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

Abstract

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 deposit contains the official code, locked protocols, calibration freezes, and all audit artifacts for the manuscript "Jev-Routed Temporal Granularity with Recovery and Hysteresis for Causal Video Generation." The work audits all three preconditions end to end on Causal-rCM, a Wan2.1-T2V-1.3B-480p causal text-to-video diffusion model, pairing a frame-wise mode (c1-1) with a chunk-3 mode (c3-3) and a near-zero-cost hold mode, and routing among them with a two-threshold controller with hysteresis and forced recovery. Routing is driven by the Jev System One oracle, a text-only decision model that answers typed difficulty, drift-risk, and motion-class questions before a segment is generated. All audits run on two GPUs with different compute characters — a 16 GB Tesla T4 and an A100-SXM4-80GB — because the cost precondition turns out to be a property of the hardware. The study is deliberately evidence-preserving: negative and failed experiments are reported as evidence, and no claim of universal adaptive superiority is made. The value is a reproducible account of when the three preconditions hold, what breaks when they do not, and a working routing mechanism for the regime in which they do. What this deposit contains Manuscript: Jev-Routed-Temporal-Granularity-with-Recovery-and-Hysteresis.pdf — the standalone paper (also included inside the ZIP under paper/). Replication artifact: zenodo_replication_artifacts.zip — the complete experimental repository with relative structure preserved: src/ — core Python packages (causal-rCM inference, controller, uncertainty, evaluation, baselines, utils) scripts/ — 50 audit, calibration, and figure-generation scripts colab/ — A100 experiment runners jobs/ — 34 reproducible experiment job definitions (phase 0 through phase 5c, plus RQ3/RQ4) config/ — locked manifests and calibration freezes (JSON/YAML) results/ — evaluation logs and summary data (JSON / JSONL / CSV) for every phase and research question paper/ — LaTeX source, bibliography, manuscript figures, and the integrity audit report Top-level README.md, requirements.txt, LICENSE, CITATION.cff, and protocol/status notes Key findings Cost is hardware-conditional (RQ2). 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 diffusion-transformer (DiT) time. The same model pair lies on opposite sides of break-even on different GPUs. The preregistered hand-crafted signal fails; a text-only oracle succeeds (RQ1). On 72 locked held-out cases the hand-crafted residual signal reaches AUROC 0.491, whereas the Jev difficulty output reaches AUROC 0.788 [0.575, 0.946] at 0.67 s per decision, computed entirely off the generation path. Recovery and hysteresis stabilize switching (RQ3). Switches drop from 0.50 → 0.38 on the T4 and from 0.92 → 0.50 under injected signal noise on the A100, without worsening regret; switching carries a directly measured seam cost. The result is a trade-off, not dominance (RQ2). 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 (RQ4). The fine-versus-coarse benefit is monotone in the Jev difficulty signal on the T4 but flat on the A100 diagnostic split, tying RQ4 to the choice of endpoint. Method (summary) Rollouts use a fixed grid (a shared coarse prefix, then segments), and before each segment the controller selects an action from {fine c1-1, coarse c3-3, hold}. The Jev oracle is queried once per decision with a composite typed prompt built from the text prompt, a BLIP caption of the pre-decision context frame, the segment index, and recent motion/warp statistics. A two-threshold controller with a hysteresis band and a forced recovery window maps the signal to an action. Every decision is recorded with its generated frames, independent quality measures, timing, and switch-boundary cost; all measurements are prompt-paired, and the internal fine-family reference is used for diagnosis only, never for the headline claims. Hardware and environment Audits run on a Kaggle Tesla T4 (16 GB, PyTorch SDPA attention) and a Google Colab A100-SXM4-80GB (FlashAttention-2). Inference is fp16 at 480×832, 16 fps, with the released c1-1 and c3-3 Causal-rCM checkpoints. Python dependencies are listed in requirements.txt. A100 runs pin upstream commit ed3cb14dd936f92cdc9f9381af7369991509b41f. External data and exclusions To keep the artifact lightweight and free of sensitive content, the deposit excludes model weights, raw video renders, virtual environments, caches, and secrets. The complete result-video dataset is hosted externally on Kaggle (see the repository links below); this deposit retains only static figures and the numeric audit artifacts needed to reproduce every reported result. Reproducibility Every reported number is traceable to a locked protocol, a calibration freeze, and a recorded audit artifact. The repository contains the exact scripts and job definitions that generated each result, and negative results are preserved alongside positive ones. See the repository README.md for step-by-step reproduction instructions. Links GitHub repository: https://github.com/j1s4nn/jev-adaptive-granularity-WM Project page: https://j1s4nn.github.io/jev-adaptive-granularity-WM/ Result figures (Kaggle): https://www.kaggle.com/datasets/ajjisan/jev-adaptive-granularity-figures Result videos (Kaggle): https://www.kaggle.com/datasets/ajjisan/jev-adaptive-granularity-videos Citation @software{Jisan_JevRouted_Temporal_Granularity_2026, author = {Hossen Md Jisan}, title = {Jev-Routed Temporal Granularity with Recovery and Hysteresis for Causal Video Generation}, year = {2026}, version = {1.0.0}, publisher = {Zenodo}, note = {Department of Artificial Intelligence, Nanjing University of Information Science and Technology (NUIST), Nanjing, China}, url = {https://github.com/j1s4nn/jev-adaptive-granularity-WM} } License Released under the MIT License. Copyright © 2026 Hossen Md Jisan.