Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Neural B-frame Video Compression with Bi-directional Reference Harmonization

Authors: Yuxi Liu, jin dengchao, Shuai Huo, Jiawen Gu, Chao Zhou, Huihui Bai, Ming Lu, Zhan Ma

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results indicate that our BRHVC outperforms previous state-of-the-art NVC methods, even surpassing the traditional coding, VTM-RA (under random access configuration), on the HEVC datasets. The source code is released at https://github.com/kwai/NVC. 5 Experiments 5.1 Settings 5.2 Comparison Results 5.3 Ablation Study and Complexity 5.4 Visulization
Researcher Affiliation Collaboration 1Nanjing University, Nanjing, China 2Kuaishou Technology, Beijing, China 3Beijing Jiaotong University, Beijing, China
Pseudocode No The paper describes methods through textual descriptions, diagrams (e.g., Figure 4, Figure 5, Figure 6), and mathematical equations, but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes The source code is released at https://github.com/kwai/NVC.
Open Datasets Yes Training. We use Vimeo-90k [50] to train BRHVC from scratch with 7-frame sequences. ... Testing. We evaluate the compression performance on HEVC Class B~E [5], UVG [34], and MCLJCV [46].
Dataset Splits Yes We use Vimeo-90k [50] to train BRHVC from scratch with 7-frame sequences. Then we fine-tune BRHVC on original Vimeo videos with 17-frame sequences following [39, 23]. ... The video frames are randomly cropped into 256 256 patches. We randomly reverse the order of sequences with a probability of 50% as data augmentation. ... The first 97 frames of all the datasets are used with Intral Period 32, i.e., coding 93 B/P-frames with 4 I-frames.
Hardware Specification Yes Table 2: Ablation study on different network designs with the corresponding computation cost. The BD-rate results are calculated on HEVC datasets. We compare average single-frame codec times tested in 1080p sequences on one Nvidia RTX 4090 GPU.
Software Dependencies No The paper mentions 'Adam W [19] is used as the optimizer' but does not specify version numbers for programming languages, libraries, or other key software components.
Experiment Setup Yes Training. We use Vimeo-90k [50] to train BRHVC from scratch with 7-frame sequences. Then we fine-tune BRHVC on original Vimeo videos with 17-frame sequences following [39, 23]. We use the multi-stage training strategy in [39]. The video frames are randomly cropped into 256 256 patches. We randomly reverse the order of sequences with a probability of 50% as data augmentation. Adam W [19] is used as the optimizer with a batch size of 8. Testing. ... The first 97 frames of all the datasets are used with Intral Period 32... --Intra Period=32 --QP={qp} --Level=6.2 --Bitstream File={bitstream_file}