doanh25032004/video-gen-physics-gallery

Paper material (assets/drive)

Two generated galleries plus a figure asset pack. All reproducible from the repo — don't hand-edit them, re-run the generator.

gallery	what it shows	generator
real_videos/	the input side of the benchmark: real GT episodes, 5 diverse tasks per (embodiment x view-track x markovian/non-markovian) cell	scripts/make_real_video_gallery.py
method_comparison/	the output side: ONE episode rendered by every acceleration method, so a single dir is a like-for-like comparison	scripts/make_method_comparison_gallery.py
figure_drawio/	teaser figure tiles: one PNG per camera slot (single-view + multi-view row per embodiment, same episode & frame within a row) — for assembling the "Dataset Collection" block in draw.io	scripts/make_figure_slots.py
figure_tasks/	per-task image folders: one dir per task, 5 uniformly-spaced timeline frames + strip.png + prompt.txt, covering all 14 measured skills across 3 embodiments — for qualitative / skill figures	scripts/make_task_image_pack.py
figure_tasks_humanoid_short/	humanoid, shortest real prompts (2 single-view + 2 multi-view): same layout as figure_tasks/ plus prompt_short.txt, for captions that must not wrap	scripts/make_humanoid_short_prompt_pack.py

Both ship pre-cut stills so figures can be assembled without touching the mp4s:

• real_videos/*/frames/<camera>_fNNNNN.png + <camera>_strip.png (filmstrip)
• method_comparison/*/frames/<method>_fNNNNN.png + compare.png (rows = methods, columns = the same rollout position)

Each clip dir also carries an info.json, and each gallery a manifest.json (machine readable) plus INDEX.md (human readable, with the per-cell tables).

Regenerate

python3 scripts/make_real_video_gallery.py            # ~675 MB
python3 scripts/make_method_comparison_gallery.py     # ~1.9 GB
python3 scripts/make_figure_slots.py --square         # ~7 MB
python3 scripts/make_task_image_pack.py               # ~25 MB
python3 scripts/make_humanoid_short_prompt_pack.py    # ~9 MB

Both are incremental — existing videos/stills are reused, so a re-run after changing only the sheet/index code is cheap. Add --dry-run to see the selection without writing anything.

Publish to the Hub

Mirrored to doanh25032004/video-gen-physics-gallery (dataset repo). Re-sync after regenerating:

HF_TOKEN=... python3 scripts/upload_drive_to_hf.py

That script commits in small batches and skips whatever is already on the Hub, so it is safe to re-run / resume. Do NOT use hf upload-large-folder here — it stalls on this tree (5300 small files).