RIFT Proposes One-Step Generative Video Compression with Residual Correction

arXiv published on November 1, 2026 the paper RIFT: Residual-Informed Flow Transfer for One-Step Generative Video Compression. The authors note that generative video compression produces a directional mismatch between codec-induced degradation and the perturbation trajectories that flow-matching models are trained on. Calibrating by degradation strength alone is not enough: compressed latent representations remain strongly aligned across codec operating modes, while restoration residuals vary greatly in magnitude and direction. RIFT introduces a Residual Geometry Inducer that turns the flow prediction into a content-adaptive restoration update using codec features visible to the decoder. The authors report that on three public video compression benchmarks RIFT outperforms competing generative methods. The authors did not release the code.

What it means

The work concerns video compression, not generation, so it does not directly overlap with the tools from the AI Video section. Our reference has no tool pages for codecs or generative video compression tools. The closest related tools are video generators. Maestro works as image-text-to-video on the WanGP pipeline, requires 6 to 24 GB of VRAM, is open source, but the code license permits only non-commercial use of the program itself. LTX-Video handles the image-to-video task, requires only 1 GB of VRAM as described by the author, the code is under Apache-2.0, and the weights are open with a revenue cap of $10 million. Imagination-Studio-by-keys-Wasserman is a studio for Mac and NVIDIA DGX Spark; its fact sheet lists the text-to-speech task and 8.3 GB of VRAM. None of these tools does video compression, and their fact sheets contain no information about using flow-matching to restore frames after encoding. RIFT is a research paper without published code, so its applicability on local hardware cannot yet be assessed.