📝 Abstract:
Classical and learned video codecs rebuild a generic visual prior at every receiver and transmit the residual it fails to predict. This thesis inverts that premise. A pretrained video-diffusion model already encodes a prior over natural video, and because every receiver holds an identical copy before any bitstream arrives, that prior is decoder- side side information rather than transmitted content. We compress a video by spe- cializing this frozen predictor to it: the Wan2.1-FLF2V-14B model, which predicts the frames between two boundary frames, stays frozen, while a rank-one adapter ap- plied to seven modules (the two layers that tokenize and de-tokenize the latent, and five conditioning embeddings) shifts it toward the target. The transmitted bitstream is this adapter, 81,360 parameters quantized to eight bits, together with sparse key frames from a standard inter-frame codec. At the decoder the adapter is merged into the frozen model and the full video is reconstructed in a single forward run of the diffusion sampler. On UVG and HEVC Class B the scheme stays below 0.007 bits per pixel, reaching 0.00155 at full resolution, and is competitive with recent learned video codecs at comparable rate