Linum · Pixel-space T2I

Pyramid-JiT (P-JiT) 2B

Linum, a two-person research lab training generative video models from scratch, has released a 2B-parameter pixel-space text-to-image model built on its Pyramid-JiT (P-JiT) architecture. It was trained for 660M samples at a compute cost of $21K, as a proof of concept on the way to larger video models. P-JiT removes the VAE and uses 32×32 patches, against 16×16 in Linum v2 and models such as FLUX-2. The post says this cuts the attention context window by 4× and converges in 11× fewer samples than Linum v2. The model uses a frozen Qwen3.5-4B text encoder and 22 shared DiT blocks. Training-only readout heads predict low-resolution versions of the image, and a PixelREPA head aligns features to DINOv3-L. Training moved from 512² to 1024² with a noise-schedule curriculum. The release is experimental, under Apache 2.0, with model code and weights published.