EVIDENCE / UPDATED 2026-11-15

AI image and video watermarking: CPU benchmarks

Measured 2026-11-15 · CJ-05. These are matched engineering-fixture measurements, not production throughput, visual-quality, durability, FPR or legal-conformance estimates.

Runtime and method

The corpus consists of two procedural PNG/JPEG images and seven silent SDR video fixtures made with FFmpeg testsrc2: 720p/1080p, 5/10/15 seconds, 24/30/30000:1001 fps, plus HEVC input. A separate 60-second clean clip checks the scan budget. Commands and source SHA-256 values are in CORPUS.json.

Complete export

Warm values are medians; brackets contain the observed min–max, n=3. Cold is a single first-job observation, n=1. CPU is CPU-seconds, not elapsed seconds or VM billed time.

Input Mode Cold wall s Warm wall s [range] Warm total CPU s Max warm kernel peak MiB
i1024_png Image 4.50 0.61 [0.61–0.63] 0.98 492.04
i2048_jpeg Image 4.63 0.64 [0.64–0.65] 1.02 619.47
v1080_10_30 Economy2s 28.31 25.23 [25.13–25.51] 35.78 1115.94
v1080_10_30 Economy5s 19.72 17.48 [17.18–17.66] 23.80 1180.43
v1080_10_30 Resilient 188.85 165.48 [164.91–165.85] 254.85 1229.31
v1080p_15_ntsc Economy2s 41.77 37.87 [37.43–38.24] 54.49 1273.50
v1080p_15_ntsc Economy5s 28.76 26.47 [26.32–26.58] 36.01 1268.01
v1080p_15_ntsc Resilient 288.51 258.59 [256.37–261.99] 392.82 1328.71
v1080p_5_24 Economy2s 14.00 11.65 [11.36–11.75] 16.42 1264.23
v1080p_5_24 Economy5s 9.44 7.45 [7.25–7.56] 10.01 1038.82
v1080p_5_24 Resilient 81.91 67.22 [67.16–67.35] 104.38 1238.08
v720_10_30 Economy2s 15.67 13.85 [13.82–13.87] 21.49 757.91
v720_10_30 Economy5s 10.72 9.16 [9.13–9.19] 13.16 726.93
v720_10_30 Resilient 115.75 107.64 [107.59–107.66] 189.35 793.12
v720_10_hevc Economy2s 15.74 14.11 [13.96–14.14] 22.38 757.03
v720_10_hevc Economy5s 11.08 9.52 [9.48–9.57] 14.15 733.16
v720_10_hevc Resilient 115.85 111.02 [110.75–111.21] 194.90 814.50
v720_15_ntsc Economy2s 23.08 21.16 [21.07–21.22] 33.34 793.58
v720_15_ntsc Economy5s 15.24 13.55 [13.55–13.81] 19.70 771.13
v720_15_ntsc Resilient 169.74 155.82 [155.37–157.11] 274.55 906.97
v720_5_24 Economy2s 8.00 6.41 [6.40–6.41] 9.91 766.65
v720_5_24 Economy5s 5.64 4.00 [3.99–4.17] 5.61 707.69
v720_5_24 Resilient 47.91 44.32 [44.27–44.39] 77.70 780.70

Complete video export

Whole-container CPU and memory

Independent verification

The primary scope is explicitly current, matching a new issuer; it does not establish absence across every historical algorithm. Video positives below use Economy5s outputs; all profile measurements are retained in the CSV.

