They Trained a Text-to-Image Model From Scratch in 24 Hours for $1,500

A team just demonstrated you don’t need a war chest of millions to build a competitive image generator. In a 24-hour speedrun on 32 H200 GPUs, with a total compute budget of roughly $1,500, they stacked a series of modern architectural tricks to train a text-to-image model from scratch. This isn’t a theoretical exercise; it’s a practical blueprint they’re open-sourcing, complete with training code and the experimental framework used to validate each component. The result shows how drastically the economics of generative AI have shifted from the early Stable Diffusion days.

The recipe ditches the standard variational autoencoder (VAE) to train directly in pixel space using x-prediction, a method from recent research that simplifies the whole pipeline. Starting straight at 512px resolution, they bypass the usual progressive growing schedule. To boost visual quality without a VAE, they bolted on classical perceptual losses—LPIPS and a DINOv2-based loss—directly to the pixel outputs. The team noted, ‘Once you predict pixels directly, everything becomes straightforward again.’ These lightweight losses add minimal overhead but provide a consistent signal that improves convergence speed.

To keep the computational cost manageable, they implemented token routing with a technique called TREAD. It randomly selects half of the tokens to bypass a large chunk of the transformer, re-injecting them later so no information is permanently dropped. This slashed the per-step cost. They also tweaked the guidance mechanism, finding that standard classifier-free guidance (CFG) made routed models look worse, especially when undertrained. Their solution was a custom self-guidance scheme to clean up the output.

The key takeaway isn’t just that a small team can now spin up a model over a weekend. It’s that the combination of pixel-space training and classic vision tools is genuinely competitive with more complex latent-space methods. I’m skeptical that this exact recipe will dethrone the best proprietary systems tomorrow, but it redefines the floor for what’s possible. The fact that the entire process—from a blank slate to a working model—cost less than a used Honda Civic signals a profound democratization of generative model research that’s worth paying very close attention to.