tl;dr: A 2-person lab trained a 2B-parameter text-to-image diffusion model (Pyramid-JiT) for ~$21K by ditching the VAE and operating directly in pixel space, enabling 32×32 token reduction (4× more aggressive than typical) and 11× faster convergence than their previous model. The architecture uses a frozen Qwen3.5-4B text encoder, readout heads along the DiT trunk for stronger early-block gradients, and DINOv3-L feature alignment. Weights are released under Apache 2.0 as a stepping stone toward their eventual video generation model; limitations include no text rendering, crowds, or complex multi-person scenes.