FaceForensics++ c23 and c40: what the compression levels are and what they cost

This page belongs to the deepfake detection hub and is a companion to the FaceForensics++ page, which covers the dataset's sources, methods and splits. Here the subject is only the compression, because it is the condition most often dropped when a FaceForensics++ score is quoted.

01What do c23 and c40 mean?

The labels are release names, not terms from the paper. The dataset README lists three folders for every sequence, "c0/raw", "c23/hq" and "c40/lq", and describes the compressed videos as "h264 compressed videos with compression rate factor" 0, 23 and 40 (FaceForensics dataset README). The constant rate factor is the H.264 encoder's quality setting: a lower value keeps more detail at a higher bitrate, a higher value discards more. Zero is lossless, so the raw release is the uncompressed source and the README notes that frames extracted from it are "100% the same as our raw images used in the paper".

The paper describes the same three levels in its own words. The authors compress with H.264 because "raw videos are rarely found on the internet" and the codec "is widely used by social networks or video-sharing websites"; the high quality level uses "constant rate quantization parameter equal to 23" and "is visually nearly lossless", and the low quality level is "produced using a quantization of 40" (arXiv 1901.08971, section 3). So c23 is the paper's HQ and c40 is its LQ, and any FaceForensics++ figure should carry one of the three labels.

02How much data is in each level?

The README gives the download sizes for all FaceForensics++ videos, original and manipulated: about 500 GB at raw, about 10 GB at c23 and about 2 GB at c40. The 1,000 source videos came from YouTube and were manipulated with four methods, DeepFakes, Face2Face, FaceSwap and NeuralTextures, giving 1,000 real and 4,000 fake videos at each level, which the dataset comparison table lists beside the other public sets.

Sizes are the README's figures for the FaceForensics++ videos alone; the Google and Jigsaw DeepFakeDetection set that the same script downloads has its own sizes.

03How much accuracy does compression cost?

The paper's Table 1 trains each baseline on all four manipulation methods and reports binary accuracy on the test split of 140 videos at each level. XceptionNet, the strongest baseline, scores 99.26 percent on raw, 95.73 on HQ and 81.00 on LQ. The drop from raw to c23 is 3.53 points; the drop from c23 to c40 is 14.73 points. Every other baseline falls further: MesoNet goes from 95.23 to 83.10 to 70.47, the method of Bayar and Stamm from 98.74 to 82.97 to 66.84, and the steganalysis features with an SVM from 97.63 to 70.97 to 55.98, close to chance at the lowest level.

The ordering of the methods also changes with the level. At raw, six of the seven baselines score above 95 percent and the steganalysis features sit fourth at 97.63; at c40 the same features are last at 55.98, and only XceptionNet stays above 80. A method that reads fine pixel statistics loses most when the encoder smooths them away, so a ranking of detectors made on raw video does not survive compression, which is one reason DeepfakeBench ranks on c23 and tests on c40 separately.

The human observers in the paper's user study, 204 participants classifying randomly selected test images as real or fake, lost less: 68.69 percent on raw, 66.57 on HQ and 58.73 on LQ. Compression hurts the detector more than the eye because the detector was reading high-frequency traces that the encoder removes, while the observers were already near their limit at every level. The paper's conclusion that detectors "clearly outperform human observers" holds at all three levels, but the margin shrinks from 30.57 points at raw to 22.27 at c40.

04Which level do benchmarks use?

c23, almost always. DeepfakeBench trains every one of its 15 detectors on FaceForensics++ c23 and lists "Only c23" for FaceForensics++, DeepFakeDetection and FaceShifter in its dataset table (DeepfakeBench README). Its results table then tests the c23-trained models on c40 as one of six within-domain columns: Xception falls from 0.9637 AUC on c23 to 0.8261 on c40, UCF from 0.9705 to 0.8399 and SPSL from 0.9610 to 0.8174. The cross-dataset scores by model table carries both columns for every detector.

That choice has two consequences for anyone comparing numbers. A FaceForensics++ score from the paper and one from DeepfakeBench are not the same measurement even on the same level, because the paper reports video-level binary accuracy after training on all four methods and DeepfakeBench reports frame-level AUC averaged over the top-3 checkpoints. And a model trained on c23 has never seen c40 artefacts, so the c40 column in DeepfakeBench is a compression generalisation test, not a within-level score. The cross-dataset page treats the change of generator the same way this page treats the change of compression.

05Why does the level matter when quoting a score?

Because the same detector on the same fakes gives three different numbers, and the highest one is the least like anything found online. The paper's own motivation for compressing was that raw video does not circulate; a score on raw describes a laboratory condition. A score on c23 describes a lightly compressed upload, and a score on c40 describes the kind of re-encoded clip a platform serves after one or two passes. The best deepfake detection model page uses the c40 column as its proxy for heavy compression for that reason, and the equal error rate page notes the same effect in audio, where ASVspoof 2021 scored compressed speech by codec and bitrate.

The quotable caveat: a FaceForensics++ accuracy or AUC quoted without its compression level cannot be compared with any other FaceForensics++ figure. The statutes described on the is a deepfake illegal page define covered content by what it appears to show, not by its bitrate, so a detector's sensitivity to compression is a limit on the evidence, not on the offence.