Linum releases open-weight text-to-image model trained on Pyramid-JiT architecture for about $20K
Linum, a two-person lab training video models, released an open-weight text-to-image model under Apache 2.0, trained for about $20K on 660M samples using its Pyramid-JiT pixel-space diffusion architecture. The company says the research is published on its website and that the architecture converges 11x faster than its latent diffusion counterpart.
Why it matters
The model is an openly licensed, pixel-space alternative to standard latent diffusion pipelines, removing the VAE stage. The company reports lower training cost and cheaper inference from reduced attention tokens, which may matter for teams budgeting image generation or training their own models.
Who should care
Ecommerce brands producing product imagery, designers, indie developers, and ML teams evaluating open image models or low-cost training approaches.
What you can do
Download the Apache 2.0 weights from Linum's site, run a set of your own product or brand prompts, and compare output quality and inference cost against your current image model.