GoKawiilLinum releases Pyramid-JiT, a decoder-only text-to-image model trained without a VAE
Linum's research team has published Pyramid-JiT (P-JiT), a new decoder-only pixel-space architecture that predicts images at multiple resolutions through the diffusion transformer trunk, building on their earlier VAE-free JiT-DDT model. The company says P-JiT matches the final FD-DINOv2 score of its Linum v2 model using 11.3 times fewer training samples and 4.3 times fewer GPU-hours, while operating at four times the pixel resolution. Code and weights have been released under an Apache 2.0 license as a research artifact rather than a finished product.
GoKawiil's interpretation of the reporting above, not reported fact.
By eliminating the VAE and predicting images progressively at multiple resolutions, Linum is attempting to reduce the token counts that make attention computation so costly in image and video generation. The company frames this work as a step toward cheaper video models, noting that current systems like Seedance require token sequences far larger than typical LLM context windows, which it argues limits affordability for animation tools. If the efficiency gains reported here hold up under independent testing, they could suggest a path toward meaningfully cheaper generative video, though this release is explicitly a research checkpoint rather than a production model.
- P-JiT is a decoder-only, VAE-free pixel-space architecture that predicts images at multiple resolutions along the DiT trunk.
- Linum reports it reaches Linum v2's FD-DINOv2 score with 11.3x fewer training samples and 4.3x fewer GPU-hours at 4x the pixel count.
- Code and model weights are released under Apache 2.0 as a research artifact, positioned as a step toward the upcoming Linum v3 model.
Source: linum.ai — Authorship Statement, 2026-10-06
Published there as: “Training Text-to-Image Models Without a VAE”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.