Nicholas0228/PACC-Website-Videos

PACC website videos

Original generated videos used by the project page for Compress to Remember: Learning Compact Memory via On-Policy Distillation for Long Video Generation.

This release contains 40 MP4s (187,088,827 bytes): 12 selected MBench cases × three methods (Full AR, Relax Forcing, PACC), plus four minute-long PACC MovieGen videos used in VBench evaluation. These are curated illustrations, not the complete evaluation suite or a representative random sample. Full AR is an unscored diagnostic with a larger context budget.

All PACC outputs use the backbone-specific CR10, checkpoint 5,000 compressor, original sparse Relax bootstrap, two raw sink frames, one complete selected compressed block including residual tokens, and one recent raw frame. The generator is frozen. Both backbones use their own native four-step sampler and three-latent chunks. All clips in this release use seed 0.

Collection	Count	Resolution	Frames / FPS	Duration
MBench: six CF cases and six Causal-rCM cases, three methods each	36	832 × 480	417 / 16	26.0625 s
MovieGen: two CF and two Causal-rCM videos	4	832 × 480	957 / 16	59.8125 s

The minute-long examples are mammoths (CF, prompt 1), Pacific coast (Causal-rCM, prompt 3), ships in a coffee cup (CF, prompt 7), and neon street (Causal-rCM, prompt 0). The coast example visibly loses detail late in the video; it is preserved here as the exact current website selection pending further curation. Publication does not enhance or re-encode the videos.

Inventory and reproduction

• videos.json: all 40 filenames, full prompts, original case/prompt IDs, backbones, methods, checkpoints, seeds, geometry, source run names, byte sizes, and SHA-256 hashes.
• SHA256SUMS: checksums of the original MP4s.
• prompts/: exact caption JSON files for the 16 examples.
• Website source and instructions: access follows the source repository's permissions.
• Generation instructions: pinned assets and portable generation commands.

Use a commit-pinned /resolve/<commit>/videos/<filename>.mp4 URL for playback. main can advance when future examples are added; old pinned releases remain reproducible. These are synthetic research outputs. Consult the paper and the original model/dataset terms for provenance and permitted uses.