[Digital Today reporter Yoonseo Lee] Linum, a small AI research group of two people, unveiled an image generation AI model called Pyramid-JiT (P-JiT) that it developed for $21,000, or about 28.5 million won.

An online media outlet, Gigazine, reported on Nov. 9 that the model focuses on reducing the high computing costs of image generation AI and the difficulty of control, and is currently being distributed as open weights.

Conventional image generation models typically compress images into latent representations through a VAE (variational autoencoder) and then restore them. P-JiT, by contrast, handles pixel data directly without such conversion.

To reduce computation, it also increased the level of token compression. While the tokenisation unit of Linum v2 was 16×16, P-JiT supports 32×32. It represents the same 512×512 image with fewer tokens, cutting the context length processed by the attention context window to one quarter. Linum explained that training was possible with samples equivalent to about one-eleventh of its previous model, and that training costs were about $21,000.

The model architecture was also adjusted for pixel-space processing. P-JiT includes a noise schedule suited to pixel-space models and a "readout mechanism" that strengthens gradient signals in early blocks of the diffusion transformer. The readout mechanism predicts low-resolution images early in the network to speed up convergence.

Text and images each go through two single-modality refinement blocks and are then combined into a single stream made up of 22 identical diffusion transformer blocks. The text encoder uses Qwen3.5-4B with parameters frozen and uses the outputs of specific layers. Images are split into 32×32 patches, then reduced to a 1,024-dimensional latent space at a 3-to-1 compression rate and projected into an image refinement block.

The training process used multiple loss functions together. It combined MSE, PixelREPA-based cosine distance, MSE of the final output head, and perceptual losses based on LPIPS and P-DINO. Training proceeded by first learning about 660 million samples on 512×512 images, then additionally training about 50 million samples on 1,024×1,024 images.

There are also limitations. The training data did not include infographics, PDFs or book pages. Linum said P-JiT therefore does not support generating images containing text, posters or infographics. It also acknowledged that because the model is relatively small, limitations remain in full-body depiction, rendering multiple objects, interactions among multiple people, and areas requiring broad knowledge.

The release is drawing attention as an attempt to lower the cost structure of image generation models. That is because reducing the latent-representation conversion stage and training directly in pixel space led to lower development costs and fewer training samples. Still, as text-based images and understanding complex scenes still need improvement, performance upgrades and follow-on training methods after the open-weight release are expected to be key points to watch.