j1s4nn/ddag-matched-frames

DDAG Matched Granularity Frame Comparisons

Matched fine, coarse and seam-repaired decoded frame comparisons

DDAG Matched Granularity Frame Comparisons

Qualitative frames for DDAG (JevGAG + DDAG programme): matched fine (c1-1, 2 steps), coarse (c3-3, 2 steps) and seam-repaired (blend 0.25) segments generated from the same boundary and noise on a causal diffusion video model, with the composed grid figure.

The modes produce visibly different content, yet no independent endpoint ranks the coarse mode worse - a key finding for adaptive temporal granularity in causal video generation and video world models.

Contents

• fig_qualitative.png: composed comparison grid
• frames/: per-prompt fine / coarse / repair decoded frames

Related resources

• Code (JevGAG programme): https://github.com/j1s4nn/JevGAG
• Code (Jev-Routed Temporal Granularity): https://github.com/j1s4nn/jev-adaptive-granularity-WM
• Project page (JevGAG): https://j1s4nn.github.io/JevGAG/
• Project page (Paper 1): https://j1s4nn.github.io/jev-adaptive-granularity-WM/
• Portfolio: https://j1s4nn.github.io
• Source (Kaggle): https://www.kaggle.com/datasets/ajjisan/ddag-matched-frames
• Preprint (JevGAG): https://doi.org/10.5281/zenodo.23693517
• Preprint (Jev-Routed Temporal Granularity): https://doi.org/10.5281/zenodo.23395414

Citation

@misc{hossen_jevgag_ddag,
  author       = {Hossen, MD Jisan},
  title        = {Regime-Aware Temporal Granularity for Causal Video World Models: When Does Adaptivity Pay?},
  year         = {2026},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.23693517},
  url          = {https://doi.org/10.5281/zenodo.23693517}
}

License

CC-BY-4.0. Maintained by Vision Intelligence (MD Jisan Hossen (NUIST)).