Linum released a 2B Pyramid-JiT text-to-image model, "a milestone towards more affordable video models," trained from scratch for only about $20K. It describes itself as "a 2-person lab researching more efficient architectures for generative image and video." The details and weights are in Linum's field notes, and the Apache 2.0 weights are on Hugging Face.
The architecture
Last month Linum published Pyramid-JiT, a pixel-space architecture that it says converged 11 times faster than its latent-diffusion counterpart. According to the model cards, it is a single decoder-only diffusion transformer that reads the noisy image as 32 by 32 pixel patches (256 tokens at 512 by 512) together with the caption, with no VAE. One card gives 2.2B parameters and the other 2B.
The checkpoint
- Cost and data: Linum's Hugging Face card for the "pjit-21k" checkpoint puts the compute at roughly $21K, on 660M training samples, against the "~$20K" in the announcement.
- Weights: about 8.74 GB in fp32, under Apache 2.0.
- Limits: the card labels it an experimental research checkpoint, not a production model. It was trained on little to no images of rendered text, so it cannot reliably write text or make posters, and it struggles with full-body shots, crowds and multi-person interactions.
- Comparison: Linum's post compares it with its own earlier Linum v2, a 2B open text-to-video model, with P-JiT scoring lower on one metric and higher on another.
The figures come from Linum's announcement and model cards, and the details beyond the announcement come from search results, since the blog post itself could not be opened.