Shape-of-Motion · DL3DV (static-real): -0.8 dB

Memory / revisit — a 4th axis (NEW)

Can a method REMEMBER what it has already seen? Single-shot & sliding-window generators have no persistent world state — when the camera revisits an observed pose or returns after wandering into unseen space, they re-hallucinate it rather than recall it. These two new benchmark families isolate exactly that. Both score only on real GT: the revisit/return legs carry true ground-truth frames; the synthetic away-legs are masked out upstream.

Two families × three sources (Kubric-4D clean-GT primary, DL3DV static, SpatialVID dynamic). M1 multi_visit = REVISIT-CONSISTENCY (GT-fidelity at a revisited observed pose + cross-visit self-consistency drift). M2 moving_back = RETURN-FIDELITY (quality back at the observed pose after a round-trip, and the PSNR degradation vs the pre-departure observed baseline — the forgetting signal). pending = method has outputs but is not yet scored on this family; methods with no run are omitted. 9 methods scored per benchmark now (incl. the dedicated spatial-memory model spmem and the warp-conditioned Wan2.1-14B variant FrameCrafter, both on all six memory benchmarks — both families × all three scene types); FlexWorld/TrajCrafter/ViewExtrapolator/Voyager and spmem on the still-generating long-horizon set remain pending.

⚠ Headline — explicit geometric caching, not learned memory, enables return-fidelity. Now confirmed across all three scene types. The moving-back degradation column (return-PSNR drop vs the method’s own pre-departure observed baseline; smaller = better memory) is the story, and the same split now holds across static-real (DL3DV), synthetic (Kubric-4D clean-GT), AND dynamic-real (SpatialVID) — so it is a real capability gap, not a domain or render artifact.

Geometric memory WORKS. Per-scene 4D-GS (Shape-of-Motion) stores the whole scene and barely degrades — −0.8 dB (DL3DV) / −1.0 dB (SpatialVID). Among forward generators, GEN3C — which keeps an explicit 3D point cache — is the best by a wide margin and is the only generator that stays low on every scene type: −17.0 dB (DL3DV) / −13.8 dB (Kubric) / −17.5 dB (SpatialVID) (return-PSNR 17.3 / 22.4 / 17.3).
Learned / neural spatial memory does NOT (yet). spmem — the dedicated spatial-memory paper, now scored on all six memory benchmarks — collapses −30.0 dB (DL3DV) / −24.1 dB (Kubric) / −32.0 dB (SpatialVID), among the worst on all three. The collapse is uniform across static, synthetic and dynamic, so it is not a domain gap; the learned memory simply isn’t recalling the observed pose — it forgets exactly like the memoryless single-shot generators.
The memoryless single-shot generators all collapse in the same band. Lyra 2.0 (−28.1 / −19.4 / −27.6), Ours(base VACE) (−28.2 / −21.7 / −29.2), NVS-Solver (−31.9 / −21.0 / −31.0), ViewCrafter (−28.5 / −25.4 / −32.5) dB — the ~−20 to −32 dB band that motivates this axis: today’s strongest extrapolators have no working memory, and spmem sits squarely inside it.

Honest caveat: Matrix-Game 3.0’s tiny degrade (−1.8 DL3DV / −9.5 Kubric / −3.7 SpatialVID dB) is not memory — it is an artifact of uniformly-low quality (return-PSNR ≈ 11 with observed-baseline ≈ 13–21 on all three), so there is little to forget. Always read the degrade column together with the absolute return-PSNR.

Headline result — return-degradation across all three scene types CAMPAIGN HEADLINE

The single number that carries this axis: return-PSNR degradation (drop vs the method’s own pre-departure observed baseline — lower bars = better memory). Six key methods × three scene types. The shape is identical everywhere: geometric caching stays near zero, every generator (learned-memory spmem included) collapses. All values are real-GT means read directly from the moving-back JSONs.

