WorldRIDE: Qualitative & Benchmark Summary
Synchronized 2×2 video comparisons and controlled benchmark tables on the identical SANA-WM backbone (1280×704), covering WorldCompass RL, faithful & 3-reward adapted Re-DMD, stepwise reward ablations, ordered-gate / local-normalization failure modes, and 60-second closed-loop world revisits.
+1.4 over Streaming, outperforming WorldCompass RL & Re-DMD on all consistency and 60s revisit metrics.
Ordered Gate blocks static/wrong-action reward hacking (0.0% leakage on degraded-quality motion traps).
1,063 ms vs. 6,552 ms for 6 independent rollouts, saving 2.86 GB VRAM (-6.13 GB with detached replay).
Single-stage 233.6 H200-h vs. 2-stage SF + WorldCompass RL (41.6 h / 307.5 H200-h); 22.5× faster DiT than Teacher.
1. Synchronized 2×2 Qualitative Video Comparisons
All four panels in each video are frame-synchronized at identical timestamps with identical initial frames, prompts, and camera action trajectories.
Streaming loses the boardwalk into sand dunes; WorldCompass RL warps wooden planks; 3-Reward Re-DMD suffers a chunk-boundary break. WorldRIDE keeps the walker centered along an intact boardwalk.
Streaming completely loses the walking student into an empty path; WorldCompass RL shrinks and drifts the subject onto grass; Re-DMD dissolves the brick path into weeds. WorldRIDE preserves both subject and path geometry.
Single-anchor 3-Reward Re-DMD suffers a severe chunk-boundary jump cut into a completely different flower garden; Streaming & WorldCompass RL deform the cliffside scene. WorldRIDE stays seamless across chunks.
Streaming clones the monk into three figures; WorldCompass RL collapses into dark pillars; 3-Reward Re-DMD jump-cuts into a circular white plaza (Seg=0). WorldRIDE completes the L-turn with intact courtyard architecture.
Under repeated left-right zigzag controls, Streaming and WorldCompass RL lose the explorer and collapse the colonnade; Re-DMD distorts column perspective. WorldRIDE keeps symmetric stone columns across turns.
Streaming and WorldCompass RL smear the fruit cart into cluttered stalls; single-anchor Re-DMD tears at chunk boundaries (Seg=0). WorldRIDE maintains crisp geometric floor tiles and coherent cart motion.
Streaming hallucinates a blue orb; w/o Action+Geo crumbles the astronaut; w/o Geometry stabilizes heading but jumps scale. Adding conditional Geometry (WorldRIDE Full) locks 3D astronaut scale and Earth curvature.
Over 573 frames, Streaming loses the elephant off-frame; w/o Action+Geo spawns duplicate elephants; w/o Geometry drifts sideways off-trail. WorldRIDE Full preserves a single coherent elephant along the sunset trail.
Streaming warps players and stadium stands; w/o Action+Geo & w/o Geometry misalign goalposts under pan. WorldRIDE Full maintains crisp pitch lines, goal geometry, and causal ball control.
Clear illustration of Global EMA motion-freeze hacking: Global EMA (top-right) freezes like a static photograph (Dyn=0); Ungated Flat Sum spawns road signs through the hiker. WorldRIDE Full executes smooth roundtrip motion.
Global EMA stalls nearly motionless despite roundtrip controls; Ungated Flat Sum hallucinates a phantom duplicate figure. WorldRIDE Full delivers dynamic travel and clean neon reflections.
In Global EMA, the child freezes tiny in the distance and never returns on the reverse camera leg; Unweighted DMD loses path benches. WorldRIDE Full smoothly completes the 573-frame roundtrip.
Unweighted DMD crashes into a blurry stone wall (Persp 0.11); Global EMA stalls before the L-turn; Ungated Flat Sum distorts the walking cat. WorldRIDE Full cleanly completes the alley turn.
960-frame closed-loop trajectory returning to visited viewpoints: Unweighted DMD loses the ocean horizon and Quality Re-DMD washes into beige haze; WorldRIDE Full preserves snow ripples and coastal icebergs.
