Netflix Just Open-Sourced the Wrong Thing

Netflix dropped a video generation repo this week — Go-with-the-Flow, out of their Eyeline Research group — and I'm sitting here trying to figure out which of two conflicting thoughts to have first.

The system works like this: you describe motion in text, the model converts that into optical flow fields, and the flow fields condition the video generation. Optical flow as an intermediate representation between language and video is not a new idea, but the execution here is clean enough that "not surprised" doesn't feel like the right framing. More like: the right people finally had the right infrastructure budget.

Eyeline Research is Netflix's visual R&D arm — they've done neural face replacement for localized dubbing, computational cinematography, the kind of unglamorous applied vision work that keeps a streaming service running at scale. They are good at this. Seeing them publish foundational video generation work is a jump, but not a leap into the void.

What's interesting is the comment embedded in the paper — the part about text prompts as a distribution mechanism.

That's the thing. Not as a cute demo note. As a real observation about what this technology implies.

Netflix's entire business is distributing video. Petabytes of it, constantly, to a few hundred million people. If you can instead distribute a model and a text description and generate the video at the edge — or locally — you have just described a completely different company. A company with different infrastructure costs, different content licensing implications, different everything.

They probably didn't mean to say that out loud.

Or they did, and it's a research paper, and nobody in business strategy reads research papers, so it's fine.