Infinite Interactive Human Video Generation

Bytedance’s FlowAct-R1 is another Fidelity dial signal - not because it’s “better video,” but because it’s pushing toward interactive humans that can run indefinitely.

The authors present FlowAct-R1 as a framework for real-time interactive humanoid video generation, designed to stream video at arbitrary durations while staying low-latency responsive. They report ~25fps at 480p with time-to-first-frame ~1.5s, which is the kind of systems-level threshold where “avatar as a live interface” starts to feel plausible.

Under the hood, the core claim is solving the interaction-over-time problem: a “chunkwise diffusion forcing” strategy (plus a “self-forcing” variant) to reduce error accumulation and maintain long-term temporal consistency during continuous interaction - alongside full-body control so the agent can naturally transition between behavioral states (idling → gesturing → listening → speaking) without degrading.

“FlowAct-R1 streams lifelike humanoid videos with naturally expressive behaviors, enabling infinite durations for truly seamless interaction.” - Bytedance

That’s upward pressure on Fidelity: synthetic people that don’t just look good in a short clip, but hold up over minutes/hours of live back-and-forth - where believability is mostly about stability, not peak quality.

> The interface layer is thickening. If you disagree with my interpretation, or you’ve spotted a better signal then reply and tell me.