24.77 | 51.78 | 26.13 | 25.64 | 25.85 | |
| MEDABackbone=HART, ηa=30%, ηc=30%, KV Size=0.80 GB, Attn. TFLOPs=7.402026.06 | 47.12 | 24.62 | 52.33 | 25.79 | 43.12 | 24.71 | 54.84 | 26.07 | 28.37 | 25.71 |