Across 60 seconds (24 chunks), Unweighted DMD drifts off structural landmarks (-4.1 dB) and Quality Re-DMD blurs roof trusses; WorldRIDE Full preserves steel pillars, brick bays, and floor rails.
4-step Streaming suffers severe limb dismemberment; 4-step WorldRIDE preserves anatomical integrity and sand tracks at 22.5× faster DiT speed than the 60-step teacher.
Streaming duplicates the dog in later chunks; WorldRIDE maintains a single coherent subject and stable garden layout matching the 60-step reference.
L-shaped camera turn across an architectural plaza: WorldRIDE achieves 0.94 Perspective Consistency vs. 0.50 for Streaming.
4-step WorldRIDE preserves boat symmetry, water reflections, and distant mountain shoreline across 384 streaming frames.
Crisp Doric column geometry under pitch-down; 4-step WorldRIDE surpasses the 60-step Causal Teacher across Total, I2V, and Quality on 5,590 videos.
Streaming hallucinates overexposed orange artifacts at borders; WorldRIDE preserves faithful cyan coral structure.
2. Qualitative Case & Failure-Mode Summary Table
Overview of representative qualitative comparisons above, contrasting baseline failure modes with WorldRIDE Full.
| Comparison Group | Scene & Trajectory | Observed Baseline / Ablation Failure Mode | WorldRIDE Full Behavior | Key Case Metric | Video |
|---|---|---|---|---|---|
| vs. RL & Re-DMD | Wooden Beach Boardwalk (L-shape, 24 s) | Streaming loses boardwalk; WC RL warps planks; Re-DMD chunk break (Seg=0) |
Continuous straight boardwalk & centered walker | Persp 0.87 · Seg 1.0 | ▶ Watch |
| vs. RL & Re-DMD | University Campus Path (Roundtrip, 24 s) | Streaming subject vanishes completely; WC RL drifts onto grass; Re-DMD loses brick path |
Subject & brick walkway preserved on roundtrip | Scene 1.0 · Subj 1.0 | ▶ Watch |
| vs. RL & Re-DMD | Floating Sky Island (Roundtrip, 24 s) | 3-Reward Re-DMD jump-cuts to unrelated flower garden (Seg=0); WC RL cliff collapse |
Seamless multi-chunk cliff, balustrade & waterfalls | Persp 0.96 · Seg 1.0 | ▶ Watch |
| vs. RL & Re-DMD | Tibetan Monastery (L-shape, 30 s) | Streaming clones 3 monks; WC RL pillar occlusion; Re-DMD jump-cuts (Seg=0) |
Clean 30 s L-turn with intact temple architecture | Persp 0.92 · Seg 1.0 | ▶ Watch |
| vs. RL & Re-DMD | Roman Ruins (Zigzag) & Bazaar (Repeat) | Streaming / WC RL lose subject & colonnade; Re-DMD chunk tear (Seg=0) |
Symmetric stone columns & crisp geometric tiles | Persp 0.89 / 0.85 | ▶ Watch |
| Stepwise Reward | Outer Space Horizon (Progressive, 24 s) | w/o Action+Geo crumbles suit; w/o Geometry jumps scale (Persp 0.34) |
Stable 3D astronaut scale & curved Earth horizon | Persp 0.95 vs. 0.34 | ▶ Watch |
| Stepwise Reward | Savanna Loop (36 s) & Soccer Pitch (18 s) | w/o Action+Geo duplicates elephant; w/o Geometry drifts off-trail / warps goalposts |
Single coherent subject & stable 3D pitch geometry | Persp 0.59 / 0.92 | ▶ Watch |
| Norm & Gate | Desert Highway & Cyberpunk Street (24 s) | Global EMA freezes like a static photo (Dyn=0); Flat Sum phantom artifacts |
Dynamic roundtrip camera travel without artifacts | Dyn 1.0 vs. 0.0 | ▶ Watch |
| Norm & Gate | Autumn Park (36 s) & Cobblestone Alley (24 s) | Global EMA stalls in distance (Dyn=0); Unweighted DMD crashes into wall |
Full 573f roundtrip return & clean alley L-turn | Persp 0.71 vs. 0.11 | ▶ Watch |
| 60s Revisit | Arctic Icefield & Industrial Warehouse (960f) | Unweighted DMD landmark drift; Quality Re-DMD color/haze washout over 60 s |
Accurate 60-second closed-loop landmark revisit | +2.8 to +4.1 dB PSNR | ▶ Watch |
3. Same-Backbone Controlled Baselines (WorldCompass RL & Re-DMD)
All 4-step methods use the identical SANA-WM Streaming backbone, training dataset, Causal ODE initialization, and unified qwen3.8-max WBench + 60-second SANA-WM 80-Scene (160 videos, 800 revisit pairs) evaluation.
