Stable Diffusion 1.4

Stable Diffusion 1.4 is a text-to-image latent diffusion model: it starts from noise and iteratively forms an image guided by a text prompt, doing much of the work in a compressed latent representation rather than directly in pixels. The cited checkpoint is published in CompVis’s repository and builds on the Latent Diffusion research from LMU Munich; its model card credits Stability AI support. It is a generative image model, not an image classifier. TT-Metal’s matrix specifies 512x512 output and batch 1, with alternative single-device runs on n150 Wormhole and p150 Blackhole. Those two hardware rows are one model checkpoint at the same image resolution, collapsed here without retaining brittle seconds-per-image measurements.