GVCCTurbo proposes bitrate planning for generative compression of video and images without

arXiv The researchers introduced GVCCTurbo, a scheduler that separates costly generative prior updates from correction transmission via a codebook. Instead of rigidly tying the number of sampler steps to the number of correction slots, the method lets you set the bitrate as an input parameter of the schedule. The scheduler is compatible with rectified-flow video (GVCC style) and diffusion-based image compression (DDCM style), requires no fine-tuning, and works with future distilled priors. In a 720p test with Wan-GVCC, the number of prior evaluations was reduced from 20 to 9, yielding about 44% savings in decoding time with a slight increase in LPIPS on dynamic scenes.

What it means

This news concerns generative video compression, not text-to-video generation. Our reference has no tools that tackle video compression via codebooks. The closest tool by name — Wan — is a video generator from Alibaba that works on the text-to-video model, not a codec. GVCCTurbo uses Wan only as a test bench for measuring decoding speed, without modifying the model itself or adding compression capabilities to it. The work has no direct impact on the tools in the video generation section. It sits in an adjacent area — ultra-low-bitrate video transmission relying on generative models — which is not yet covered by individual tool pages in our reference.