Video Generation on VR-Bench
79.28Overall ScoreLumos-Nexus
Evaluation Results
| Method | Links | |
|---|---|---|
| Lumos-NexusInference Latency=40.5 s/step, Inference FLOPs=4.568 P/step, Inference GPU Usage=82.2 GB, Training Latency=2.25 s/it, Training GPU Usage=26.3 GB2026.05 | 79.28 | |
| Wan2.1-14BInference Latency=35.1 s/step, Inference FLOPs=3.462 P/step, Inference GPU Usage=45.7 GB, Training Latency=9.72 s/it, Training GPU Usage=94.7 GB2026.05 | 78.23 | |
| Omni-VideoInference Latency=7.8 s/step, Inference FLOPs=1.014 P/step, Inference GPU Usage=39.6 GB, Training Latency=2.25 s/it, Training GPU Usage=26.3 GB2026.05 | 72.78 | |
| Omni-Video*Inference Latency=36.1 s/step, Inference FLOPs=3.513 P/step, Inference GPU Usage=60.3 GB, Training Latency=7.21 s/it, Training GPU Usage=72.8 GB, Description=Variant replaces original 1.3B generator with Wan2.1-14B and is fine-tuned via LoRA2026.05 | 70.63 |