Diagnosing Under-Development of Irreversible Processes in Video Generation

2026-08-01

AI Summary

This work addresses the challenge that current text-to-video generation models struggle to accurately simulate irreversible physical processes—such as ice melting or paper burning—often producing underdeveloped sequences rather than temporally reversed ones. To diagnose this issue, the authors propose a two-stage evaluation protocol that quantifies temporal progression through attribute evolution correlation and stasis rate, complemented by a statistically verifiable diagnostic mechanism based on null hypothesis testing. They further introduce constructive monotonicity constraints in a disentangled attribute latent space, replacing post-hoc guidance methods prone to gaming. Experiments across seven state-of-the-art models reveal that while real videos exhibit an average progression correlation of +0.40 and a stasis rate of 35%, generated videos show near-zero progression and stasis rates of 92%–100%. Human evaluations corroborate these findings, yielding significantly higher scores for real videos (2.75 vs. 0.99), confirming both the prevalence of the problem and the efficacy of the proposed approach.

Abstract

Many physical attributes are \emph{irreversible}: ice melts but does not re-freeze, paper chars but does not un-burn. Do video generators respect this? We show the question is hard to measure, and that what can be measured reliably is \emph{development} rather than reversal. Metrics of local reversal are null-degenerate: a per-clip violation rate scores $0.50$ on pure noise, and a variance-normalized reversal residual sits at its noise ceiling. What survives null-testing is a two-part protocol: progress (a directional attribute correlation) and a stasis rate. Under this protocol, generated video separates cleanly from real footage, and the gap is human-validated. Across seven text-to-video models, real reference footage advances ($ρ{=}{+}0.40$, $35\%$ static) while every generator shows near-zero progress and $92$--$100\%$ stasis; nine annotators rate real footage far above generated ($2.75$ vs.\ $0.99$ on a $0$--$4$ scale). The reliable finding is \emph{under-development}: generators barely advance irreversible attributes rather than reversing them. As a complementary mechanism, we show that post-hoc readout guidance is gameable, whereas enforcing monotonicity by construction in a disentangled attribute latent removes the gameable readout, validated in controlled and semi-synthetic settings.

Problem

• irreversible processes
• video generation
• under-development
• attribute progression
• temporal consistency

Innovation

• irreversible processes
• video generation
• monotonicity by construction
• attribute disentanglement
• progress-stasis protocol