Linum Releases P-JiT Image Model With Training Cost Kept At $21,000
Linum, an AI research organization run by two AI researchers, has released the image generation model Pyramid-JiT (P-JiT) with open weights.
1 source 15m ago
Reporting from 1 source: GIGAZINE.
Why it matters
P-JiT shows that a pixel-space image model can converge on about one-eleventh the samples and roughly $21,000 in training spend, which shifts the cost argument for researchers who cannot match large-lab budgets.
Key Facts
- Linum released the image generation model Pyramid-JiT (P-JiT) with open weights.
- P-JiT processes images directly as pixel data without converting them into latent representations.
- P-JiT training cost approximately $21,000.
- P-JiT converged with about one-eleventh the number of samples compared with Linum v2.
- P-JiT does not support generation of images containing text, posters, or infographics.
P-JiT processes images directly as pixel data instead of converting them into latent representations, and its training stayed at roughly $21,000. The model reached convergence with about one-eleventh the samples of the earlier Linum v2 by compressing tokens harder.
P-JiT skips the usual VAE step. The model runs on 32x32 image patches rather than latent representations, which cuts the attention context window enough to train on far fewer samples than Linum v2. The architecture pairs a text encoder built from Qwen3.5-4B with a stack of Diffusion Transformer blocks, split into two single-modality refiner blocks and 22 identical blocks. Training ran in stages, with around 660 million samples at 512x512 before roughly 50 million at 1024x1024. Linum states the model does not handle text inside images, posters, or infographics.
Written with AI by Yomimono from the 1 cited source below, including Japanese-language reporting where cited. See our editorial policy.
Add Yomimono as a preferred source on Google