Show HN: A decent text-to-image model trained for ~$20K (Apache 2.0 Weights)
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
Training a text-to-image model for just $21,000 is now viable using a pixel-space architecture that achieves a 32x32 token reduction, cutting the attention context window by 4x and accelerating training convergence by 11x. For production engineers, this Apache 2.0 release proves you can bypass complex latent-space VAE setups to deploy custom, fast-converging image and video generation pipelines at a fraction of standard GPU costs.
A functional text-to-image model trained from scratch for $21K using a pixel-space architecture (P-JiT) that represents a 512² image in the token count others use for 256² — a 4× smaller attention context window and 11× faster convergence than their prior DiT baseline, with Apache 2.0 weights. If the token-reduction claims hold, this points to dramatically cheaper diffusion inference and training, meaning you could self-host or fine-tune a capable image model on a modest budget rather than paying API rates — though it's an experimental release and a stepping stone, so expect quality and control gaps versus FLUX-class models.