Twenty-one thousand dollars. That is the entire training bill Linum says it ran up building Pyramid-JiT, a 2-billion-parameter text-to-image model, and the two-person lab has now posted the weights on Hugging Face under Apache 2.0, with the model code on GitHub.
The gallery on their field notes page shows 249 samples of what it does, and you can pull the whole thing down today for roughly the price of a mid-range car.
What the money bought
The spec sheet, straight from Linum’s field notes: 2 billion parameters, trained on 660 million samples across 512 and 1024 pixel resolutions, on 64 H100 GPUs in bf16 with FSDP2. The team is explicit that this is a proof of concept on the road to their real goal, cheap generative video.
It cannot render text, and it struggles with full-body poses, multiple entities, and complex interactions. The team admits as much in their own writeup, which honestly makes the rest of it easier to trust. Nobody at Linum is calling this a Midjourney killer.
The sting is in the number. For two years the default assumption said competing in generative image models costs nine figures. This one cost $21K. The gap between frontier and good enough is where entire business models live, and that gap just got cheaper to cross.
The technical bet: no VAE
Under the hood, PJT makes a stranger and more interesting choice than the price tag. Nearly every diffusion image pipeline you have used compresses images through a VAE, a variational autoencoder, before the model ever sees them. Linum skipped it and works directly in pixel space.
Their argument: VAEs hit an empirical ceiling on compression, and better VAE reconstruction can make downstream generations worse. Instead of fixing the compressor, they attacked cost from inside the model, with token reduction 4x more aggressive than their previous version, plus intermediate prediction heads that keep gradients flowing through the deeper stack.
The reported result is convergence on 11x fewer samples than before. Treat that multiplier as a lab-reported figure, since nobody has reproduced it independently. Still, my read is that a lot of the compute in generative training was buying convenience, not capability.
Why this number matters more than this model
Models are news for a week. Price floors are news for a decade.
If a two-person lab can produce a competent text-to-image model for $21K, then good-enough image generation is collapsing toward something a startup can put on a credit card. Competent, Apache 2.0, yours to fine-tune, though nowhere near frontier grade.
That has consequences for anyone whose product depends on image generation being expensive. If your moat is “we have the model,” the moat is melting from the bottom. What survives is everything around the model: distribution, workflow, brand, integration depth. Midjourney has taste and community. Adobe has the creative suite your team already lives in. A raw image API has a price list, and price lists are the first thing commoditized.
I keep coming back to what this does to build-versus-buy decisions. A year ago, building your own image model was laughable, so everyone rented an API. The honest answer now is that the crossover point moves every few months, and nobody renting should treat their API bill as fixed physics. It is a market price for something that keeps getting cheaper to own.
What to do with this today
Run the numbers yourself. The weights are on Hugging Face and the code is on GitHub, both Apache 2.0. If your product ships more than a few million generated images a year, spend an afternoon benchmarking a fine-tuned PJT-class model against your current API spend. Even if you stay on the API, you will negotiate better knowing the alternative exists.
Recheck the math quarterly. The cost floor moved by an order of magnitude while most teams were not looking. Whatever build-versus-buy math you did last year is stale.
Watch the video domino. Linum says images are the waypoint and video is the destination. If the image floor sits at $21K today, the same compression tricks applying to video is the uncomfortable question every stock footage and video tool vendor should be asking.
The caveats, stated plainly
This is a lab’s field-notes post, not a peer-reviewed paper, and nobody has independently reproduced the training cost yet. The 11x convergence claim is self-reported. The weak spots, meaning text rendering, poses, and multi-entity scenes, are real and disqualifying for plenty of use cases.
And one $21K run does not break an economy. Frontier quality still costs frontier money, and the labs spending it know exactly what they are buying.
But the direction is one-way. Every few months, some small team posts weights that would have been a funded startup’s entire pitch two years earlier. If your plan assumes image generation stays expensive, that assumption now has an expiry date.
