How does causal-AnyFlow compare to causal-rCM?

[fy-zhang-ctrl]

I noticed that rCM has released a causal version and claims that its 1-step frame-wise model achieves a higher VBench score than 4-step chunk-wise AnyFlow.

I am curious whether AnyFlow can also achieve performance comparable to rCM in 1-step generation. If so, should we simply use AnyFlow instead of rCM in the future, since AnyFlow supports arbitrary-step inference?

[guyuchao]

Hi, I don't believe the VBench score is an accurate reflection of actual quality. I will be looking into their model soon.