A text-to-image model built entirely from scratch — our own VAE and diffusion transformer, trained from zero on a self-curated ~3.2M-image set, on one desktop GPU. No borrowed backbone, no licence, no royalties.
One person. One GPU. A foundation image model built from zero — no fine-tune, no borrowed backbone.
A 16-channel latent autoencoder with self-attention — ~12:1 compression, not a borrowed Stable-Diffusion VAE.
An attention-based diffusion transformer, flow-matching, dual-stream text conditioning — written and trained from scratch.
~3.2M images curated from dozens of sources, captioned and filtered in-house (2.25M active in the current stage) — the corpus is ours too.
EMA weights + REPA representation alignment for clean, stable convergence at long horizons.
Full self-attention with 2D axial RoPE positions and dual-stream text↔image attention, plus register tokens — memory-efficient SDPA kernels keep every layer inside one card’s VRAM.
This is the work of a lab and a cluster — done on a desk, by one hand.
This entire stack — VAE, transformer, and data pipeline — was built and trained from scratch by one person on a single GPU. The hard, uncertain part — proving the architecture learns from zero — is already done. The training set has since grown from 1.1M to ~3.2M curated images and quality keeps climbing with every step; it scales directly with more compute and more data. What’s left is no longer a research risk. It’s resources and people.
The same recipe carries to billions of parameters and a richer latent — sharper faces, hands and fine detail. It just needs the compute.
Multi-GPU training reaches 1024px+ and cuts weeks of convergence to days.
The edit / reference-image stack is already architected on this base — a direct, near-term path to a revenue-generating fashion try-on product.
One person got it this far. With a few strong engineers, the same stack becomes a shipping product — everything a generative-media company needs to own.
If this is what one GPU makes, imagine the next order of magnitude.