Stable Diffusion KMP

What makes SDAI stand apart is that it treats the provider as a swappable slot rather than the product. In one app you can queue a job against your own Automatic1111 instance, switch to Stability AI or Fal.ai for a higher-quality render, drop to AI Horde or Hugging Face Inference when you don't want to spend credits, or cut to on-device inference via Android ONNX, MediaPipe, stable-diffusion.cpp SDXL, iOS Silicon Diffusion Core ML, or PrismML Bonsai Image 4B when you have no connection and want nothing leaving the device. That breadth is genuinely rare on mobile and is the reason to pick SDAI over a single-backend client. The controls are also deeper than most mobile generators allow. You get tagged positive and negative prompts, dimension and sampler selection, seed, CFG scale, sampling steps, restore faces, batch size, and provider-specific options, plus LoRA, embeddings and hypernetworks where the backend supports them. Img2Img and Inpaint are built in, and any gallery image can be sent back into either mode to reuse its prompt and context. The built-in gallery keeps generation details inspectable and supports export and sharing, so the loop from prompt to finished image never leaves the phone. There are real edges. It is mobile-only — there is no web or desktop UI, so it complements rather than replaces a desktop ComfyUI or A1111 workspace. On-device output quality tracks your hardware, and model support varies by platform and engine. The app leans toward users who already understand the Stable Diffusion vocabulary — CFG, samplers, embeddings — and low-end phones will struggle with local inference. Advanced tools beyond generation, like ControlNet or upscalers, are not part of the feature set the vendor documents. Where it shines is the middle ground: creators who run their own server and want phone access, privacy-minded users who want local generation with no cloud upload, and open-source contributors, since the Kotlin Multiplatform codebase and its API documentation — expanded on May 26, 2026 to cover every module from app and core/* through domain, network, storage, and the feature/* engines — are all public on GitHub. Tinkerers get the most here; one-tap users will not.