Mine StableDiffusion

Mine StableDiffusion is a native AI art generation application that operates completely offline, designed to run Stable Diffusion models efficiently on Android and desktop devices. Developed with Kotlin Multiplatform and Compose, it utilizes a C++ engine to deliver optimized performance with hardware acceleration (Vulkan and Metal), guaranteeing privacy by processing all operations locally on the user's device.

★★★★★

4.0(1 ratings)

• Text-to-image generation. This core function allows creating images from textual descriptions. The user enters a prompt, and the application, using the loaded Stable Diffusion model, generates a unique visual representation. It supports a wide range of models, from fast ones like SD-Turbo to high-quality ones like SDXL and FLUX, offering control over resolution and inference parameters.
• Support for custom models (GGUF). The application allows loading and using your own models in GGUF format. This provides the flexibility to employ specific fine-tunes, VAE adjustments, and LoRAs, allowing users to specialize generation in particular styles or themes beyond the included base models.
• Advanced parameter control. It offers a detailed configuration panel to fine-tune the generation process. Users can modify the number of inference steps, the CFG scale (which determines fidelity to the prompt), select different samplers, and control the random seed to be able to reproduce exact images or explore systematic variations.
• Offloading management (CPU/GPU). Includes options to offload specific model components, such as the CLIP encoder or the VAE decoder, from the GPU to the CPU. This functionality is critical for devices with limited video memory (VRAM), as it frees up GPU resources for the main part of the diffusion process at the cost of slightly slower speed.
• Weight type selection (wtype). Allows choosing the precision with which the model weights are stored in memory, from FP32 (maximum precision) down to 2-bit quantizations like Q2_K. This selection manages the trade-off between RAM/VRAM usage and final image quality, with the Auto option being recommended for most users.
• MMAP (Memory Mapping) activation. By activating this function, the model weights are mapped directly from disk instead of being fully loaded into RAM. This reduces initial memory consumption, although it may increase disk read activity during generation. On Android, it is enabled by default and can be disabled if slowness is experienced.
• Direct Concurrency Control (Direct Convolution). This is an experimental option that changes the convolution algorithm used in the diffusion model. On some specific hardware, it can provide a performance increase, although its effect may vary. It is recommended to test it to see if it improves speed on your particular device.
• Batch generation. The application supports generating multiple images in a single run. This allows the user to efficiently create several variants of the same prompt, saving time by not having to start the process manually for each desired image.
• Support for multiple aspect ratios. Before generation, the user can select from several predefined image ratios (such as square, landscape, or portrait). This flexibility is essential for adapting creations to different uses, such as wallpapers, illustrations, or formats for social media.
• Native cross-platform compatibility. The application is compiled natively for Android, Windows, Linux, and macOS. This means that on each operating system it leverages the specific graphics capabilities (Vulkan on Android, Linux, and Windows; Metal on macOS) to obtain the maximum possible performance from the available hardware.
• History and seed management. It keeps a record of generated images along with the parameters and seed used. This function is key to the creative workflow, as it allows you to regenerate an exact image or use a seed that produced an interesting result as a starting point for new explorations.
• Architecture based on C++ and JNI. The generation core is implemented in C++ using the stable-diffusion.cpp engine, which communicates with the Kotlin user interface through JNI (Java Native Interface). This separation ensures that the intensive processing is extremely fast and efficient, regardless of the platform.

The development of Mine StableDiffusion began in July 2025, with the first commit in the repository. The application is created and maintained primarily by the developer known as Onion99. It is programmed using a shared code approach: the business logic and user interface are written in Kotlin Multiplatform with Compose Multiplatform, while the inference engine and generation backend are written in C++ (approximately 5.7% of the code base).

Alternatives to Mine StableDiffusion:

Ollie IDE

Ollie is a sovereign creative suite designed for developers, writers, and creators who want the power of AI without sacrificing privacy or paying monthly subscriptions.

Price: $19   Size: 180 MB   Version: 2026.2.28   OS: Windows, Linux, MacOS

guIDE

guIDE is the first truly native, local IDE for artificial intelligence.

Price: Free   Size: 360 MB   Version: 2.0.0   OS: Windows, Linux

LM Studio

LM Studio is a desktop application that allows you to run large language models (LLMs) such as GPT-oss, Qwen3, Gemma3, and DeepSeek directly on a personal computer.

Price: Free   Size: 568 MB   Version: 0.4.8-1   OS: Windows, Linux, MacOS