Why doesn't the model context explode?

[chengjinze]

When feeding the generated video frames and the context into the model, the token count nearly exceeded 120k; I’m really curious how it managed to run smoothly.

[oso95]

Because the video never enter the model context at all, the model only reads a few small, deliberately-composed images.