Linum open-sources Pyramid-JiT: A $21K text-to-image model

If you told someone a few months ago that a two-person team could train a solid image generation model entirely from scratch for about the cost of a used Honda Civic, they'd laugh you out of the room.

Yet that is exactly what research lab Linum just pulled off. The duo open-sourced Pyramid-JiT, an experimental 2B-parameter text-to-image model that cost a mere $21,000 in compute to train from the ground up.

In an industry where foundation model training budgets usually start at seven figures, Linum's run feels like a major wake-up call.

Ditching the VAE entirely

Almost every modern diffusion model relies on a Variational Autoencoder (VAE) to crunch high-resolution images into a smaller latent space before running the diffusion process. Linum decided that pipeline is fundamentally broken.

According to the team, VAEs hit an empirical compression ceiling and can actually degrade downstream generation quality. Instead of leaning on another model's latent representation, Pyramid-JiT operates directly in pixel space, forcing the diffusion model to learn its own compression and generation simultaneously.

To make pixel-space computation affordable, Linum built a pyramid architecture. They added intermediate readouts along the DiT trunk that predict the clean image at quarter and half resolutions. This injects strong gradient signals into earlier layers, helping the model converge 11 times faster than their previous architecture with four times more aggressive token reduction.

Using 64 H100 GPUs across eight nodes, the team trained through roughly 610 million samples at 512px resolution before stepping up to 1024px. The total bill came out to $21K flat.

Small budget, clear limits

Given the footprint, the visual quality is impressive across realistic portraits, nature scenery, and stylized concept art. But a 2B model trained on a shoestring naturally comes with compromises.

For starters, Pyramid-JiT cannot render text at all. Because Linum's ultimate roadmap points toward video animation tools, they deliberately stripped out all text-heavy images, infographics, and book pages from their dataset.

It also struggles with full-body poses, fine limb anatomy on starting blocks, dense crowds, and complex multi-object prompts like counting six cars.

Still, the release is remarkably open. The weights weigh in at roughly 8.74GB on Hugging Face under a permissive Apache 2.0 license, meaning anyone with a decent consumer GPU can download and run it locally for commercial or personal experiments.

Why it matters

For independent developers and small AI labs, Pyramid-JiT proves that pre-training custom diffusion models is no longer the exclusive playground of venture-backed giants. At $21,000, specialized teams can afford to train task-specific base models from scratch, tailoring data pipelines exactly to their needs instead of wrestling with fine-tunes of generic baselines.

For the wider industry, this project puts direct pixel-space generation back on the map. By showing that smart architectural tricks like multi-scale readouts and curriculum noise scheduling can bypass the traditional VAE bottleneck, Linum challenges the conventional wisdom that latents are mandatory for efficient diffusion.

We think Linum's approach is brilliantly pragmatic. Pyramid-JiT is not trying to beat Midjourney at poster design; it is simply a proof-of-concept for their upcoming Linum v3 video engine. Instead of chasing bloated parameter counts, the two founders focused on slashing inference costs and training overhead. While the model has real flaws in anatomy and text, its architectural efficiency offers an inspiring blueprint for how lean research labs can compete in modern AI.