What comes after Stable Diffusion? Stable Cascade could be Stability AI’s future text-to-image generative AI model

Stability AI, the company behind Stable Diffusion text-to-image AI, is teasing a new innovation dubbed Stable Cascade that promises to level up image generation. This novel architecture, inspired by the Würstchen model, aims to be more agile and precise than its predecessor, the SDXL lineup. The secret sauce? A latent diffusion technique that crafts a skeletal yet thorough image roadmap, cutting the computational mustard for top-notch results.

Unlike its monolithic ancestor, Stable Cascade provides a trio of smaller modules (Stages A, B, and C) that work together to generate hi-res images from bite-sized text prompts. By parting ways with the text-to-image fusion and image decoding tasks, Stable Cascade flexes its training muscles more efficiently. It’s like splitting a hefty task into bite-sized chunks; in this case, it’s not just efficient—it’s sixteen times more cost-effective!

With Direct Preference Optimization (DPO), image quality may improve further—an innovative tuning approach harnessed by Stability AI’s Emad Mostaque. Stable Cascade outshone incumbent leaders like SDXL, showcasing superior image quality and prompt accuracy while brandishing a surprisingly swifter inference speed despite being heavier by 1.4 billion parameters. Stable Cascade generates words into images flawlessly, setting a gold standard in the text-to-image arena.

Stable Cascade doesn’t stop at just image creation; it can remix images as well. From retooling existing images to advanced skills like in-painting and super-resolution.