A two-person research lab has trained a text-to-image model for roughly $21,000 — and released it under a fully permissive license. Linum, founded by co-CEOs Sahil Chopra and Manu Chopra, published the results on November 4 in a technical field-notes post, detailing a 2-billion-parameter Pixel-JiT (P-JiT) model trained on 660 million samples as a proof-of-concept on the road to larger video models. Both the model code and the weights are available under Apache 2.0, a license that permits commercial use without restriction. It is a striking data point in the debate over how much frontier-quality AI research actually costs, a topic we follow in our AI research coverage.

Why Skip the VAE?

Most modern image-generation pipelines compress images using a separate variational autoencoder (VAE): the VAE squeezes an image into a compact latent space, and a diffusion model generates inside that compressed space. Linum's team argues this division of labor has hit a ceiling. In their write-up, they note that VAEs "seem to have an empirical ceiling on the amount of compression they can achieve," that pushing compression further requires substantial surgery to the traditional architecture, and that improving a VAE's reconstruction can actually worsen the downstream diffusion model's generations.

Their answer is to remove the VAE altogether. P-JiT is a pixel-space model — the diffusion model operates directly on pixels and learns a latent purpose-built for its own use case, compressing and generating at the same time. The bet is that a single model trained end-to-end can squeeze more out of the same compute than a hand-off between two specialized networks.

The Technical Payoff

The architecture's name refers to its pyramid-shaped tokenization. Where most current models — Linum cites FLUX-2 as a typical example — use a 16×16 token reduction, P-JiT achieves a 32×32 token reduction, representing a 512×512 image in the token budget previously needed for a 256×256 image. That translates to a 4× reduction in the attention context window, and attention is precisely the O(n²) bottleneck that makes both training and inference expensive for generative video.

The results, by the team's own measurements: P-JiT converges in 11× fewer samples than Linum v2, the lab's previous model built on the more standard Wan VAE plus DiT diffusion recipe. The $21K run — 660 million training samples — produced a model the team describes as surprisingly competent given its size.

What $21K Buys

The release includes 249 sample images with their prompts, spanning people (57), animals (42), buildings (20), products (43), landscapes (19), and stylized renders (68). The team is candid about positioning: this is an "experimental release" and a "stepping stone" toward their real target, a video model. The samples demonstrate that the architecture trains stably and produces coherent images at a fraction of conventional training budgets, not that it matches frontier image models on quality.

The write-up is also notable for how the team documents failure modes alongside results. They enumerate their model's limitations directly in the field-notes post rather than burying them, an increasingly common practice among smaller labs whose reputations depend on credibility rather than marketing budgets. The post is the fourth entry in the team's "Road to Linum v3" series, building on an earlier iteration where the P-JiT architecture was first described — a running technical log that lets outside researchers follow each design decision, including the ones that did not work.

For context on the cost figure: Linum's $21K covered the compute for a 660-million-sample training run of the 2B-parameter model. That is not directly comparable to frontier-lab budgets, which span data acquisition, staffing, and experimentation over many aborted runs. What it does establish is the marginal cost of one successful, openly documented training run — a number individual researchers and university labs can now realistically plan around.

Linum's stated ambition is creative tools that democratize access to animation. The team's critique of today's generative video is that it is too slow and expensive, and lacks fine-grained controls — a combination they argue renders it "effectively useless for storytelling," since creative work is inherently iterative: you make, you tweak, you get better. Cheaper training and faster inference are, in their framing, the precondition for video generation that actually fits an editing workflow.

Small Labs, Big Signals

The economics on display here matter beyond one startup. Training runs that once defined the field's frontier — measured in millions of dollars — are now being matched, in narrow domains, by teams of two spending less than a typical senior engineer's annual salary. Open-weights releases at this scale give independent researchers the ability to audit, fine-tune, and build on production-grade architectures without institutional backing.

The caveats are fair to state: a 2B text-to-image model is far from the frontier of image quality, and Linum's convergence claims are self-reported rather than independently benchmarked. But the release is fully reproducible — code and weights under Apache 2.0 — which means anyone skeptical can measure the claims directly. That, more than any single benchmark, is what makes small-lab releases like this one worth watching: the proof is downloadable.

Stay Ahead of AI

Research that once required a corporate lab now fits on a credit card. Track every breakthrough in our AI research news.

Read more AI news →