About

物作り. Monozukuri. Making things — as a discipline.

We build generative models that render the world. Diffusion. Image. Video. The aim is to push what currently looks like a sample of reality until you can't tell the difference.

In the day-to-day, the work is what happens when the model wants more memory than the card has. When attention is too slow. When the data we need doesn't exist in any clean form. When no eval will tell us whether we're winning.

Most labs route around those. We treat them as the next thing to build.

We needed to run three diffusion training jobs on a card that comfortably fits one. The brute-force answer was buy more cards. We did the math. The cost curve grew faster than the work, and every dollar in extra hardware was a dollar not going to data, evals, and the people who actually move science forward. So we invented ABST — stream the frozen backbone between CPU and GPU one block at a time, bit-identical gradients, three to four times more concurrent training per card. The constraint became the brief. The brief produced the method.

The bet underneath all of it: the labs that build the next generation of generative models won't be the ones with the largest clusters. They'll be the ones that learned to design under constraint — because the model that wins is the one trained on the right signal, evaluated against the right standard, on hardware whose limits forced cleaner architecture. Money buys scale. It doesn't buy the methods that make scale matter.

Three positions we don't move on:

Taste is trainable. The field treats taste as subjective and therefore untrainable. We disagree. The same rigor labs apply to alignment for safety, we apply to alignment for creative quality. There are no public datasets for that, no agreed evals, no shared vocabulary. We're building all three.

Honest data over clean data. Models trained on polished data produce polished mediocrity. The signal we want is messy, contradictory, unstructured — what real people say, buy, and ignore. The pipelines for that don't exist off the shelf. We built our own.

Constraint is the design brief. Cheaper compute doesn't make better models. Tighter constraints force better designs. Our fleet is small on purpose. Every method we publish started as a wall — and is the tool we built to keep going.

We're a lab, not a product team. The output is research — the model, the method, the eval. Published, reproducible, ours.

— The Caistro Labs Team