| 4-Step Model (Same SANA-WM Backbone & Data) | Paradigm (Anchors / Norm / Reward) | WBench Overall ↑ | SegCont ↑ | NavTraj ↑ | 60s VBench Total (%) ↑ | 60s Camera RotErr (°) ↓ |
|---|---|---|---|---|---|---|
| SANA-WM Streaming (official 4-step) | K=1 / Unweighted / None | 72.5 | 91.8 | 81.9 | 69.07 | 22.72 |
| Unweighted DMD (same training data & init) | K=1 / Equal (w=1) / None | 72.8 | 99.4 | 81.3 | 68.80 | 20.33 |
| Self-Forcing + WorldCompass RL (2-stage) | Chunk RL (G=12, top/bot 3) | 72.8 | 90.5 | 82.0 | 68.67 | 19.39 |
| Single-Anchor Quality Re-DMD (faithful Re-DMD) | K=1 / Global EMA / HPSv3 | 73.4 | 94.9 | 81.3 | 67.69 | 18.34 |
| Single-Anchor 3-Reward Re-DMD (adapted Re-DMD) | K=1 / Global EMA / Q+A+Geo | 73.6 | 92.4 | 81.3 | 68.63 | 19.33 |
| WorldRIDE Full (Ours) | K=3 / Group Softmax / Ordered Gate | 73.9 | 98.1 | 82.6 | 69.14 | 16.58 |
4. Controlled Mechanism & Stepwise Reward Ablations
Table 2A: Isolating Ordered Gating, Multi-Anchor Rollout & Local Group Softmax (+ 192-Group / 384-Clip Counterfactual Audit)
| Isolated Mechanism | Controlled 4-Step Variant | Anchors / Norm / Gate | Wrong-Action Trap False Upweight ↓ | WBench Overall ↑ | SegCont ↑ | DynDeg ↑ | Setting ↑ | 60s VBench Total (%) ↑ |
|---|---|---|---|---|---|---|---|---|
| Baseline ref. | SANA-WM Streaming (official) | K=1 / — / — | — | 72.5 | 91.8 | 98.7 | 71.7 | 69.07 |
| Single-anchor equal | Unweighted DMD (same training data) | K=1 / Equal (w=1) / — | — | 72.8 | 99.4 | 97.5 | 71.9 | 68.80 |
| Contrib. 2 (Multi-anchor) | w/o Action + Geometry (multi-anchor equal) | K=3 / Equal (w=1) / — | — | 73.6 | 98.7 | 99.4 | 71.7 | 69.23 |
| Contrib. 1 (Ungated sum) | Multi-Anchor + Ungated Flat Sum | K=3 / Group Softmax / Flat Sum | 25.0% | 74.0* | 97.5 | 94.9 (-1.9) | 70.4 | 68.58 |
| Contrib. 3 (Global norm) | Multi-Anchor + Global EMA | K=3 / Global EMA / Ordered | 3.1% | 74.1* | 98.7 | 82.9 (-13.9!) | 70.9 | 67.42 |
| Full Method (Ours) | WorldRIDE Full | K=3 / Group Softmax / Ordered | 3.1% (8× lower) | 73.9 | 98.1 | 96.8 | 72.3 | 69.14 |
*Asterisks denote short-clip static consistency inflation from under-motion (Global EMA suffers a -13.9 collapse in Dynamic Degree and drops 60s VBench Total to 67.42%; see Desert Highway & Cyberpunk Street videos above).
