Behind the buildings.
Building Studio runs SD-Turbo (lighter) or Z-Image-Turbo (best quality) locally with WebGPU. It downloads software and model files when you start the model. Briefs, prompts and generated images are processed on your device.
- Z-Image-Turbo by Tongyi-MAI, licensed under Apache 2.0.
- Browser ONNX adaptation by the WebNN team, pinned to revision
02e1d8c. The prompt encoder and image transformer use Q4F16 quantisation (4-bit weights with float16 computation); decoding and image screening use float16. Includes the model’s image screening components. - SD-Turbo by Stability AI, in the browser ONNX edition by Microsoft, pinned to revision
fceebfa, under the Stability AI Non-Commercial Research Community License. Its CLIP tokenizer files come from the same repository. Image screening reuses the Z-Image-Turbo screening models. - Pipelines and scheduler adapted from the Microsoft WebNN Developer Preview (Z-Image-Turbo and SD-Turbo samples), commit
7dea1d0, under the MIT license. - ONNX Runtime Web 1.29.0, MIT license; Transformers.js 3.8.1, Apache 2.0 license, for tokenisation.
- Tokenizer assets from the Microsoft reference sample, used for the Qwen3 prompt encoder.
- Studio page and engine first built for the Kasterpillar Image Studio.
Model and runtime downloads contact Hugging Face and jsDelivr. Browser caching is best effort. Image generation needs enough GPU memory and may fail on devices that pass the preliminary support check.