Method	Return-PSNR degrade↓ (dB, lower = better memory)			memory mechanism → verdict
Method	DL3DV static-real	Kubric-4D synthetic	SpatialVID dyn-real	memory mechanism → verdict
Shape-of-Motion · 4D-GS per-scene 4D-GS	−0.83	—	−1.01	geometric — stores the scene
GEN3C · Cosmos-7B 3D point cache	−17.04	−13.85	−17.52	explicit 3D cache — best generator
Ours · VACE-14B warp re-derive	−28.18	−21.65	−29.21	memoryless (re-derives from warp)
Lyra 2.0 · Wan2.1-14B single-shot	−28.12	−19.37	−27.62	memoryless
spmem · spatial-memory learned memory	−29.99	−24.08	−32.00	learned memory — collapses too
NVS-Solver single-shot	−31.93	−21.03	−30.97	memoryless

Shape-of-Motion is a per-scene 4D-GS reconstruction (the no-forgetting upper bound, not a forward generator) and is N/A on Kubric-4D where the OURS row is the warp-anchored VACE generator. Read across the three columns: the only method that holds on all three scene types is GEN3C’s explicit 3D cache; the dedicated learned-memory model spmem collapses hardest of the generators on every type (−24 to −32 dB), confirming the finding is about caching mechanism, not domain or motion.

M1 · multi_visit — revisit-consistency

Revisit a fixed observed pose 3× over the trajectory. ★ Revisit GT-fidelity = PSNR/SSIM/LPIPS vs the real GT frame at that pose (does the revisit match reality?). Cross-visit drift = self-PSNR/SSIM/LPIPS between the method's own visits to the same pose (GT-free — does it stay consistent with itself?). High self-PSNR with low GT-PSNR = consistently wrong; per-scene 4D-GS shows near-perfect self-consistency (it renders one fixed scene) as the reference.

M1a · Kubric-4D clean-GT PRIMARY — revisit a fixed observed pose (no warp confound)

Method	★ Revisit GT-fidelity (vs real GT at revisited pose)			Cross-visit drift (GT-free self-consistency)			n
Method	PSNR↑	SSIM↑	LPIPS↓	self-PSNR↑	self-SSIM↑	self-LPIPS↓	n
Ours · Wan2.1-VACE-14B OURS	19.46	0.772	0.229	33.69	0.955	0.026	30
GEN3C · Cosmos-7B	20.41	0.832	0.246	25.68	0.888	0.122	30
ViewCrafter	13.58	0.471	0.584	14.24	0.513	0.500	30
FrameCrafter · Wan2.1-14B	16.33	0.417	0.459	18.88	0.612	0.386	27
NVS-Solver	20.41	0.762	0.241	28.09	0.896	0.081	30
Lyra 2.0 · Wan2.1-14B	19.39	0.787	0.252	26.91	0.899	0.082	30
spmem · spatial-memory memory	15.87	0.495	0.511	21.05	0.712	0.320	30
Matrix-Game 3.0	11.10	0.484	0.735	15.89	0.724	0.330	30

M1b · DL3DV static — revisit a fixed observed pose 3×

Method	★ Revisit GT-fidelity (vs real GT at revisited pose)			Cross-visit drift (GT-free self-consistency)			n
Method	PSNR↑	SSIM↑	LPIPS↓	self-PSNR↑	self-SSIM↑	self-LPIPS↓	n
GEN3C · Cosmos-7B	15.86	0.484	0.485	15.61	0.477	0.469	14
ViewCrafter	12.50	0.323	0.636	13.65	0.379	0.535	14
FrameCrafter · Wan2.1-14B	16.69	0.481	0.389	16.99	0.519	0.381	12
NVS-Solver	11.78	0.349	0.726	12.82	0.437	0.618	14
Shape-of-Motion · 4D-GS per-scene 4D-GS	19.59	0.628	0.479	57.81	0.998	0.001	14
Lyra 2.0 · Wan2.1-14B	13.42	0.385	0.643	14.92	0.462	0.564	14
spmem · spatial-memory memory	11.99	0.351	0.730	14.87	0.483	0.527	12
Matrix-Game 3.0	10.81	0.353	0.761	13.16	0.491	0.607	14

