How Real-Time AI Video Is Changing How Creators Work

影片總結

fal’s H3 Max makes a 5-second video in 1.5 seconds, enabling continuous AI streams and more controllable Hollywood workflows.

fal co-founder Gorkem Yurtseven and Head of Engineering Batuhan Taskaya explain how fal post-trained MiniMax’s open-weight H3 video model, combining fewer diffusion steps with kernel, systems and hardware optimization. H3 Max delivers over 10 times the speed of the original while preserving quality; its public Turbo version generates a 5-second clip in 1.5 seconds at roughly half the cost. The launch sparked spontaneous Twitch and web experiments, followed by H3 Max Director, which can maintain up to two minutes of video memory and support continuous streams up to 60 minutes. fal says H3 Max has become its platform’s most popular video model, at more than double the volume of the next model. The discussion also turns to professional production: Blender references enable precise scene control, while fal is developing camera, lighting, lip-sync and motion controls, aiming for 99.9% reliability. Hollywood is fal’s fastest-growing segment, with studios seeking targeted tools and secure, IP-aware workflows.

章節

1. 0:00Meet fal & the H3 Max Launch: MiniMax’s Open Video Model Lets fal Post-Train
2. 2:45From Inference Platform to Post-Training: Compute Pressure and a 35× H3 Max Speedup
3. 6:33The Secret Sauce: 50-to-20 Diffusion Steps, 70–80% Utilization, and Eight-GPU Nodes
4. 10:44The Twitch Moment: H3 Max Turbo Generates Five Seconds of Video in 1.5 Seconds
5. 13:11How the Model Actually Remembers What Happened in a Scene: H3 Max’s Viral Live Experiments
6. 15:47The Economics: H3 Max Director Remembers Two Minutes and Streams for up to 60
7. 22:02Beyond Consumer: H3 Max’s Voice-Directed Live Shows, Scene Memory, and Consumer-Hardware Plans
8. 28:37Prompting a Director: Blender-to-H3 Max Workflows and Structured Camera Controls
9. 35:43Where Video Models Go Next: Studio IP, US-Hosted Seance, and 10x–100x AI Adoption