Exporting Remotion videos on the GPU: a layered render, measured against Chrome

Remotion renders React compositions to video by capturing each frame from headless Chrome. When the composition embeds camera footage, most of the CPU goes to decoding that footage and capturing frames, not to React. We describe remotion-gpu, an open tool with three paths: P1 (Remotion's own and NVENC), P2 (a layered render: the framing written in React is exported as a per-frame plan and executed on the GPU by ffmpeg with Vulkan Video, libplacebo and NVENC, while Chrome draws only the graphics) and P3 (P2 plus skipping repeated graphics frames). On a 75 s production video, P3 used 112 s of CPU (N=4) against 460 s for P1 (N=1); the layered output matched P1 at mean SSIM 0.978 (N=1). Every number comes with its run count and machine load; the cross-video benchmark is a pilot (N=1, 4 s). The contribution is systems work and measurement, not a new algorithm. We also report what did not work. Code, data and benchmark: https://github.com/konecz/remotion-gpu (MIT).Files: remotion-gpu-paper-en.pdf (paper) and remotion-gpu-artigo-pt.pdf (Portuguese version).