Year: 2026
Type: dissertation
Abstract: Video codecs, classical and learned alike, rebuild a generic model of natural video at every receiver and still spend bits on what it cannot predict. A pretrained video-diffusion model has already learned what natural video looks like, and both ends can hold the same copy, so it need not be paid for twice. This thesis compresses a video by specializing such a model to it rather than describing it. The Wan2.1-FLF2V-14B model, which predicts the frames between two boundary frames, stays frozen. We fit a rank-one adapter of 81,360 parameters, quantized to eight bits, to one target video and apply it where information enters the model, leaves it, or conditions it: two interface layers, and five modules that supply the text context and the per-timestep modulation. The bitstream is that adapter plus sparse key frames from a standard codec, which anchor the reconstruction to the target's pixels; the decoder merges the adapter into the frozen weights and generates the video between successive key frames. We evaluate on UVG and HEVC Class B at two resolutions, at rates below 0.007 bits per pixel and as low as 0.00155. At 832×480 every method we compare against spends ten to thirty-four percent more bits and reconstructs less faithfully on both perceptual measures; at full resolution our middle operating point undercuts the lowest rate they publish. Ablations indicate that the adapter carries information the frozen weights cannot supply, and that the pixel anchor keeps it small.
Author: Küçük, Furkan
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214509347