The problem. High-quality video compression at very low bitrates often produces blurry or fake-looking videos. Generative models can create realistic details, but they risk adding incorrect information that doesn't match the original scene. The key is balancing faithfulness to the source and visual realism.
What was done. The researchers identified that a critical decision is when to start the denoising process in a , a type of generative model that progressively removes noise. They developed a system, ADE-GVC, that uses a lightweight predictor to adaptively pick the optimal start time for each video sequence, instead of using a fixed start time for all videos.
Main finding and why it matters. Compared to the state-of-the-art baseline GLC-Video, their method saved an average of 40.99% in bitrate for a perceptual quality metric (DISTS) and 95.24% for a traditional pixel-accuracy metric (PSNR). This shows that adaptively selecting the denoising entry point can dramatically improve both realism and fidelity in very low bitrate video compression.