How Does AV1 Stack Up Against x264, x265, and VP9 in Short Video Encoding?

Introduction

AV1 has attracted attention since 2017 for its superior compression performance. Industry evaluations (e.g., Facebook, Netflix) have shown that its encoding complexity has dropped from nearly a thousand times that of VP9 to a few hundred times. To verify AV1's performance on short videos, Meitu's audio‑video team conducted a full‑scale assessment in November 2018 using the Top 500 Meipai short videos, benchmarking against the mainstream encoders x264, x265, and VP9 used in production.

Overall Performance

At equal visual quality, AV1 achieves bitrate gains of 36.0% over x264 high profile, 26.9% over x265 main profile, and 31.8% over VP9; the advantage grows with higher resolution.

Encoding time for AV1 is 395× that of x264 high profile, 36× that of x265 main profile, and 156× that of VP9.

Research Background

AV1 (AOMedia Video 1) is the newest video coding standard since 2017, praised for high compression. Tests by Moscow State University, Facebook, and Netflix confirm AV1 surpasses H.265 and VP9 in compression ratio, though early versions suffered from extreme complexity (up to 1000× slower than VP9). Recent efforts by the AOM standardization group and hardware vendors (Intel, Google, ARM, Qualcomm, Samsung, Sony) have dramatically improved speed, bringing AV1 close to practical use.

Software and platforms increasingly support AV1 playback, including Firefox, Chromium, Windows 10, and Android Q, while major hardware manufacturers are developing AV1 decoders, suggesting rapid mobile hardware adoption by 2020.

Comprehensive Encoding Performance Evaluation

The experiment used libaom v1.0.0 to encode 523 short videos selected from Meipai's Top 500 and top creators. Encoders for comparison were ffmpeg 4.0.2 builds of x264, x265, and libvpx‑vp9. Quality metrics included PSNR, SSIM, and VMAF‑Phone.

Video Test Set Selection

The test set consists of 523 videos with the following characteristics:

Mostly captured on smartphones (photo videos, screen recordings, official releases).

Compressed source videos.

Predominantly SD/HD (480p/540p/720p), not UHD/4K/8K.

Frame rate around 30 fps.

Aspect ratio 16:9.

Duration between 10 s and 60 s.

Complexity Analysis of Test Sequences

Each video’s spatial (SI) and temporal (TI) complexity was computed following ITU‑T P.910. Both maximum and average SI/TI values were used because many videos contain frequent scene changes.

Figures 1 and 2 (included below) show the SI‑TI distribution of the first 6 s of the Top 500 videos, indicating that most belong to categories A (single person, limited detail), B (single person with graphics/detail), or C (multiple persons), matching Meipai’s content profile.

Encoder Choices

AV1 was encoded with the AOM reference software (AOM/libaom). H.264/AVC, H.265/HEVC, and VP9 were encoded using ffmpeg 4.0.2 libraries. The specific versions are listed in Table 1 (image below).

Encoder Parameter Configuration

Two bitrate‑control modes were used: CRF and 2‑pass ABR. Table 2 (image) shows the CRF values selected for each encoder; the resulting bitrates served as targets for the ABR runs. Detailed per‑encoder settings are in Table 3.

Evaluation Results

BD‑Rate was used to measure compression efficiency, while encoding time ratios expressed complexity. Results are presented for both CRF/QP and ABR modes.

CRF/QP Results

Figures 3 and 4 (below) display bitrate savings of AV1 over the other encoders at equal PSNR and SSIM, respectively.

At PSNR, AV1 saves on average 35.88% (vs. x264), 26.17% (vs. x265), and 36.00% (vs. VP9). The gain increases with resolution; x265 consistently outperforms x264 and VP9, while VP9 approaches x265 at higher resolutions.

At SSIM, AV1 saves 35.33% (vs. x264), 22.24% (vs. x265), and 68.52% (vs. VP9). The SSIM‑based BD‑Rate for VP9 is unreliable because the RD curves are nearly parallel, making direct comparison impossible. BD‑PSNR should be used alongside BD‑Rate for a more accurate assessment.

VMAF‑Phone results for 1080p indicate that the metric struggles to differentiate AV1, x264, and x265, while VP9 performs noticeably worse.

Encoding time for AV1 is 370.89× (vs. x264), 44.34× (vs. x265), and 168.88× (vs. VP9). At higher resolutions, VP9’s encoding time drops to about one‑fifth of x265’s while maintaining comparable RD performance.

ABR Results

Figures 11 and 12 show BD‑Rate savings under ABR control for PSNR and SSIM.

At PSNR, AV1 saves 36.21% (vs. x264), 27.64% (vs. x265), and 27.69% (vs. VP9). Unlike CRF, VP9’s compression under ABR surpasses x264 and rivals x265, even exceeding x265 at high resolutions.

At SSIM, AV1 saves 32.96% (vs. x264), 21.59% (vs. x265), and 62.35% (vs. VP9). RD curves (Figures 14‑17) confirm the same ranking: AV1 > x265 > x264 > VP9.

Using VMAF‑Phone, AV1’s RD performance is close to x264 and x265, while VP9 remains inferior.

Encoding time under ABR: AV1 is 420.50× (vs. x264), 28.69× (vs. x265), and 143.53× (vs. VP9).

Conclusions

AV1 delivers 30%‑40% bitrate savings over VP9 and x264 high profile at equal quality, and about 27% savings over x265 main profile. Encoding speed has improved dramatically: the v1.0.0 AV1 encoder is 395.7× slower than x264, 156.2× slower than VP9, and 36.5× slower than x265.

VP9 shows weaker performance under SSIM and VMAF, but under ABR it narrows the gap with x265 and encodes roughly five times faster than x265.

Key take‑aways for encoder evaluation:

Native VMAF‑Phone is not suitable for distinguishing codec quality in short‑video scenarios.

Combining BD‑Rate with BD‑PSNR provides a more reliable assessment of RD performance.

Future Work

Meitu will continue monitoring AV1 decoder support on mobile devices and browsers, prioritize AV1 for high‑resolution user‑generated content, and extend evaluations to include Intel’s open‑source SVT‑HEVC and SVT‑AV1 encoders.