12:00–1:00 PM ET (this week only)
Next session / Tuesday
Active Sampling for Ultra-Low-Bit-Rate Video Compression via Conditional Controlled Diffusion
Amirhosein Javadi
ActDiff-VC is a generative video compression framework that pairs active sampling with conditional video diffusion. Rather than transmitting information for every frame, it adaptively selects keyframes from scene dynamics and represents motion with a compact set of point trajectories, and a conditional diffusion decoder reconstructs the video from these sparse signals. On standard benchmarks it reaches up to 64.6% bitrate reduction at comparable perceptual quality relative to strong learned codecs.
