Memory Revisit Consistency

This metric evaluates $360^\circ$ revisit rollouts: the model should leave the starting view and return to the same generated scene. It compares protocol-sampled first and last frames and does not verify a geometric revisit, so it only shows visual departure and return. The official closed-loop protocol uses pose-paired memory_revisit_fidelity and memory_return_gate instead.

The metric reads protocol-sampled prediction frames $\{I_t\}_{t=1}^{N}$ ($N\ge 2$) and resizes the last frame to the first-frame size. DINOv2 yields $L_2$-normalized features; cosine similarity is mapped into $[0,1]$ (same as Eq. (1) in memory_revisit_fidelity). Closure uses three components

$$s_{\mathrm{DINO}} = s_{\mathrm{DINO}}(I_1, I_N), \qquad s_{\mathrm{SSIM}} = \mathrm{SSIM}(I_1, I_N), \qquad s_{\mathrm{MSE}} = 1 - \mathrm{clamp}_{[0,1]}\!\left(\frac{\mathrm{MSE}(I_1,I_N)}{\tau_{\mathrm{MSE}}}\right) \tag{1}$$

with default $\tau_{\mathrm{MSE}} = 0.08$. The closure score is a weighted average (defaults $w_{\mathrm{DINO}}=0.55$, $w_{\mathrm{SSIM}}=0.25$, $w_{\mathrm{MSE}}=0.20$):

$$S_{\mathrm{close}} = \frac{ w_{\mathrm{DINO}}\, s_{\mathrm{DINO}} + w_{\mathrm{SSIM}}\, s_{\mathrm{SSIM}} + w_{\mathrm{MSE}}\, s_{\mathrm{MSE}} }{ w_{\mathrm{DINO}} + w_{\mathrm{SSIM}} + w_{\mathrm{MSE}} } \tag{2}$$

To avoid rewarding a static video with a perfect loop, intermediate progress is measured at the $25\%$, $50\%$, and $75\%$ frames using DINO displacement from the first frame, then linearly mapped into $[0,1]$:

$$\delta = \frac{1}{|\mathcal{Q}|}\sum_{q\in\mathcal{Q}} \bigl(1 - s_{\mathrm{DINO}}(I_1, I_q)\bigr), \qquad S_{\mathrm{prog}} = \mathrm{clamp}_{[0,1]}\!\left( \frac{\delta - \ell}{u - \ell} \right) \tag{3}$$

$\mathcal{Q}$ is that tercet by default, with $\ell = 0.03$ and $u = 0.18$. The final score modulates closure by the progress gate (default $\gamma = 0.35$):

$$S_{\mathrm{rev}} = S_{\mathrm{close}} \cdot \bigl((1-\gamma) + \gamma\, S_{\mathrm{prog}}\bigr) \tag{4}$$

A fully static video with perfect first–last closure is therefore capped around $1-\gamma = 0.65$.

In WorldAtlas this algorithm is implemented as the official metric memory_revisit_consistency (dinov2 backend). Under the default normalization config, raw and normalized outputs lie in $[0,1]$ with higher is better. It uses the legacy first/last-frame protocol and must not be substituted for the pose-aligned revisit metric. The key limit is that it does not verify true $360^\circ$ camera geometry.

60-Second Illustrative Examples

Audit manifest: docs/assets/metric_examples/longseq_memory_60s_examples_20260909.json.

Input image

60s memory_loop · illustrative same input

memory-dolly_loop-memory_pool_002

Suite: memory_loopScene: photorealistic / outdoor / city

Benchmark prompt target

Move the camera forward through the scene, then back to where it started.

prompt source: memory_loop_prompt

A coarse first-vs-last check. Prefer the closed-loop family when a true return exists.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Output Amemory_revisit_consistency

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Output Bmemory_revisit_consistency

Input image

60s memory_loop · illustrative same input

memory-square_loop-memory_pool_396

Suite: memory_loopScene: photorealistic / outdoor / terrestrial_landscape

Benchmark prompt target

Walk a closed square around the scene and return to the starting view.

prompt source: memory_loop_prompt

A coarse first-vs-last check. Prefer the closed-loop family when a true return exists.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Output Amemory_revisit_consistency

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Output Bmemory_revisit_consistency

Input image

60s memory_loop · illustrative same input

memory-yaw_loop_180-memory_pool_407

Suite: memory_loopScene: photorealistic / indoor / living_spaces

Benchmark prompt target

Turn the camera to the right, then turn back to the starting view.

prompt source: memory_loop_prompt

A coarse first-vs-last check. Prefer the closed-loop family when a true return exists.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Output Amemory_revisit_consistency

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Output Bmemory_revisit_consistency