HyperFlow Infinity: an open agent loop that keeps generating video until you stop it

Video Rebirth, the team behind the HyperFlow 8-step LoRA for MiniMax H3, announced HyperFlow Infinity on Oct 25 (6:20 AM MDT). The post passed 300 likes and 370 bookmarks in a day, a lot for a 1.5K-follower account. The code is Apache-2.0. You give it a story seed: a world, an opening situation, characters and a reference image. It then runs in a loop. H3 renders a 5-second segment, a vision-language model watches it and writes what changed into a scene graph, a 'director' step picks the next beat, and H3 continues from the last few latents of the previous clip. No extra model is trained.

Per the README, the scene graph is the memory. It tracks characters, places and recent events, and folds older events into chapter summaries, which is how the same Alice stays on screen across the full 60-minute demo at 1344×768. The director also rejects beats that repeat recent ones. You can pause a run, branch from any segment with a new instruction ('explore the cave instead'), and resume. The repo compares it with Echo-Infinity and argues its plot stays open-ended instead of following a preset path.

The requirements rule out most studios for now. The README asks for at least 4 NVIDIA GPUs (3 if the vision-language model runs through OpenRouter), and the only tested setup is H200 141 GB cards. Its own numbers on an 8×H200 node are about 49–78 s per 5-second segment, depending on the Qwen backend. Licensing matters more: the code is Apache-2.0, but the HyperFlow LoRA weights fall under the MiniMax H3 Community License, which according to the README excludes the EU, UK, South Korea and the US unless MiniMax authorises it.

Why it's on ReelStack: this is the first open, readable version of the 'agent that watches its own footage and directs the next shot' pattern. Even if you never run H3 at this scale, the scene-graph memory, the anti-repeat rule and branch-from-segment are ideas you can copy into a Seedance or Kling pipeline driven by Claude Code.

Our take: treat the 60-minute film as a technical demo, not a storytelling benchmark. Check where you're based against the weights licence before building client work on it.