Stable Cascade

Stable Cascade, announced by Stability AI in February 2024, took a different architectural bet than the rest of the SD family: a three-stage cascade built on the Würstchen architecture. Stage C converts the text prompt into tiny 24x24 latents, a compression factor of 42, versus 8x for Stable Diffusion, and Stages B and A decode those into the final image. The payoff is cheaper training and faster inference: Stability reported finished images in about 10 seconds for 30 steps against SDXL's 22 seconds for 50. It shipped in a research preview with a 3.6B-parameter Stage C variant recommended for best quality, plus full training and inference code on the Stability-AI GitHub.

The catch is the license: weights are under the stable-cascade-nc-community license, non-commercial only. Stability's own human evaluation (parti-prompts plus aesthetic prompts, 30 steps) claimed it beat Playground v2, SDXL, SDXL Turbo, and Würstchen v2 on prompt alignment and aesthetics in most comparisons, though those numbers are provider-reported and the model never got a clean independent rerun. Development went quiet as Stability pivoted toward SD3 and SD3.5, and the ~20GB VRAM requirement kept it out of the casual hobbyist tier. It remains an important architecture reference even though few people run it today.

Provider-reported human eval on parti-prompts + aesthetic prompts: best in prompt alignment and aesthetics vs Playground v2, SDXL, SDXL Turbo, Würstchen v2 in most comparisons (Stability AI, 2024)