Tensor-Optimized Generative Video for Efficient SME Workflows

Short description of the innovation study

Generating video with AI is still too expensive for most European SMEs in the creative and media sectors: as clips get longer, the GPU memory and energy needed grow faster than the video length, which in practice limits SMEs to clips of a few seconds. TENSOR-VID, carried out by Multiverse Computing, takes an open-weights text-to-video model and makes it much cheaper to run. The model is first compressed with Multiverse's CompactifAI technology and then retrained on high-performance computing (HPC) systems to recover the smooth motion that compression disrupts. Every version is measured against a fixed benchmark of prompts, random seeds and clip durations. The goal is 30–50% less compute per generated second of video with at most a 5–10% loss in quality, so that longer, coherent clips can be produced on studio-grade hardware.

Organisations involved: 

Main SME: Multiverse Computing S.L.