Concyclics's picture
Add 30,492 Cosmos3-Super benchmark videos across nine methods
8cd359f verified
|
Raw History Blame Contribute Delete
6.64 kB

Cosmos3-Super complete benchmark videos

30,492 videos: 650 RBench scenes + 1,044 PAI-Bench-G scenes, nine methods, two fixed seeds per scene. This adds Super to the existing Edge/Nano release. The original 320 × 192, 61-frame, 24 FPS, 35-step, CFG 6 exports are copied without re-encoding. Each package is independently usable. Total ZIP size: 11,294,910,657 bytes (11.295 GB).

Benchmark Method Package Videos Size
paibench BF16 super_paibench_bf16.zip 2,088 0.722 GB
paibench FP8 super_paibench_fp8.zip 2,088 0.768 GB
paibench NVFP4-AWQ W4A16 reference super_paibench_awq_reference.zip 2,088 0.734 GB
paibench Plain W4A4 super_paibench_fp4_base.zip 2,088 0.925 GB
paibench ARCQuant 50% super_paibench_arcquant.zip 2,088 0.780 GB
paibench ARCQuant 100% super_paibench_arcquant100.zip 2,088 0.747 GB
paibench SharQ super_paibench_sharq_perf.zip 2,088 0.782 GB
paibench ThunderWorld 1.5× super_paibench_b1.5_protected_n.zip 2,088 0.833 GB
paibench ThunderWorld 2× super_paibench_b2_joint.zip 2,088 0.772 GB
rbench BF16 super_rbench_bf16.zip 1,300 0.421 GB
rbench FP8 super_rbench_fp8.zip 1,300 0.462 GB
rbench NVFP4-AWQ W4A16 reference super_rbench_awq_reference.zip 1,300 0.428 GB
rbench Plain W4A4 super_rbench_fp4_base.zip 1,300 0.579 GB
rbench ARCQuant 50% super_rbench_arcquant.zip 1,300 0.468 GB
rbench ARCQuant 100% super_rbench_arcquant100.zip 1,300 0.439 GB
rbench SharQ super_rbench_sharq_perf.zip 1,300 0.473 GB
rbench ThunderWorld 1.5× super_rbench_b1.5_protected_n.zip 1,300 0.502 GB
rbench ThunderWorld 2× super_rbench_b2_joint.zip 1,300 0.459 GB

Download and verify

hf download Concyclics/thunderworld-benchmark --type dataset --include 'super/*' --local-dir thunderworld-benchmark
cd thunderworld-benchmark/super
sha256sum -c SHA256SUMS
unzip super_rbench_b2_joint.zip
cd super_rbench_b2_joint
sha256sum -c SHA256SUMS

To download one method, replace the include pattern with super/*_b2_joint.*. packages.json records exact archive sizes, counts, hashes and validation results. sample_index.jsonl.gz maps all 30,492 samples to archives and internal video/metadata paths without downloading the ZIPs. Each archive contains full judge_index.jsonl and method-label-free judge_inputs.jsonl, with paths relative to its extracted directory.

Contents and protocol

Every package includes original MP4 exports, conditioning images, exact prompts/seeds, original per-video metadata, generation receipts and referenced generation logs. Frozen model/method definitions and export provenance are included under provenance/. The nine methods are BF16, native W8A8 FP8, plain W4A4, NVFP4-AWQ W4A16 quality reference, ARCQuant 50%, ARCQuant 100%, SharQ, ThunderWorld 1.5× and ThunderWorld 2×. These are the study's Cosmos adaptations; method_spec and frozen model manifests define the actual execution. AWQ is a decoded quality reference, not a qualified native-kernel speed result. Full-corpus results include previous development scenes; they are not held-out scene generalization. Generation/judging completed on 2026-10-10.

All MP4/input bytes were checked against export hashes. Every archived member was read back for SHA256/ZIP-CRC verification; one video per method and benchmark was decoded for dimensions, frame count and FPS (18 representatives). Packaging did not change the generation or encoding. The original logs include shared-device timings; these are not exclusive-GPU performance measurements.

MP4 uses H.264/yuv420p CRF15 and is lossy. Super NPZ, lossless MKV, model weights and AR tensor caches are not included in these video packages. The existing repository npz/ contains the earlier Edge/Nano inventory, not Super NPZ. Use original floating-point/lossless outputs for strict pixel consistency metrics. Upstream benchmark-input and model terms remain applicable.

Code · Paper · Models