Md Jisan Hossen’s JevGAG asks Jev to help choose how a video generator produces its next segment: frame by frame, in larger chunks, or by holding still. It also tests Jev as a judge of completed clips.

The repository separates the quality of that decision signal from whether the supposedly cheaper generation mode actually saves compute.

Images weakened this routing signal

The author releases 80 paired segments evaluated with text, image and hybrid inputs to Jev 1.14. Recomputing the released predictions reproduced these AUROC scores, which measure how well a signal ranks risky segments above safer ones:

  • Difficulty: 0.777 text, 0.667 image, 0.687 hybrid.
  • Predicted drift: 0.755 text, 0.647 image, 0.656 hybrid.
  • Motion category used as a risk score: 0.706 text, 0.510 image, 0.568 hybrid.

The author’s paired uncertainty intervals cross zero for the first two image-versus-text differences. The motion difference’s interval stays below zero. This is evidence about this setup, with these prompts and inputs.

Check the cheaper action before building a router

  • On the author’s recorded A100 runs, changing the coarse mode’s sampling steps reverses its compute advantage over frame-wise generation.
  • The headline adaptive timing comparison uses adaptive_warp, a local motion-based policy. It does not measure a Jev router’s end-to-end speedup.
  • Those figures measure time inside the video-generation network. Remote calls, captioning and decoding require separate accounting.
  • Re-running the released confidence-adjusted simulation showed similarly small changes to the difficulty-only policy. Its rule adds an uncertainty term to difficulty; it is a narrow heuristic test.

A perfect rank correlation has a small denominator

  • The author reports 48 clips grouped into six generation policies for the separate completed-video evaluation.
  • Recomputing the Jev means and comparing them with the released CLIP means reproduced Spearman correlation 1.0 for subject preservation.
  • That is agreement between six policy averages and a machine-derived image-similarity measure. It does not establish per-clip human agreement or calibrated success probabilities.

Verification: released-prediction calculations were reproduced, and the repository’s stored-result checks returned 18/18 PASS. GPU generation and live Jev inference were not repeated.

Several reconstruction scripts depend on a prior project at an author-local path. Treat the release as inspectable experimental evidence with incomplete clean-clone reproduction. Typed output still needs a validated decision rule.