Krea releases open real-time 14B video model
Krea has released Krea Realtime 14B, an autoregressive video model it says is over 10x larger than existing real-time video models, letting users steer generation as frames stream back.
Krea announced Krea Realtime 14B, a 14-billion-parameter model for real-time, long-form video generation. The company describes it as an effort to build the first publicly available autoregressive video generation model.
The model is distilled from Alibaba's Wan 2.1 14B text-to-video model using Self-Forcing, a technique that turns bidirectional video diffusion models into autoregressive ones. Krea says it generates text-to-video at 11fps using 4 inference steps on a single NVIDIA B200 GPU, with first frames arriving in as little as one second.
According to Krea, existing open-source real-time video models are built on smaller bases such as Wan 2.1 1.3B. It says those models struggle with complex motion, high-frequency detail, text-to-video and long generations.
Because each frame becomes fixed once it is denoised, frames can stream to the user right away. Krea says this lets users change prompts mid-generation, restyle video streams on the fly and edit video-to-video in real time.
The write-up focuses on exposure bias. Autoregressive models trained on ground-truth frames must rely on their own imperfect outputs at inference, so errors build up over time. Self-Forcing addresses this by generating autoregressively during training. On top of that, Krea introduces two inference techniques, KV Cache Recomputation and KV Cache Attention Bias, to limit error accumulation. It also describes memory optimizations for training large autoregressive video models.
Krea's evidence for quality comes mainly from example videos. The material provided does not include benchmark comparisons or license terms.