Jev-Routed Temporal Granularity with Recovery and Hysteresis for Causal Video Generation

Hossen Md Jisan · Department of Artificial Intelligence, Nanjing University of Information Science and Technology (NUIST), Nanjing, China

GitHub repository Manuscript (PDF) Result figures Result videos

Overview

Causal video generators produce a rollout one temporal chunk at a time, which invites a simple efficiency idea: spend fine per-frame computation only where a segment is hard, and use cheaper coarse chunks, or even reuse, elsewhere. Realizing this idea requires three preconditions that are usually assumed rather than measured: a mode that is genuinely cheaper on the deployment hardware, a signal that predicts hard segments before generation, and an independent endpoint that certifies the trade. This project audits all three on Causal-rCM, a Wan2.1-1.3B text-to-video causal diffusion model, pairing a frame-wise mode with a chunk-3 mode and a near-zero-cost hold mode, and routing among them with a two-threshold controller with hysteresis and forced recovery. The audit runs on a Tesla T4 and an A100-SXM4-80GB, because the cost precondition turns out to be a property of the hardware, not of the model. Negative and failed experiments are preserved as evidence.

Key findings

Hardware-conditional cost inversion. On the T4 each coarse forward is 3.01× a fine forward against a 2.33× break-even point, so coarse is never cheaper; on the A100 the ratio falls to 1.60–1.75 and coarse is 1.46× faster in DiT time.
A preregistered hand-crafted signal fails; a text-only oracle succeeds. The locked residual signal reaches AUROC 0.491, whereas the Jev System One oracle reaches AUROC 0.788 [0.575, 0.946] at 0.67 s per decision, off the generation path.
Recovery and hysteresis stabilize switching. Switches drop from 0.50 to 0.38 on the T4 and from 0.92 to 0.50 under signal noise on the A100, without worsening regret.
A trade-off, not dominance. On 24 fresh A100 prompts a single-decision Jev policy saves 1.78 s of DiT time and 7.39 s of total time (both significant) at a small significant −0.020 CLIP subject-consistency cost.
Monotonicity is signal- and endpoint-dependent. The fine-versus-coarse benefit is monotone in the Jev difficulty signal on the T4 but flat on the A100 diagnostic split.

Results

Rollout DiT latency and the coarse/fine per-forward ratio against the 2.33 break-even point on T4 and A100. — Cost precondition across hardware. The same model pair lies above break-even on the T4 (coarse slower) and below it on the A100 (coarse faster).

Cost precondition across hardware and context regime.
Setting	fine s/fwd	coarse s/fwd	ratio	coarse cheaper?
T4, clean context	1.32	3.58	3.01	no
T4, noisy context	0.71	2.10	2.94	no
A100, timing arm	0.125	0.200	1.60	yes
A100, full-decode arm	0.112	0.196	1.75	yes

Six-mode cost-quality sweep on the Tesla T4 and the A100-80GB. — Six-mode cost–quality sweep. Fewer NFEs do not imply lower DiT time: on the T4 every coarse mode sits above and to the right of fine c1-1 s1, while on the A100 the coarse family occupies a cheaper, less accurate frontier point.

Held-out AUROC of hand-crafted signals versus Jev outputs. — Signal discrimination. The Jev outputs outperform every hand-crafted signal; the dashed line marks the preregistered residual result (0.491) and the dotted line is chance.

Per-segment Jev routing traces over eight prompts. — Jev routing traces (F = fine, C = coarse, H = hold). Static content is held, fast motion is routed to fine, and moderate motion to coarse.

Controller ablation: switch counts and seam cost on T4 and A100. — Controller ablation. Recovery plus hysteresis cuts switches on both GPUs and reduces the noisy-arm seam cost at switch boundaries.

Prompt-paired bootstrap differences against fixed fine on 24 fresh A100 prompts. — Confirmatory study. Both Jev policies pay a small, significant subject-consistency cost; the DiT and total-time savings are significant for the single-decision policy.

Representative decoded frames from matched rollouts across policies. — Representative decoded frames from matched rollouts. Fast-motion prompts preserve the subject under fine and Jev-routed policies, while the coarse rollout drifts.

Citation

@article{jisan2026jev,
  title   = {Jev-Routed Temporal Granularity with Recovery and Hysteresis for Causal Video Generation},
  author  = {Jisan, Hossen Md},
  year    = {2026},
  note    = {Manuscript prepared for IEEE Transactions on Circuits and Systems for Video Technology}
}