Researchers have developed a novel video codec that stores video and audio as the weights of a neural network, rather than compressed pixel data. This approach uses a sinusoidal representation network (SIREN) that maps space-time coordinates to video and audio values. After training, the network is compressed using knowledge distillation, quantization, and LZMA2 encoding. The resulting compressed representation achieved a significant reduction in file size while maintaining a PSNR of 28.72 dB for video and 24.18 dB for audio, outperforming traditional codecs like H.264 and HEVC in terms of compression ratio. AI
IMPACT This research presents a novel approach to video and audio compression, potentially offering significant file size reductions compared to traditional codecs.
RANK_REASON Academic paper detailing a novel method for video and audio compression using neural networks. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