Table 2B: Step-by-Step Gains of Action Following & Conditional 3D Geometry Rewards
| 4-Step Student Variant (Same SANA-WM Backbone) | Reward Configuration | WBench Overall ↑ | NavTraj (5s) ↑ | Perspective ↑ | 60s Camera RotErr (°) ↓ | 60s Revisit PSNR (dB) ↑ | 60s Revisit LPIPS ↓ |
|---|---|---|---|---|---|---|---|
| SANA-WM Streaming (official 4-step) | Unweighted (K=1) | 72.5 | 81.9 | 49.6 | 22.72 | 14.26 | 0.5697 |
| Unweighted DMD (same training data & init) | Equal (w=1, K=1) | 72.8 | 81.3 | 60.4 | 20.33 | 14.31 | 0.5444 |
| w/o Action + Geometry (equal-weight multi-anchor) | Equal (w=1, K=3) | 73.6 | 82.9 | 62.7 | 15.81 | 15.08 | 0.5372 |
| w/o Geometry (+ quality-gated Action) | Q → Action (K=3) | 73.5 | 82.7 | 62.6 | 13.69 (-13.4%) | 14.94 | 0.5402 |
| WorldRIDE Full (+ conditional Geometry, Ours) | Q → A → Geo (K=3) | 73.9 | 82.6 | 67.0 (+4.4) | 16.58 | 15.26 | 0.5343 |
5. Training Time, GPU-Hours, Memory & Streaming Inference Speed
Table 3A: H200 CUDA-Synchronized Step Latency, Shared-Prefix vs. Independent Rollouts, Peak VRAM & 720p Inference FPS
| Rollout & Training Paradigm (Same SANA-WM Backbone) | Spine (ms) | 6-Branch Sampling (ms) ↓ | Step Time (s/step) ↓ | Peak Trainer VRAM (GB) ↓ | 720p Infer FPS ↑ (DiT / Stage-1) |
|---|---|---|---|---|---|
| SANA-WM Chunk-Causal (60-step Teacher) | — | — | — | — | 1.56 / 1.41 |
| Self-Forcing (single-anchor, unrewarded) | 2,972 | 0 | 5.84 | 39.17 | 35.06 / 10.99 |
| Self-Forcing + WorldCompass RL | — | G=12 rollouts | 5.84 + 20.80 | 39.17 / 35.24 | 34.90 / 10.97 |
| Global Full-Seq Re-DMD (M=2 full-seq decode) | 5,944 | 0 (2× full-seq) | 20.06 | 54.07 | 35.01 / 10.91 |
| Independent Rollouts (6 separate prefixes) | 3,019 | 6,552 | 16.38 | 56.93 | 35.01 / 10.91 |
| WorldRIDE w/ 4-Step Diff. Unroll (no detach) | 3,015 | 1,063 | 18.08 | 60.20 | 35.01 / 10.91 |
| WorldRIDE (Shared-Prefix + Detached, Ours) | 3,015 | 1,063 (6.2× faster) | 15.23 (-26.8% vs. RL) | 54.07 (-2.86 / -6.13 GB) | 35.01 / 10.91 |
Table 3B: End-to-End 1280×704 Training Wall-Clock & Total Active GPU-Hours on One 1×8 NVIDIA H200 Node
| Method (1280×704, 1×8 H200 Node) | Stages (Post-ODE) | Wall-Clock Time ↓ | Active H200 GPU-h ↓ | Speedup vs. SF + WorldCompass RL ↑ |
|---|---|---|---|---|
| Self-Forcing (unweighted DMD) | 1 Stage | 12.7 h | 76.2 | (Unrewarded baseline) |
| Self-Forcing + WorldCompass RL | 2 Stages | 41.6 h (12.7h + 28.9h) | 307.5 (332.9 node) | 1.00× (2-stage RL baseline) |
| WorldRIDE (Ours) | 1 Stage | 29.2 h | 233.6 | 1.42× faster (-30% time, -24% GPU-h) |
All methods share the same one-time Causal ODE initialization (1.3 h / 10.4 H200-h); GPU-hours include rollout generation, CausalVAE candidate decoding, frozen HPSv3 + DA3 evaluation, and DMD/RL updates.