TL;DR Linum v2 was bottlenecked by its attention context window: a 720p, 5-second clip cost 110K tokens. Most generative systems are Latent Diffusion Models (LDMs) with separate VAE and DiT. Pixel-space models like JiT combine them but struggle with fine details. We propose JiT-DDT, a novel encoder-decoder architecture that recovers detail and trains efficiently. Against Linum v2, JiT-DDT trains a text-to-image model with 3.6× fewer GPU-hours, even generating images with 4× the pixels. Linum v2 (256×256, 2.0B latent-space DiT + VAE, 256 latent tokens) vs. JiT-DDT (512×512, 2.5B active pixel-space DiT, 320 pixel tokens = 64 encoder + 256 decoder). Code and weights are available under Apache 2.0. This is a research artifact, not a full model release. Stay tuned for Linum v3.

Almost all generative image and video models are LDMs, using a VAE for compression and a DiT for generation. VAEs are trained separately and frozen, capping token reduction at 16×16 for models like FLUX, Ideogram, and Z-Image. Last fall, Tianhong Li and Kaiming He published JiT, achieving 32×32 token reduction by discarding the VAE.