Video Area Home >> Video Codecs Comparisons >> MSU Video Codecs Comparison 2025

MSU Neural Video Codecs Evaluation

Recently, Microsoft published MLVC, a neural video codec designed specifically for NPU deployment and tested in Microsoft Teams. This represents an important milestone in the development and practical adoption of neural video codecs (NVCs). Several highly efficient codecs based on the DCVC framework have emerged in recent years, demonstrating impressive progress compared with conventional video codecs. However, we believe that it is important to evaluate these solutions not only against standard reference implementations, but also against the world's leading state-of-the-art encoders, many of which are extensively evaluated in the MSU Video Codecs Comparisons. Therefore, we present a short evaluation, including subjective comparisons of MLVC, DCVC-RT, and DCVC-UF against some of the world's best video encoders.

Below, we evaluate these NVCs against encoders from two MSU benchmarks: the MSU FullHD Video Codecs Comparison, which focuses on software encoders, and the MSU Hardware Video Codecs Comparison, which evaluates encoding efficiency across different hardware platforms, including GPUs, VPUs, and FPGAs. The first comparison allows us to assess how modern NVCs perform against frontier software codecs in terms of compression efficiency. The second addresses a more practical question: if GPU resources are available for efficient video encoding, is a neural video codec already the best choice, or do modern hardware-accelerated conventional codecs remain competitive?

FullHD Software comparison

The MSU FullHD Video Codecs Comparison includes software encoders evaluated under aligned encoding-speed requirements. It covers both well-known open-source codecs and submissions from the world's leading codec developers, providing a competitive benchmark for state-of-the-art video compression technologies. The test set consists of 50+ videos selected from more than 30,000 high-quality samples to provide broad coverage of the spatial and temporal complexity (SI/TI) space.

For the evaluation presented here, the closest scenario is the VOD use case (5+ FPS). As shown below, the aggregated encoding speed of the tested NVCs across the test set is close to that of most encoders in this category. For the NVCs, encoding was performed on NVIDIA GeForce RTX 4070 Ti GPU and for all other encoders on Intel Core i7 12700K CPU.

Most of the evaluated codecs outperform MLVC in terms of PSNR 6:1:1, while a substantial number also outperform DCVC-RT and DCVC-UF. A similar picture can be observed for other metrics, such as SSIM 6:1:1 and VMAF Y. According to the corresponding leaderboards, DCVC-RT achieves approximately 10-20% better BSQ-rate than the reference x265 encoder, meaning that it requires roughly 10-20% less bitrate to achieve the same quality. In contrast, MLVC requires more than 30% additional bitrate.

Overall, when NVCs are compared with industrial implementations of conventional video coding standards, their compression efficiency is generally comparable to that of open-source HEVC and AV1 implementations. However, across most evaluated quality metrics, they still substantially underperform the leading proprietary implementations of VVC, AV1, and HEVC.

We recognize that the types and distributions of distortions produced by neural video codecs can differ substantially from those of conventional codecs, and that traditional objective quality metrics may therefore not always provide a fully representative comparison. We have previously investigated this issue in a dedicated benchmark study, identifying the quality metrics that are most suitable for evaluating neural and conventional video codecs. For this reason, we additionally included MLVC and DCVC-RT in our subjective evaluation. The comparison covers 8 videos and includes more than 20,000 pairwise subjective assessments, which were aggregated into quality scores using the Bradley-Terry model.

In the subjective evaluation, however, both neural codecs perform considerably better than their objective-metric results would suggest. DCVC-RT is outperformed only by the leading proprietary codec implementations, while MLVC achieves subjective quality comparable to VVenC and SVT-AV1.

These results once again highlight the limited relevance of conventional objective quality metrics for evaluating neural video codecs, whose distortion characteristics can differ substantially from those of traditional codecs. At the same time, the strong subjective performance demonstrates the impressive progress of NVCs, which are already becoming competitive with state-of-the-art conventional video coding solutions.

FullHD Hardware comparison

In the MSU Hardware Video Codecs Comparison, we evaluate high-speed video encoding solutions running on a wide range of hardware platforms. In particular, the latest 2025 comparison included 14 unique devices, enabling us to compare encoding efficiency across different hardware architectures and acceleration technologies.

Since, unlike most hardware-accelerated conventional codecs, the evaluated NVCs are currently unable to reach the higher encoding-speed ranges, we focus below on the 30+ FPS use case, which provides the closest practical point of comparison.

Overall, DCVC-RT demonstrates strong competitiveness in this comparison, outperforming all evaluated GPU- and CPU-based codecs except proprietary codecs Streamlake-200, Tencent canghaiV1 and Tencent canghaiV2, which run on VPU and FPGA platforms, respectively. Particularly notable is that DCVC-RT outperforms solutions based on dedicated VPUs from NETINT and AMD, highlighting the competitiveness of neural video coding even against specialized hardware designed specifically for video encoding.

Takeouts

Based on the current evaluation, we can draw the following conclusions:

Neural video codecs are already competitive with conventional video coding solutions, with DCVC-RT showing particularly strong performance. MLVC is considerably less efficient; however, even the strongest NVCs still lag substantially, sometimes by several times in bitrate, behind the best industrial implementations of conventional codecs.

Subjective evaluation provides a more representative picture of NVC performance. Conventional objective metrics do not fully capture the perceptual characteristics of neural compression artifacts, as we have also demonstrated in our previous benchmark study. In subjective tests, both DCVC-RT and MLVC therefore perform noticeably better relative to conventional codecs than objective metrics suggest.

DCVC-RT outperforms all tested integrated GPU video encoders across the evaluated quality metrics, demonstrating that NVCs can already be highly competitive with conventional GPU-based hardware encoding. The remaining gap is primarily to the strongest specialized and proprietary solutions.

by Nickolay Safonov