Research map / Video / Video generation
GVCC: Zero-Shot Video Compression via Codebook-Driven Stochastic Rectified Flow
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GVCC, a zero-shot video codec that turns a pretrained rectified-flow video model (Wan 2.1 14B) into a decoder by converting its ODE sampler to a marginal-preserving SDE and transmitting codebook atom indices per step.
Problem
Ultra-low-bitrate video compression suffers from oversmoothing; deterministic rectified-flow video samplers leave no per-step stochastic channel for transmitting information, so generative models are usually used only as post-hoc refiners.
Approach
Convert the deterministic RF probability-flow ODE into an equivalent reverse-time SDE at inference time (score derived from velocity), and replace per-step Gaussian noise with top-M codebook atoms selected by correlation with the encoder-side residual, so transmitted indices and signs specify the generation trajectory without retraining.
Key results
- On full-UVG 1080p, GVCC-T2V reaches LPIPS-Alex 0.1329 at 0.00268 bpp, versus DCVC-RT at 0.2913 at 0.00533 bpp per the paper's unified evaluation.
- On UVG 720p (3 GOPs), GVCC-I2V averages PSNR 31.97 / LPIPS 0.074 at ~0.054 bpp, and GVCC-FLF2V averages PSNR 31.40 / LPIPS 0.096 at ~0.013 bpp.
- User study preferred GVCC-T2V over DCVC-RT and GNVC-VD in 97% and 88% of comparisons, respectively.
Surprising findings
- GVCC-I2V obtains nearly the same VMAF as GVCC-FLF2V (67.59 vs 67.80) despite higher bitrate and lower LPIPS, suggesting the tail residual does not uniformly improve reference-aligned fidelity.
- Without a tail residual, I2V becomes unstable from GOP 3 on Jockey, so the residual is not a smooth rate-control knob.
Limitations
- Encoding a 33-frame GOP takes ~90 s and decoding ~60 s on one RTX 6000 Ada with a 14B DiT, which is much slower than conventional codecs.
- Evaluation is limited to local atom-count sweeps and no matched-rate or global BD-rate claim is made; I2V lies in a separate rate region and is treated as exploratory.
- Quality degrades on out-of-distribution content (PSNR several dB lower) and g_scale has a narrow stable window (2-3).
Experimental setup
Datasets: UVG
Benchmarks: UVG
Metrics: PSNR MS-SSIM LPIPS DISTS BPP VMAF FVD FID KID E_warp
Baselines: DCVC-RT GLC-Video GNVC-VD Free-GVC HEVC VVC DCVC-FM GLVC
Builds on: DDCM Turbo-DDCM NCS score-SDE rectified flow Wan 2.1
Closest papers by meaning
- Taming generative video models for zero-shot optical flow extraction (NeurIPS 2025 · similarity 0.86)
- Once-for-All: Controllable Generative Image Compression with Dynamic Granularity Adaptation (ICLR 2025 · similarity 0.85)
- TI2V-Zero: Zero-Shot Image Conditioning for Text-to-Video Diffusion Models (CVPR 2024 · similarity 0.85)
- Towards image compression with perfect realism at ultra-low bitrates (ICLR 2024 · similarity 0.85)
- FlowVid: Taming Imperfect Optical Flows for Consistent Video-to-Video Synthesis (CVPR 2024 · similarity 0.85)
- Pyramidal Flow Matching for Efficient Video Generative Modeling (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.