PJiT

PJiT (Pyramid-JiT, P-JiT) is an experimental text-to-image model from Linum AI, a two-person research lab that trains generative video models from scratch. It has 2B parameters and works directly in pixel space, with no VAE. Linum describes the roughly $21K training run as a proof of concept on the way to its next video model.

The architecture patches images at 32×32 and projects each patch into 1024 dimensions, which gives 32×32 token reduction. Linum says this is 4× more aggressive than the 16×16 reduction in Linum v2 and cuts the attention context window by 4×. Text and image each pass through two refiner DiT blocks, then merge into a 22-block DiT trunk. Text conditioning comes from a frozen Qwen3.5-4B model. Training uses intermediate low-resolution image readouts, a PixelREPA alignment to DINOv3-L, a noise-schedule curriculum, and perceptual losses (LPIPS and P-DINO). The model was trained on about 600M samples.

The weights are released under Apache 2.0 (about 8.74 GB, per a YouTube review). The release is explicitly experimental. A reviewer reported that it struggles with complex scenes and text rendering. The announcement page shows a gallery of 249 samples with their prompts.