Isekai Engine

Loading worlds…

How this works

Each world is a JSON file: characters, style, an opening, and for plotted worlds a story spine of acts and major events with entry flags, variants, and a few pivotal choices. Every turn runs two model calls. The narrator writes the scene from the world card, the current event's brief, the story flags, recent memory notes, and the last turns. The director then returns a small JSON record: whether the event's goal was met, which flags to set, a one-line memory note, and an image prompt for a future scene illustrator. Ordinary choices change the scene; only pivot choices change which version of later events you get. The illustrator runs LCM Dreamshaper v7, a Stable Diffusion 1.5 model distilled to 4 steps (MIT and CreativeML OpenRAIL-M licenses), through ONNX Runtime on WebGPU, and draws one picture per turn from the director's image prompt. It needs a GPU with 16-bit float support and about 8 GB of memory alongside the text model. Use a recommended model: anything under about 2B parameters cannot follow the story brief and will wander. With the in-browser engine the weights download once from the MLC CDN, are cached, and run on your GPU through WebGPU, so nothing leaves your machine. Saves live in this browser; use Export to keep one.

Storage on this device

Everything below lives in this browser only. Models are cached so the next load is instant; clearing one means downloading it again.