FlowAct-R1 streams lifelike AI avatars with infinite duration

Most AI video avatars you see in live streams today feel painfully fake. They are either stuck in a five-second looping animation or take an awkward multi-second pause before answering a chat prompt.

Generating human video in real time without lag or identity collapse has been an uphill battle. Researchers just introduced FlowAct-R1, an interactive humanoid video generation model that aims to fix that.

According to the project release, FlowAct-R1 achieves 25fps real-time 480p video generation with a time-to-first-frame latency of only about 1.5 seconds. Crucially, it can run for infinite durations without breaking character.

In human preference evaluations, voters consistently favored FlowAct-R1 over existing SOTA methods across real-time streaming, natural behavior, and continuous generation.

How it turns heavy DiT into a streaming engine

Standard Diffusion Transformers deliver great visuals, but their full-attention mechanisms are far too slow for live, two-way interaction.

FlowAct-R1 gets around this by converting the base DiT into a streaming autoregressive model. The team combined this autoregressive adaptation with multi-stage diffusion distillation to slash inference steps.

To make the character look alive, the model uses joint audio-motion finetuning. Instead of just flapping its lips to sound cues, the avatar coordinates mouth movements with natural head and body language.

Long-running video generation usually suffers from drift, where faces distort after a few hundred frames. FlowAct-R1 solves this with a structured memory bank during inference.

By separating reference, long-term, and short-term memory alongside a dedicated denoising stream, the system uses chunkwise generation and memory refinement to keep the identity consistent forever.

Why it matters

We think FlowAct-R1 marks a practical turning point for digital humans. Rather than chasing ultra-high-resolution demos that take minutes to render, the team focused on what real deployment actually demands: low latency and stability.

A 480p resolution at 25fps might sound modest on paper, but hitting a 1.5-second initial latency changes the entire user experience. In mobile shopping streams or live customer support, an avatar that responds instantly feels genuinely interactive, while a 4K avatar that freezes for five seconds breaks the illusion immediately.

For creators and brands, this opens the door to truly autonomous, 24/7 interactive broadcasts that answer questions on the fly rather than playing canned video clips.

Across the industry, this puts heavy pressure on older digital human pipelines. The race is shifting from offline visual fidelity to real-time streaming efficiency. The real hurdle now is serving cost: keeping continuous autoregressive diffusion cheap enough to scale across thousands of concurrent live streams.