M1c · SpatialVID in-the-wild dynamic — revisit

Method	★ Revisit GT-fidelity (vs real GT at revisited pose)			Cross-visit drift (GT-free self-consistency)			n
Method	PSNR↑	SSIM↑	LPIPS↓	self-PSNR↑	self-SSIM↑	self-LPIPS↓	n
Ours · Wan2.1-VACE-14B OURS	15.33	0.513	0.507	14.71	0.498	0.525	27
GEN3C · Cosmos-7B	16.86	0.567	0.406	17.59	0.591	0.378	30
ViewCrafter	9.10	0.258	0.798	9.86	0.257	0.733	29
FrameCrafter · Wan2.1-14B	17.60	0.533	0.362	18.51	0.611	0.340	28
NVS-Solver	12.83	0.441	0.634	15.52	0.603	0.468	30
Shape-of-Motion · 4D-GS per-scene 4D-GS	17.54	0.598	0.557	52.83	0.996	0.002	28
Lyra 2.0 · Wan2.1-14B	15.42	0.495	0.441	18.15	0.578	0.289	30
spmem · spatial-memory memory	11.18	0.437	0.689	13.35	0.611	0.523	27
Matrix-Game 3.0	10.13	0.391	0.721	13.22	0.525	0.556	30

Self-consistency (cross-visit) is trivially near-perfect for the per-scene 4D-GS reconstructor (it renders ONE optimized scene every visit — self-PSNR ~50–58 dB) — that is the geometric-memory upper bound, not a generative result. Among forward generators, GT-fidelity on Kubric-4D clean-GT separates the memory-equipped (Lyra 2.0 19.4 dB) from the memoryless (Matrix-Game 3.0 11.1 dB). Ours(base VACE) is competitive on the primary clean-GT revisit (19.5 dB) by re-deriving the pose from warp conditioning.

M2 · moving_back — return-fidelity

Move away into unseen space, then return to a previously-observed pose. ★ Return fidelity = PSNR/SSIM/LPIPS at the returned pose vs GT. obs-PSNR = the method's own quality at that pose before departing (the baseline). degrade = the PSNR drop after the round-trip (the memory/forgetting metric — smaller is better). away = mean camera distance travelled before returning.

M2a · Kubric-4D clean-GT PRIMARY — move away into unseen space, then return

Method	★ Return fidelity (back at the observed pose)			Return context			n
Method	PSNR↑	SSIM↑	LPIPS↓	obs-PSNR↑	degrade↓	away	n
Ours · Wan2.1-VACE-14B OURS	18.42	0.723	0.274	40.07	+21.65	19.92	30
GEN3C · Cosmos-7B	22.44	0.879	0.185	36.28	+13.85	19.92	30
ViewCrafter	13.64	0.472	0.573	39.00	+25.35	19.92	30
FrameCrafter · Wan2.1-14B	16.08	0.376	0.513	40.07	+24.00	19.92	30
NVS-Solver	19.05	0.769	0.243	40.08	+21.03	19.86	24
Lyra 2.0 · Wan2.1-14B	20.69	0.836	0.199	40.06	+19.37	19.57	29
spmem · spatial-memory memory	16.00	0.496	0.514	40.08	+24.08	19.92	30
Matrix-Game 3.0	11.06	0.494	0.752	20.61	+9.55	19.92	30

M2b · DL3DV static — move away, then return to the observed pose

Method	★ Return fidelity (back at the observed pose)			Return context			n
Method	PSNR↑	SSIM↑	LPIPS↓	obs-PSNR↑	degrade↓	away	n
GEN3C · Cosmos-7B	17.30	0.519	0.399	34.33	+17.04	4.66	14
ViewCrafter	12.26	0.312	0.659	40.71	+28.45	4.66	14
FrameCrafter · Wan2.1-14B	10.72	0.244	0.730	41.72	+31.00	4.81	12
NVS-Solver	9.88	0.301	0.787	41.82	+31.93	4.66	14
Shape-of-Motion · 4D-GS per-scene 4D-GS	20.02	0.642	0.442	20.86	+0.83	4.66	14
Lyra 2.0 · Wan2.1-14B	13.69	0.379	0.633	41.82	+28.12	4.66	14
spmem · spatial-memory memory	11.73	0.344	0.738	41.72	+29.99	4.81	12
Matrix-Game 3.0	11.25	0.343	0.732	13.06	+1.82	4.66	14

