Announcement
BudgetPix is pitched as an image generator that can use as few as 10% of the tokens
Its creator says the pixel-space model adjusts computation to image content at inference time.
TLDR
BudgetPix’s creator describes a pixel-space image generator that adapts its computation to image content during inference. They say it can use from 100% down to 10% of the tokens while keeping the prompt and noise the same.
Combined views
13.7K
1 Source, first seen ago
