Linum

Quick answers

What is the Pyramid-JiT (P-JiT) architecture?

P-JiT is a pixel-space diffusion model architecture that achieves 32×32 token reduction (4× more aggressive than Linum v2) and converges in 11× fewer samples. It eliminates the separate VAE compression step and learns compression directly within the diffusion model, enabling more aggressive compression while learning faster.

How much did it cost to train the Pyramid-JiT model?

The complete training and post-training cost approximately $21,038, comprised of $20,348 for pretraining on 610M samples at 512² and 50M samples at 1024² resolution, plus $690 for post-training with Finite Difference Flow Optimization (FDFO). The pretraining consumed 8,139 H100-hours at $2.50 per hour.

What are the model's limitations?

The model cannot generate text or infographics as it was not trained on rendered text data. As a 2B-parameter model, it struggles with full-body generation, scenes with many entities, multi-person interactions, and world knowledge. Training a larger model and additional post-training would likely improve performance in these areas.

Why is Linum building generative models from scratch?

Linum builds models from scratch because artistic controls cannot be effectively added through post-hoc fine-tuning or workflows. The team believes control primitives must be embedded directly into the model training process itself to create tools that capture creative intent and make animation production accessible.

What text encoder does P-JiT use?

P-JiT uses a frozen Qwen 3.5-4B text encoder. The text representation is read from three full-attention layers (layers 7, 15, and 27) which are concatenated and projected into the text refiner blocks.