M2c · SpatialVID in-the-wild dynamic — move-back

Method	★ Return fidelity (back at the observed pose)			Return context			n
Method	PSNR↑	SSIM↑	LPIPS↓	obs-PSNR↑	degrade↓	away	n
Ours · Wan2.1-VACE-14B OURS	13.03	0.440	0.616	42.24	+29.21	5.25	27
GEN3C · Cosmos-7B	17.30	0.581	0.365	34.81	+17.52	5.16	30
ViewCrafter	9.07	0.262	0.776	41.60	+32.54	5.18	29
FrameCrafter · Wan2.1-14B	9.63	0.214	0.764	42.19	+32.56	5.16	30
NVS-Solver	11.22	0.370	0.706	42.19	+30.97	5.16	30
Shape-of-Motion · 4D-GS per-scene 4D-GS	17.00	0.583	0.569	18.00	+1.01	5.25	28
Lyra 2.0 · Wan2.1-14B	14.57	0.475	0.521	42.20	+27.62	5.16	30
spmem · spatial-memory memory	10.24	0.409	0.743	42.24	+32.00	5.25	27
Matrix-Game 3.0	11.33	0.376	0.620	15.01	+3.68	5.16	30

M3 · Kubric-4D return-fidelity by away-distance bin PRIMARY clean-GT

The angle bin = how far the camera travelled into unseen space before returning (30° nearby → 180° full back-and-return). Return-PSNR holds roughly flat across bins for the warp-anchored methods (they re-derive the observed region from the conditioning), exposing whether quality is driven by memory or by re-warping.

Method	30°		60°		90°		120°		180°
Method	ret-PSNR↑	degrade↓	ret-PSNR↑	degrade↓	ret-PSNR↑	degrade↓	ret-PSNR↑	degrade↓	ret-PSNR↑	degrade↓
Ours · Wan2.1-VACE-14B OURS	18.36	+21.7	18.49	+21.6	18.42	+21.7	18.40	+21.7	18.44	+21.6
GEN3C · Cosmos-7B	22.62	+13.7	22.49	+13.8	22.57	+13.7	22.31	+14.0	22.19	+14.1
ViewCrafter	13.64	+25.4	13.64	+25.4	13.64	+25.4	13.64	+25.4	13.64	+25.4
FrameCrafter · Wan2.1-14B	16.08	+24.0	16.08	+24.0	16.08	+24.0	16.08	+24.0	16.08	+24.0
NVS-Solver	18.75	+21.3	18.95	+21.1	19.64	+20.3	19.07	+21.0	18.96	+21.1
Lyra 2.0 · Wan2.1-14B	20.43	+19.7	20.58	+19.5	20.59	+19.5	20.56	+19.5	21.39	+18.6
Matrix-Game 3.0	11.06	+9.5	11.06	+9.5	11.06	+9.5	11.06	+9.5	11.06	+9.5

⚠ How to read ‘degrade’: a large positive degrade means the method was sharp at the pose before leaving but forgot it after wandering — pure generative re-hallucination, no memory. Per-scene 4D-GS (~−0.8 to −1.0 dB) is the no-forgetting reference (it stores the whole scene); the explicit 3D-cache generator GEN3C (−17.0 / −13.8 dB) is the best forward method, while the single-shot generators’ −20 to −32 dB collapse is the open problem this axis exposes. Crucially, the dedicated learned spatial-memory model spmem collapses just as hard (−30.0 DL3DV / −24.1 Kubric / −32.0 SpatialVID dB) on all three scene types — learned memory is not yet recalling the observed pose; explicit geometric caching is. Honest scope: these benches are young — 9 methods scored per source now (spmem and FrameCrafter both folded in across all six memory benchmarks, confirming the collapse on static-real, synthetic AND dynamic-real; FlexWorld/TrajCrafter/ViewExtrapolator/Voyager and spmem on the still-generating long-horizon set remain pending); the numbers shown are real-GT means, never fabricated.