README.md

metadata

license: unknown
task_categories:
  - text-to-video
tags:
  - video-generation
  - distillation
  - dmd2
  - repa

repa_vs_dmd2

Qualitative comparison of two 4-step distilled video generators: a DMD2 baseline and the same setup with a REPA loss added on top of the DMD2 objective.

Both models were sampled on the prompts of MovieGen Video Bench (extended), with four seeds per prompt, so the two archives are seed- and prompt-aligned: the same file name in each archive corresponds to the same prompt and the same seed.

Contents

File	Method	Videos
video_Repa_dmd2.tar.gz	DMD2 loss (baseline)	280
video_Repa_1k4.tar.gz	DMD2 + REPA loss	280

Each archive extracts to:

video_Repa_{dmd2,1k4}/
  MovieGenVideoBench_extended/
    student_step4_{prompt_index}_seed_student_{seed}.mp4

• prompt_index — 0-based index of the prompt in the MovieGen Video Bench (extended) prompt list (line prompt_index + 1 of the prompt file).
• seed — 1 through 4.
• student_step4 — the 4-step student sampler.

The prompt list itself is not included in this dataset; index into your own copy of MovieGenVideoBench_extended.txt to recover the text for a given prompt_index.

Usage

huggingface-cli download csusupergear/repa_vs_dmd2 --repo-type dataset --local-dir .
tar -xzf video_Repa_dmd2.tar.gz
tar -xzf video_Repa_1k4.tar.gz

To compare a single prompt across both methods:

# prompt 17, all four seeds, both methods
ls video_Repa_dmd2/MovieGenVideoBench_extended/student_step4_17_seed_student_*.mp4
ls video_Repa_1k4/MovieGenVideoBench_extended/student_step4_17_seed_student_*.mp4

Notes

The archives are gzipped tarballs of already-compressed MP4 files, so gzip provides essentially no size reduction — they exist to bundle 280 files into one download, not to save space. The video bitstreams are byte-identical to the originals.