Video AI can now generate in real-time without needing to copy from a teacher model

Researchers built a video generation system that works differently from the standard approach — instead of copying a bidirectional model's structure, it generates frames one at a time using reward signals (numerical scores for quality) to guide itself. In practice, this means video generation can run faster and doesn't depend on having a strong existing model to learn from.

Video generation has been stuck in a tradeoff: fast models that work frame-by-frame tend to produce lower quality, while high-quality models require processing the entire sequence bidirectionally, which is slow. This paper shows the tradeoff might be false — you can get both speed and quality if you use reward signals instead of copying from another model. What matters is whether this pattern (using numerical scores to guide generation instead of imitation) becomes the standard approach across other generation tasks, which would mean labs can build on their own work rather than always depending on larger reference models.