Input Check Cold wall s Warm wall s [range] Warm total CPU s Outcomes over cold + warm
i1024_png verify_positive 6.09 0.16 [0.16–0.17] 0.26 {"correct_id": 4}
i1024_png verify_metadata_loss 4.03 0.14 [0.14–0.14] 0.24 {"correct_id": 4}
i1024_png verify_negative 4.13 0.13 [0.13–0.13] 0.24 {"completed_negative": 4}
i2048_jpeg verify_positive 4.13 0.21 [0.20–0.21] 0.30 {"correct_id": 4}
i2048_jpeg verify_metadata_loss 4.08 0.19 [0.19–0.19] 0.31 {"correct_id": 4}
i2048_jpeg verify_negative 4.21 0.18 [0.18–0.18] 0.29 {"completed_negative": 4}
v720_5_24 verify_positive 6.76 1.70 [1.69–1.79] 2.66 {"correct_id": 4}
v720_5_24 verify_metadata_loss 6.72 1.68 [1.67–1.81] 2.64 {"correct_id": 4}
v720_5_24 verify_negative 25.28 18.36 [18.12–18.73] 32.53 {"completed_negative": 4}
v720_10_30 verify_positive 7.12 2.04 [2.03–2.21] 3.07 {"correct_id": 4}
v720_10_30 verify_metadata_loss 7.07 2.01 [2.00–2.07] 3.05 {"correct_id": 4}
v720_10_30 verify_negative 50.31 40.32 [40.27–40.67] 73.15 {"completed_negative": 4}
v720_15_ntsc verify_positive 7.29 2.39 [2.29–2.57] 3.50 {"correct_id": 4}
v720_15_ntsc verify_metadata_loss 7.35 2.26 [2.25–2.30] 3.39 {"correct_id": 4}
v720_15_ntsc verify_negative 70.85 59.36 [59.30–60.07] 108.24 {"completed_negative": 4}
v1080p_5_24 verify_positive 8.02 2.90 [2.86–3.02] 4.09 {"correct_id": 4}
v1080p_5_24 verify_metadata_loss 8.16 2.88 [2.83–4.82] 4.08 {"correct_id": 4}
v1080p_5_24 verify_negative 48.17 40.23 [40.14–40.27] 57.49 {"completed_negative": 4}
v1080_10_30 verify_positive 8.92 3.65 [3.59–3.73] 5.02 {"correct_id": 4}
v1080_10_30 verify_metadata_loss 8.63 3.68 [3.62–3.73] 5.08 {"correct_id": 4}
v1080_10_30 verify_negative 101.10 87.36 [85.98–87.57] 127.42 {"completed_negative": 4}
v1080p_15_ntsc verify_positive 9.48 4.24 [4.22–4.24] 5.81 {"correct_id": 4}
v1080p_15_ntsc verify_metadata_loss 9.55 4.26 [4.21–4.28] 5.85 {"correct_id": 4}
v1080p_15_ntsc verify_negative 128.53 124.45 [124.34–124.53] 177.74 {"incomplete": 4}
v720_10_hevc verify_positive 7.25 2.05 [1.99–2.13] 3.07 {"correct_id": 4}
v720_10_hevc verify_metadata_loss 7.15 2.02 [2.01–2.05] 3.08 {"correct_id": 4}
v720_10_hevc verify_negative 50.11 41.48 [41.36–41.62] 75.42 {"completed_negative": 4}

Metadata-loss video is a stream-copy remux with equal compressed essence hash; images are lossless PNG re-saves with equal decoded RGB hash. For JPEG this is JPEG→PNG, not byte-identical repackaging. A recovered original manifest does not authenticate an edited asset.

Verification cost

Controls and limits

The earlier client pilot remains a separate, narrow integration observation. It exported 11 engineering fixtures in a resident worker and checked read-back, recovery and signed HTTP reports, but its Python SELF CPU and lifetime maxrss were incomplete resource measures. The matched whole-container series above supplies the missing resource scope; it does not retroactively change those pilot timings or turn either corpus into production evidence.

Historical 0.7.1/0.7.2 benchmarks remain versioned. Some already measured cgroup CPU: the accounting limitation applies to the client-pilot SELF metrics, not to every historical record. Do not infer a speedup by comparing different versions, machines, sizes or sources.

CPU accounting

Encoder timing boundary

The encoder receipt timer measures ONNX calls during embedding. The rest of full-export wall time also includes preprocessing/postprocessing, IPC, decode/encode, signing, registry and read-back. It is not separately attributed to one of those stages. CPU seconds and this wall timer are different units.

For a deployment decision, add representative content, visual/flicker quality, delivery transforms, concurrency, storage, queueing, retries and production signing/TSA to this baseline. A faster profile name does not establish a statutory minimum cadence.

Reference measurement harness and source hashes retain the exact experimental code.

Historical 0.7.2a1 measurements · Choose a video profile · Estimate operating costs

Browse documentation ↗