The Next Frontier of AI Video Is Control

Generative AI is rapidly evolving, and one of the most exciting frontiers is video generation. The ability to create high-quality, controllable video content at speed is transforming industries from Hollywood to individual creators. In a recent A16Z podcast, Falco co-founder Gokcen Gorkem and Head of Engineering Batuhan Taskaya delved into the intricacies of their new generative video model, H3 Max, and the broader landscape of generative video.

The Genesis of H3 Max: Speed and Efficiency

Gorkem opened the discussion by highlighting the core motivation behind H3 Max: addressing the "token market fit" in generative media. This refers to the demand for generating a large volume of tokens (i.e., video content) efficiently. He explained that the industry, including Falco, has been compute-constrained, constantly seeking ways to improve efficiency. "We are growing as much as we are adding compute," Gorkem stated, emphasizing the need for models that can deliver more output with less computational cost.

Taskaya elaborated on the technical breakthroughs that led to H3 Max's impressive performance. He revealed that H3 Max is built upon Minimax H3, the first truly open-source, highly capable, latest-generation video model. This open-source foundation was crucial, as it allowed Falco to build upon a robust architecture. The team focused on a "compounding effect of multiple different optimization variables." This included post-training the model to be more efficient, reducing the number of steps required to generate video (e.g., from 50 steps to 20 steps), and optimizing the underlying systems architecture. Taskaya noted that while reducing steps can sometimes lead to a loss in quality, their approach involved improving quality first and then applying optimizations. This meticulous process resulted in a model that is not only significantly faster (up to 35x speedup compared to the original Minimax H3 endpoint) but also maintains or even improves quality, as measured by ELO scores. The cost is also dramatically reduced, by up to 2x less than previous models.