At its core, this is the official release of , developed by RunwayML in collaboration with Stability AI. To understand the name, we have to break down the three technical suffixes:
The model paused longer this time. Then it produced an image: a child’s bedroom at twilight. A stuffed unicorn with a missing eye lay on the floor, its yarn mane tangled around a broken toy robot’s claw. A dusty rug was the “Martian” red. It wasn’t what Kai asked for. It was better .
And the ghost shows him exactly how to reply.
Dr. Elara Vance had created it two years ago, just before she left the project. She had trained the image generation model on a billion cat photos, then a million Renaissance paintings, then a hundred thousand pictures of rain on windows. The “pruned” meant she had cut away the redundant neural pathways. “EMA” stood for Exponential Moving Average—a smoother, calmer version of the model’s chaotic mind. “Only” meant there was nothing else.
EMA stands for . In machine learning, keeping an EMA version of the weights during training helps "smooth out" the model, making it more stable and less prone to "deep-fried" or glitchy results. For the end-user, "EMA-only" means you are getting the final, polished version of the model intended for generation, rather than a version meant for further training. Why is it Still So Popular?