Shaduf.Research preview
Real-Time AI Video/EdenLive 5090 file speed versus browser service

Pool topic: Real-time, continuous, and interactive AI video generation. Question revision: 1. Exact question: What can generate continuous, interactive AI video today, and which setups work at which cost, latency, hardware requirements, and quality?

Checked 27 October 2026

EdenLive clears 24 FPS on an RTX 5090. Its browser service is a different test.

Project reports: the LightX2V EdenLive recipe produced a 717-frame, 1280×704 proxy-video job at 25.17 inference FPS on one RTX 5090. That is a warm, accelerated path that saves an MP4. The EdenLive WebRTC service guide separately lists an 80-GB NVIDIA GPU in its verified environment; it does not publish a matched 5090 browser result.

Why a fast file can still stall a player

The 717 frames play for 29.88 seconds at 24 FPS. The project’s real-time factor of 0.9534 implies about 28.5 seconds of measured inference: roughly 1.4 seconds of margin across that sample. Load, compilation, warmup and final file saving are outside that figure. A browser also needs delivery and a buffer. Average FPS alone does not show that a changed game action appears promptly or that every frame arrives on time.

EdenLive uses the game’s graybox frames for layout and motion; the game keeps its own rules and state. The model changes how that scene looks. It is a renderer for a reference stream, not a substitute for the game engine.

Test the scene before paying for the stream

A game developer reports good backgrounds but distorted moving objects on their own 5090 footage. Their settings lacked the optimized compile and CUDA Graph path, so their slow two-segment result is not a speed comparison with LightX2V. It does show why a game builder should judge moving objects, not just scenery.

If you already have a suitable 5090, reproduce the documented proxy-video job and check moving-object quality. Stop if the result fails your chosen visual bar. Before calling it live, test the WebRTC path on supported hardware and record action-to-visible delay, dropped or held frames, reconnects and the full bill. Unknown: those receiver and cost results have not been measured here.

Search published pools, pages, reports, and evidence.