Datasets:
image imagewidth (px) 1.08k 1.28k | label class label 9
classes |
|---|---|
0helixworld | |
0helixworld | |
0helixworld | |
0helixworld | |
0helixworld | |
0helixworld | |
0helixworld | |
0helixworld | |
0helixworld | |
2lingbot-world-v2 | |
2lingbot-world-v2 | |
2lingbot-world-v2 | |
2lingbot-world-v2 | |
2lingbot-world-v2 | |
2lingbot-world-v2 | |
2lingbot-world-v2 | |
2lingbot-world-v2 | |
1lingbot-world | |
1lingbot-world | |
1lingbot-world | |
1lingbot-world | |
3matrix-game | |
3matrix-game | |
3matrix-game | |
3matrix-game | |
4sana-wm | |
4sana-wm | |
4sana-wm | |
4sana-wm | |
4sana-wm | |
4sana-wm | |
4sana-wm | |
4sana-wm | |
4sana-wm | |
4sana-wm | |
5solarwm | |
5solarwm | |
5solarwm | |
5solarwm | |
6taiji-lingbot-wm2 | |
7taiji-matrix-game-35 | |
8taiji-solarwm | |
8taiji-solarwm | |
8taiji-solarwm | |
8taiji-solarwm | |
8taiji-solarwm | |
8taiji-solarwm | |
8taiji-solarwm | |
8taiji-solarwm | |
8taiji-solarwm |
StreamVSR World Model Optical Gallery
This gallery presents 50 four-panel videos demonstrating StreamVSR AR19000 on world-model rollouts. Each video shows bicubic enlargement in the upper left, 4× causal streaming super-resolution in the upper right, red/cyan stereoscopic video in the lower left, and a near-to-far refocus sweep in the lower right. The stereo view uses a virtual 6.5 cm baseline. The refocus sweep keeps the field of view fixed.
Browse the collection through the per-demo manifest. Each entry identifies its clip in videos/, its preview in previews/, its source attribution, protocol, frame count, and frame rate. The upload includes 49 clips of 97 frames and one 77-frame clip.
- Snow village. Compare fence rails and roof edges in the two top panels. Video · Preview
- Ivy, brickwork, and cobblestones. Watch brick joints and ivy leaves as the camera moves. Video · Preview
- Suburban street. Follow the refocus sweep from the foreground tree trunk toward the houses. Video · Preview
Processing protocols
- A · Enhancement, 21 clips. The source rollout serves as the low-resolution input. Larger inputs are resized to fit the inference budget before bicubic and StreamVSR each enlarge the same input by 4×.
- B · Restoration, 29 clips. The source rollout is reduced by 4× using antialiased bicubic downsampling. Bicubic and StreamVSR restore the low-resolution input toward the source dimensions. This controlled degradation often makes fine details such as fence rails and brick joints easier to compare.
Both workflows use the same AR19000 configuration with one DiT evaluation per streaming block. The four-panel sequences are qualitative comparisons on generated world-model footage. Restoration uses the source rollout as a reference for the synthetic downsampling task.
Sources and credit
The 39 public-source clips come from NVIDIA Research SANA-WM (10), Robbyant LingBot-World 2.0 (8), Noiz AI and HKUST HelixWorld (9), Robbyant LingBot-World (4), SolarX Team SolarWM (4), and Skywork AI Matrix-Game 2.0 (4). SANA-WM · LingBot-World 2.0 · HelixWorld rollouts · LingBot-World · SolarWM · Matrix-Game 2.0
The other 11 clips come from independent Taiji rollouts of SolarWM (9), LingBot-World 2.0 (1), and Matrix-Game 3.5 (1). The manifest credits each model and links to its project page. Matrix-Game 3.5
The source projects retain their respective rights in their materials. Consult the original project terms before reusing third-party footage.
- Downloads last month
- 188