Linum releases experimental image model from $21K training run

Linum released an experimental image-generation model from a $21K training run, describing it as a proof of concept for its next video model. The two-person lab says its Pyramid-JiT architecture converges in 11× fewer samples than Linum v2, with 4× more aggressive token reduction. The model generates directly in pixel space and uses intermediate predictions of lower-resolution images to strengthen learning earlier in the network. It cannot write text or generate posters because text-heavy training images were excluded, and it struggles with crowds, full bodies and interactions between people.