Two-minute worlds · Generated worlds
What happens if we keep going?
Two minutes in a generated world, with plenty of turns along the way.
These models get different information. Full future camera path + caption versus Next action block. These clips don't support a matched-input ranking.
Clip files & individual measurements
SANA-WM Streaming
NVIDIA / SANA authors
- AI landmark retention
- 25.0 %
- AI landmark contradiction
- 66.7 %
- AI landmark uncertainty
- 8.3 %
- Return judgments completed
- 100.0 %
- Judge first-attempt coverage
- 100.0 %
- Turn pairs tracked
- 99.4 %
- Rotation pairs qualified
- 99.4 %
- Commanded warp discrepancy
- 1.289 % diagonal
- Qualified rotation error
- 1.877 degrees
- Qualified yaw direction
- 100.0 %
- Idle warp discrepancy
- 0.144 % diagonal
- First decoded worker chunk
- 17.3 seconds
- Steady worker throughput
- 9.885 frames/s
- Peak allocated GPU memory
- 49.454 GiB
- Session process time
- 326.6 seconds
- Late / early worker latency
- Native SANA export provides session/first/steady summaries, not per-chunk latency. No inferred drift measurement.
Matrix-Game 3 Base
Skywork
- AI landmark retention
- 83.3 %
- AI landmark contradiction
- 12.5 %
- AI landmark uncertainty
- 4.2 %
- Return judgments completed
- 100.0 %
- Judge first-attempt coverage
- 100.0 %
- Turn pairs tracked
- 85.8 %
- Rotation pairs qualified
- 85.8 %
- Commanded warp discrepancy
- 1.180 % diagonal
- Qualified rotation error
- 1.599 degrees
- Qualified yaw direction
- 94.0 %
- Idle warp discrepancy
- 0.019 % diagonal
- First decoded worker chunk
- 164.5 seconds
- Steady worker throughput
- 0.235 frames/s
- Peak allocated GPU memory
- 31.196 GiB
- Session process time
- 8383.7 seconds
- Late / early worker latency
- 1.000 ratio
Each clip shows one registered input and seed. The full study includes every seed and failure.
Look for thisCompare the early returns with the late ones. Watch for objects shifting, disappearing, or becoming hard to recognize.
Every run, including failures (data file) (opens in a new tab)
Choose a scene or change the models
Examples stay in their registered order. Every seed and any failed runs remain in the full results.
Compare the numbers across every run
This small stress test covers just two scenes and one random seed. Its AI landmark labels describe these clips, not a general winner.
Results corrected on 2026-10-16
Preserve original raw videos and worker snapshots. For SANA score registered camera timestamp f using saved output frame f-1; evaluate interframe motion against camera poses[1:1+n_output]. MG3 retains frame f and poses[0:n_output]. This source-defined alignment correction preceded all primary scoring; original outputs are retained.
Analysis and original results (opens in a new tab)How we tested it — results, method & every run
Two-minute world stress sessions
Which released configuration preserves recognizable scene landmarks across repeated one- and two-minute turn-away/return sessions, follows camera actions, and sustains generation throughput?
2 cases · 1 seed · 3 compared configurations · completed 2026-10-16 · descriptive stress subset
What we found. Two scenes, one seed each: this stress subset describes these outputs only and has no confidence intervals. Base received 83.3% retained-landmark labels at 0.236 worker fps; Distilled received 47.9% at 5.117 fps, with 50.0% unclear labels. These scores combine scene visibility and persistence; they do not isolate memory. Landmark scores are AI estimates, not validated game-state memory. Compare MG3 Base and Distilled within their shared next-action interface; SANA receives a future camera path and caption. Worker speed excludes browser delivery.
Watch the turn-away/return cycles and compare the starting scene with later returns. Scores concern three declared visible landmarks, not every object. SANA has extra future-camera and caption inputs; cross-group clips are illustrative, without paired rankings.
| Model configuration | AI landmark retention% · higher is better | AI landmark contradiction% · lower is better | AI landmark uncertainty% · lower is better | Return judgments completed% · higher is better | Judge first-attempt coverage% · higher is better | Turn pairs tracked% · higher is better | Rotation pairs qualified% · higher is better | Commanded warp discrepancy% diagonal · lower is better | Qualified rotation errordegrees · lower is better | Qualified yaw direction% · higher is better | Idle warp discrepancy% diagonal · lower is better | First decoded worker chunkseconds · lower is better | Steady worker throughputframes/s · higher is better | Peak allocated GPU memory |
|---|
