Plate III · Video generation
VideoForge
Clips rendered on racks of cards that were never meant to see a diffusion model, at a price per second that makes a long cut affordable.
In service
Video is where old cards look worst on paper and best in practice. One clip is one long computation, and a long computation on a cheap card is still cheap. The work was in making the model fit at all.
Two lanes run today. A draft lane returns a short clip in about eighty seconds for judging a shot. A long lane renders the full-length take, at full resolution, from two keyframes.
What was hard
The model wanted more memory than a card of this generation carries. We split the attention and the modulation work into chunks that fit, and proved the output was identical to the unsplit run byte for byte before shipping any of it. A clip of two hundred and twenty-six frames is the longest that fits in memory on this hardware; two hundred and forty-three frames also fits, and runs two point two times slower, because the driver starts moving memory in and out of the card.
The specification
- Length: up to two hundred and twenty-six frames from two keyframes on the long lane.
- Draft: about eighty seconds for a judging clip.
- Method: every speed change is checked against the unsplit render. A change that alters the picture is not a speed change, it is a different film.