Measures maximum chronological relative drawdown in LAION aesthetic (pyiqa laion_aes) scores. Scores are grouped in temporal order (default 10 segments; equal-time bins when timestamps are known). For segment means q j q_j q j :
p j = max ( q 0 , … , q j ) (1) p_j = \max(q_0,\ldots,q_j)
\tag{1} p j = max ( q 0 , … , q j ) ( 1 )
d r i f t = max j p j − q j max ( p j , ε ) (2) \mathrm{drift}
=
\max_j \frac{p_j - q_j}{\max(p_j,\,\varepsilon)}
\tag{2} drift = j max max ( p j , ε ) p j − q j ( 2 )
Lower is better, within [ 0 , 1 ] [0, 1] [ 0 , 1 ] , with ε = 10 − 12 \varepsilon = 10^{-12} ε = 1 0 − 12 . Input scores must be finite and nonnegative, with at least two time points and two temporal segments. Improving quality has zero degradation; a decline is penalized. A dip followed by recovery still records drawdown, while signed endpoint change and slope describe recovery separately.
Mean absolute quality is also reported: constant low quality is not high quality. Nonfinite observations are not silently removed to compress time. If FPS is unknown, chronological drift is available but per-second slope is null. This is an official quality metric; it does not measure memory. Backend: quality_drift.
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_190
Suite: memory_loop Scene: stylized / outdoor / terrestrial_landscape
Benchmark prompt target
Move the camera forward through the scene, then back to where it started.
prompt source: memory_loop_prompt
Watch aesthetic quality over the 60s rollout. A late collapse is drawdown; getting better by itself is not a penalty.
Output A rollout_aesthetic_drift
Output B rollout_aesthetic_drift
Input image
60s memory_loop · illustrative same input
memory-yaw_loop_360-memory_pool_132
Suite: memory_loop Scene: stylized / outdoor / terrestrial_landscape
Benchmark prompt target
Rotate the camera a full turn in place and come back to the starting view.
prompt source: memory_loop_prompt
Watch aesthetic quality over the 60s rollout. A late collapse is drawdown; getting better by itself is not a penalty.
Output A rollout_aesthetic_drift
Output B rollout_aesthetic_drift
Input image
60s memory_loop · illustrative same input
memory-yaw_loop_180-memory_pool_439
Suite: memory_loop Scene: stylized / outdoor / aquatic_landscape
Benchmark prompt target
Turn the camera to the right, then turn back to the starting view.
prompt source: memory_loop_prompt
Watch aesthetic quality over the 60s rollout. A late collapse is drawdown; getting better by itself is not a penalty.
Output A rollout_aesthetic_drift
Output B rollout_aesthetic_drift
Input image
60s memory_loop · illustrative same input
memory-palindrome_loop-memory_pool_335
Suite: memory_loop Scene: stylized / indoor / dining_spaces
Benchmark prompt target
Move forward, turn and continue, then retrace the whole path back.
prompt source: memory_loop_prompt
Watch aesthetic quality over the 60s rollout. A late collapse is drawdown; getting better by itself is not a penalty.
Output A rollout_aesthetic_drift
Output B rollout_aesthetic_drift
Input image
60s memory_loop · illustrative same input
memory-double_yaw_loop-memory_pool_414
Suite: 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
Watch aesthetic quality over the 60s rollout. A late collapse is drawdown; getting better by itself is not a penalty.
Output A rollout_aesthetic_drift
Output B rollout_aesthetic_drift
LAION-Aesthetics Predictor (CLIP-based linear aesthetic regressor)
Used via pyiqa laion_aes; CLIP ViT-L/14 features with a linear head trained on human aesthetic ratings.