DiffVC-ONE: One-Step Video Compression with Video Diffusion Transformer

arXiv has published the paper DiffVC-ONE — a framework for generative video compression based on a one-step Video Diffusion Transformer. The authors describe three components: a Unified Unidirectional Latent Compressor for compressing latent slices, a One-Step Diffusion Enhancer for spatial-temporal enhancement of an entire group of frames in a single step, and a Hybrid Condition Generator that extracts structural, force, and semantic conditions. As described by the author, state-of-the-art perceptual quality and temporal consistency are achieved on standard benchmarks at low inference cost. The authors have not released the code.

What it means

This work relates to video compression, not generation from text or visual prompts, so it only intersects indirectly with the tools in the video generators section — through the shared architectural foundation of the Video Diffusion Transformer. Among our fact sheets, the closest in architectural terms is LTX-Video: its repository is tagged with the dit tag, and the authors describe it as a diffusion-based generator. However, LTX-Video solves the image-to-video task, not compression, and its fact sheet contains no components for latent compression or one-step enhancement. CogVideoX and Pyramid Flow are text-to-video generators and have nothing to do with video compression. There is currently no practical overlap between DiffVC-ONE and our reference: there is no code, the task is different, and none of the tools we describe solve the task of generative video compression. If the authors publish an implementation, we can return to comparing memory requirements and speed, but for now this is a research work without available code.