One-minute worlds · Generated worlds
Still the same place a minute later?
Keep looking away and coming back. See what stays put over a longer visit.
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
- 33.3 %
- AI landmark contradiction
- 58.3 %
- AI landmark uncertainty
- 8.3 %
- Return judgments completed
- 100.0 %
- Judge first-attempt coverage
- 100.0 %
- Turn pairs tracked
- 100.0 %
- Rotation pairs qualified
- 100.0 %
- Commanded warp discrepancy
- 1.075 % diagonal
- Qualified rotation error
- 1.577 degrees
- Qualified yaw direction
- 100.0 %
- Idle warp discrepancy
- 0.152 % diagonal
- First decoded worker chunk
- 74.4 seconds
- Steady worker throughput
- 10.024 frames/s
- Peak allocated GPU memory
- 49.287 GiB
- Session process time
- 277.4 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
- 100.0 %
- AI landmark contradiction
- 0.0 %
- AI landmark uncertainty
- 0.0 %
- Return judgments completed
- 100.0 %
- Judge first-attempt coverage
- 100.0 %
- Turn pairs tracked
- 94.3 %
- Rotation pairs qualified
- 94.3 %
- Commanded warp discrepancy
- 1.053 % diagonal
- Qualified rotation error
- 1.724 degrees
- Qualified yaw direction
- 95.2 %
- Idle warp discrepancy
- 0.018 % diagonal
- First decoded worker chunk
- 166.5 seconds
- Steady worker throughput
- 0.237 frames/s
- Peak allocated GPU memory
- 31.015 GiB
- Session process time
- 4265.1 seconds
- Late / early worker latency
- 0.999 ratio
Each clip shows one registered input and seed. The full study includes every seed and failure.
Look for thisPick a landmark in the opening view. Follow it each time the camera comes back.
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
Landmark scores are AI estimates, not independently reviewed memory tests. One run ended at our rental deadline; incomplete averages are withheld.
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)Results corrected on 2026-10-16
Post-registration execution amendment: one judge request failed before returning a label. One separately recorded retry used identical images and configuration. Both attempts, first-attempt coverage and the unknown-cost reservation are preserved; no generation was replaced.
Analysis and original results (opens in a new tab)How we tested it — results, method & every run
Do generated worlds remember the way back?
Which released configuration preserves recognizable scene landmarks across repeated one- and two-minute turn-away/return sessions, follows camera actions, and sustains generation throughput?
4 cases · 2 seeds · 3 compared configurations · completed 2026-10-16
What we found. Four familiar scenes and two seeds test repeated returns over a minute. 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. 1 session hit the operator rental deadline; its partial evidence is retained and incomplete balanced averages are withheld. This is not a measured model-quality failure.
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 memoryGiB · lower is better | Session process time |
|---|
