Human Preference Evaluation on ImageReward
1.0533Average ScoreSparVAR
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SparVARSkip scale strategy=w/o skip scales, Model size=8B2026.02 | 1.0533 | — | |
| Infinity-8BSkip scale strategy=w/o skip scales, Model size=8B2026.02 | 1.0529 | — | |
| Infinity-8BBackbone=Infinity-8B, Latency(s)↓=5.31, Speedup=1.0 ×2026.02 | 1.04 | — | |
| +ToProVARBackbone=Infinity-8B, Latency(s)↓=1.75, Speedup=3.0 ×2026.02 | 1.04 | — | |
| FastVARSkip scale strategy=w/o skip scales, Model size=8B2026.02 | 1.038 | — | |
| SparVARSkip scale strategy=w/ skip last 2 scales, Model size=8B2026.02 | 1.0377 | — | |
| ScaleKVSkip scale strategy=w/o skip scales, Model size=8B2026.02 | 1.035 | — | |
| SparVARSkip scale strategy=w/ dynamic skip last scales, Model size=8B2026.02 | 1.0305 | — | |
| +SkipVARBackbone=Infinity-8B, Latency(s)↓=2.65, Speedup=2.0 ×2026.02 | 1.03 | — | |
| SkipVAR-8BSkip scale strategy=w/ dynamic skip last scales, Model size=8B2026.02 | 1.0297 | — | |
| FastVARSkip scale strategy=w/ skip last 2 scales, Model size=8B2026.02 | 1.0276 | — | |
| +FastVARBackbone=Infinity-8B, Latency(s)↓=1.97, Speedup=2.6 ×2026.02 | 1.02 | — | |
| Infinity2026.04 | 0.9515 | 2,716 | |
| SparVARSkip scale strategy=w/ dynamic skip last scales, Model size=2B2026.02 | 0.9455 | — | |
| Infinity-2BSkip scale strategy=w/o skip scales, Model size=2B2026.02 | 0.9443 | — | |
| SparVARSkip scale strategy=w/o skip scales, Model size=2B2026.02 | 0.9416 | — | |
| Infinity-2BBackbone=Infinity-2B, Latency(s)↓=1.57, Speedup=1.0 ×2026.02 | 0.94 | — | |
| SkipVAR-2BSkip scale strategy=w/ dynamic skip last scales, Model size=2B2026.02 | 0.9399 | — | |
| SkipVARBackbone=Infinity, threshold=@0.84, requires additional training=true2026.04 | 0.9376 | 1,744 | |
| FastVARSkip scale strategy=w/o skip scales, Model size=2B2026.02 | 0.9327 | — | |
| ScaleKVSkip scale strategy=w/o skip scales, Model size=2B2026.02 | 0.9327 | — | |
| +SkipVARBackbone=Infinity-2B, Latency(s)↓=0.87, Speedup=1.8 ×2026.02 | 0.93 | — | |
| +ToProVARBackbone=Infinity-2B, Latency(s)↓=0.58, Speedup=2.7 ×2026.02 | 0.93 | — | |
| DepthVARBackbone=Infinity, R=92026.04 | 0.9254 | 1,616 | |
| SparVARSkip scale strategy=w/ skip last 2 scales, Model size=2B2026.02 | 0.9201 | — | |
| +FastVARBackbone=Infinity-2B, Latency(s)↓=0.62, Speedup=2.5 ×2026.02 | 0.92 | — | |
| FastVARSkip scale strategy=w/ skip last 2 scales, Model size=2B2026.02 | 0.9191 | — | |
| DepthVARBackbone=Infinity, R=82026.04 | 0.9171 | 1,303 | |
| FastVARBackbone=Infinity, w/o last two scales=true2026.04 | 0.9116 | 1,036 | |
| DepthVARBackbone=Infinity, R=72026.04 | 0.9088 | 1,174 | |
| HART2026.04 | 0.9013 | 1,103 | |
| HART-EE-21early exit=true, exit layer=212026.04 | 0.9004 | 985 | |
| DepthVARBackbone=Infinity, w/o last two scales=true, R=72026.04 | 0.8996 | 876 | |
| Infinity-EE-26early exit=true, exit layer=262026.04 | 0.8965 | 2,206 | |
| SparseVAR-0.7Backbone=Infinity2026.04 | 0.8936 | 1,301 | |
| DepthVARBackbone=HART, R=102026.04 | 0.7909 | 880 | |
| ToMeBackbone=Infinity, config={0.5, 0.5}2026.04 | 0.784 | 1,287 | |
| FastVARBackbone=HART, w/o FlashAttn=true2026.04 | 0.7448 | 1,209 | |
| DepthVARBackbone=HART, R=92026.04 | 0.6573 | 727 | |
| SparseVAR-0.7Backbone=HART2026.04 | 0.5737 | 679 |