Two Brothers Trained a Text-to-Image Model for $20,000 and Released the Weights Under Apache 2.0

Conversations about building a custom generative model usually end when the word “budget” comes up. Two brothers from the Linum lab decided to find out what a decent text-to-image model actually costs. The answer: about $20,000. They released the weights under Apache 2.0 and published the research on their website.

What They Built

The industry standard is latent diffusion. First, a VAE compresses an image into a compact latent space; then, a diffusion model learns to generate within that space. That means two training stages, two models, and two opportunities for something to go wrong.

Linum removed the VAE. The diffusion model operates directly in pixel space and learns the representation best suited to its task on its own. The architecture is called Pyramid-JiT. The model was trained for roughly a month as an extended experiment: ~$20K and 660 million samples.

Why It Matters

The authors cite three benefits. First, the architecture is simpler: there is no need to train two components sequentially. Second, according to their data, convergence is 11 times faster than with the latent counterpart. Third, inference is cheaper: the attention context can be reduced 4 times more aggressively without a noticeable decline in quality.

Imagine a startup generating marketing illustrations. In the past, a custom model was a luxury: cluster rental, a team of researchers, and months of work. Now there is a public recipe that costs about as much as one small round of experiments. It is not a replacement for Midjourney, but it is already a sufficient foundation for a narrowly specialized model.

There is a more ambitious goal behind the work as well. Linum is building video models, and almost all video models begin as image models. The authors joke that models learn nouns first and verbs later. If images become inexpensive to generate, video will become cheaper too—and speed and cost are precisely what currently prevent video from being used in iterative creative workflows.

Limitations

A few caveats up front. This is a Show HN project from a tiny team, not an independent benchmark. The authors themselves describe the quality as “decent” and say they were surprised by the result; the post contains no comparison tables against market leaders. The claims of an 11-fold speedup and a 4-fold token reduction come from their own comparisons with their own latent variant, so they should not be generalized to other tasks. The $20K figure also covers only the final run: judging by the description, it excludes the cost of earlier experiments and architecture selection.

The Hacker News discussion remains modest so far: around 110 points and 11 comments. There is no strong consensus, so this is better viewed as an interesting signal than a proven breakthrough.

What to Do With It

If you are building an image-generation product, download the weights and test them on your own prompts: the Apache 2.0 license allows unrestricted commercial use. If you are a researcher or ML engineer, read the Pyramid-JiT paper and consider whether your next project actually needs a VAE. And if you are a product leader, remember the main takeaway: the barrier to entry for custom generative models has fallen again, and “it is too expensive” is becoming a weaker argument.