Published baselines only. Full is the official task-pass ranking.
101 / 170tasks passed
77main tasks uniquely passed
8models in this setting
850samples / model
Omni-Reason-1
59.4%
Full task pass rate · 101 of 170 main tasks
vs
Strongest Full baseline
11.2%
MiniMax H3 · 19 main tasks
Task completion from Table 1. Click a column to sort.
Model comparison
Compare only results reported on the same split and metric.
By family
Ours vs the strongest other model in that family. Click a row to sort the table.
Performance is shown by capability family, using the task-level Full comparison or the sample-level Mini comparison as indicated. Select a family to sort the main table and jump to the corresponding task catalogue.
The blue bar is Omni-Reason-1. The tick is the best other published score in that family.
Understanding the metrics
How to read the reported numbers
A sample passes when the main functional outcome is useful and mostly complete. The judge checks identity, counts, action order, and correspondence to the input. Missing or substantially incorrect operations fail; valid alternatives and minor rendering artifacts may pass. Completion is separate from PSNR / MSE / SSIM / MS-SSIM / LPIPS / CLIP similarity to a particular target.
Full · Table 1170 tasks · 850 samples · eight models. A task passes when at least 3 of 5 samples pass. Full also informed branch selection.
Mini · Table 1141 common samples · twelve models. One predetermined first released sample per task. Sample pass rate, not Full task pass rate. Ours 84 (59.6%) vs Kling O3 Standard 34 (24.1%), a 35.5-point gap.
JudgeGPT-6 Astra. Ours: CFG 1.5 · 50 steps. HY-OmniWeaving and MiniMax H3 miss some reference inputs (98 and 30 Full samples).
492 / 850sample Pass judgments for Omni-Reason-1. Strongest Full baseline reaches 106.
21.43 dBPSNR vs 14.40 dB baseline max. These do not decide the table.
ConsistencyOurs leads all six target-consistency metrics. Wan VACE can beat MiniMax H3 on those metrics (6 vs 19 task passes), so Target agreement is not the ranking.
Known gaps on this bench
A head-camera view can look plausible and still miss the intended stacks. Tracking can mark points without keeping identity. View synthesis passes 4/12 tasks; future/action passes 5/16.
November 1 manuscript: Table 1 and Figure 6 use the same final output batch. Full informed development and branch selection.
Transfer
November 1 · Section 5.2 · same checkpoint, no further fine-tuning
849 planned samples; 827 evaluable and 22 unsupported. Related pretraining exposure is unverified. Appendix F.2 retains the complete results. Understand transfer →
