curvedinf/exodus-18m-training/decoder-video-gen-src - Storage Bucket

Exodus 18M Decoder Progression Video Bundle

This bundle is checkpoint-free and contains only:

• Cached projection data (uv_step*.npy, c4_step*.npy)
• Forward-derived completion text cache
• Training log (for step-wise val-loss overlay)
• Music track used in the published render
• Standalone render script

No model checkpoints are included.

Directory Layout

• data/cache/: projection caches used for rendering
• data/completions.json: per-step completion text
• data/training_log.txt: log used for step-gated val_loss overlay
• data/music/hydrostatic_mind.mp3: soundtrack
• scripts/render_video_from_cache.py: standalone renderer
• outputs/: generated videos + metadata

Python Dependencies

Install dependencies:

python3 -m pip install -r requirements.txt

You also need ffmpeg available on PATH.

Reproduce Default Video

From this bundle root:

python3 scripts/render_video_from_cache.py \
  --cache-dir data/cache \
  --completion-cache-json data/completions.json \
  --training-log data/training_log.txt \
  --music data/music/hydrostatic_mind.mp3 \
  --output outputs/subspace_projection_progression_video.mp4

Or use the helper script:

./scripts/render_default.sh

Defaults reproduce:

• 1920x1080
• 60 fps
• 158s transition + 5s hold
• text overlays (tokens trained, step-gated val-loss, completion line)
• music at 80% volume with 5s outro fade

Useful Variants

Higher bitrate:

python3 scripts/render_video_from_cache.py --video-bitrate 56M --output outputs/video_56mbps.mp4

Lossless 4:4:4:

python3 scripts/render_video_from_cache.py \
  --video-bitrate \"\" \
  --ffmpeg-crf 0 \
  --ffmpeg-pix-fmt yuv444p \
  --output outputs/video_lossless444.mp4

Delete frame PNGs after encode:

python3 scripts/render_video_from_cache.py --cleanup-frame-pngs

Notes

• The script renders from projection caches only (no checkpoints needed).
• Frame PNGs are kept by default for auditability/re-encoding.
• Metadata JSON is written alongside output video.