Linum
Open‑source text‑to‑image model for cheap inference
About
Linum is a text‑to‑image diffusion model released by a two‑person research lab. It abandons the traditional VAE‑based latent diffusion pipeline in favor of an efficient pixel‑space architecture, called Pyramid‑JiT, that converges 11× faster and reduces inference cost 4×. The model was trained on 660 million samples for roughly $20 k and is released under Apache 2.0 weights for the community.
The case
Who it is for
ML researchers and engineers building generative video tools who need a fast, low‑cost image backbone.
What is broken today
Current video generation models are slow, expensive, and lack fine‑grained control, making iterative creative workflows impractical.
What it does about it
Linum replaces the standard VAE‑driven latent diffusion pipeline with a pixel‑space diffusion model. The Pyramid‑JiT architecture removes the two‑stage training, speeds convergence 11×, and cuts inference cost 4× by aggressively reducing token counts in attention windows.
Why not the obvious alternative
Unlike latent diffusion models such as Stable Diffusion, Linum operates directly in pixel space, eliminating the VAE bottleneck. This yields faster training, cheaper inference, and higher quality control without sacrificing output fidelity.
Details
Questions
- What is the cost of training Linum?
- The model was trained for approximately $20 k on 660 million samples.
- Under what license are the weights released?
- The weights are released under the Apache 2.0 license.
- How does Linum differ from latent diffusion models?
- Linum removes the VAE stage, trains directly in pixel space, converges 11× faster, and reduces inference cost 4×.
- Is Linum suitable for video generation?
- The lab views Linum as a stepping stone toward cheaper video models, since most video models start from an image backbone.
No comments yet. Be the first to give the maker feedback.