long_sequence_motion_smoothness measures long-horizon temporal interpolatability of a prediction video. The long_sequence backend splits the full video into short clips. AMT-S then interpolates odd frames from even-sampled neighbors. Low reconstruction error means the clip still moves continuously after a long rollout—not that the motion is semantically correct or the right magnitude.
s s m o o t h = clamp ( 255 − mean ( ∣ I m i d r e a l − I m i d i n t e r p ∣ ) 255 , 0 , 1 ) (1) s_{\mathrm{smooth}}
=
\operatorname{clamp}
\left(
\frac{255-\operatorname{mean}(|I^{\mathrm{real}}_{\mathrm{mid}}-I^{\mathrm{interp}}_{\mathrm{mid}}|)}{255},
0,
1
\right)
\tag{1} s smooth = clamp ( 255 255 − mean ( ∣ I mid real − I mid interp ∣ ) , 0 , 1 ) ( 1 ) The video score is the mean of clip scores. Higher is better. Prediction-only; no reference rollout. On memory_loop this is an official metric. It is not the short-horizon motion_smoothnessLong-Sequence Metrics .
Audit manifest: docs/assets/metric_examples/longseq_memory_60s_examples_20260909.json.
The scores below come from diagnostic 60-second evaluation reports.
Input image
60s memory_loop · same benchmark case · gap 0.0363
memory-dolly_loop-memory_pool_368Suite: memory_loop Scene: photorealistic / indoor / dining_spaces
Benchmark prompt target
Move the camera forward through the scene, then back to where it started.
prompt source: memory_loop_prompt
The smoother 60s rollout interpolates cleanly between neighboring frames; the weaker one jitters or tears across clip boundaries.
Higher metric score Score 0.9930 · diagnostic
Lower metric score Score 0.9567 · diagnostic
Input image
60s memory_loop · same benchmark case · gap 0.0622
memory-double_yaw_loop-memory_pool_029Suite: memory_loop Scene: photorealistic / outdoor / terrestrial_landscape
Benchmark prompt target
Turn away and back to the start twice in a row.
prompt source: memory_loop_prompt
The smoother 60s rollout interpolates cleanly between neighboring frames; the weaker one jitters or tears across clip boundaries.
Higher metric score Score 0.9874 · diagnostic
Lower metric score Score 0.9252 · diagnostic
Input image
60s memory_loop · same benchmark case · gap 0.0617
memory-yaw_loop_360-memory_pool_028Suite: memory_loop Scene: photorealistic / outdoor / city
Benchmark prompt target
Rotate the camera a full turn in place and come back to the starting view.
prompt source: memory_loop_prompt
The smoother 60s rollout interpolates cleanly between neighboring frames; the weaker one jitters or tears across clip boundaries.
Higher metric score Score 0.9940 · diagnostic
Lower metric score Score 0.9323 · diagnostic
Input image
60s memory_loop · same benchmark case · gap 0.0614
memory-yaw_loop_180-memory_pool_037Suite: memory_loop Scene: photorealistic / outdoor / verdant_landscape
Benchmark prompt target
Turn the camera to the right, then turn back to the starting view.
prompt source: memory_loop_prompt
The smoother 60s rollout interpolates cleanly between neighboring frames; the weaker one jitters or tears across clip boundaries.
Higher metric score Score 0.9894 · diagnostic
Lower metric score Score 0.9280 · diagnostic
Input image
60s memory_loop · same benchmark case · gap 0.0609
memory-palindrome_loop-memory_pool_003Suite: memory_loop Scene: photorealistic / outdoor / terrestrial_landscape
Benchmark prompt target
Move forward, turn and continue, then retrace the whole path back.
prompt source: memory_loop_prompt
The smoother 60s rollout interpolates cleanly between neighboring frames; the weaker one jitters or tears across clip boundaries.
Higher metric score Score 0.9934 · diagnostic
Lower metric score Score 0.9325 · diagnostic
AMT: All-Pairs Multi-Field Transforms for Efficient Frame Interpolation
Zhen Li, Zuo-Liang Zhu, Litong Han, Qibin Hou, Chun-Le Guo, Ming-Ming Cheng.https://arxiv.org/abs/2304.09790
