You don't need Midjourney. You don't need a $20/month subscription. You don't need to upload your prompts to someone else's server. Your Android phone can run Stable Diffusion entirely on device and generate images in 5 to 10 seconds.
Off Grid is a free, open-source app that runs AI image generation on your phone's hardware. After you download a model once, the app never contacts any server again. Your prompts and your images exist only on your phone.
What You Need
Minimum: 6GB RAM, any recent ARM64 processor. CPU-only generation works but expect 30 to 60 seconds per image.
Recommended: Snapdragon 8 Gen 1 or newer. The NPU (Neural Processing Unit) is a dedicated AI chip on the Snapdragon that accelerates image generation dramatically. What takes 30 seconds on CPU takes 5 to 10 seconds on the NPU.
Storage: Stable Diffusion models range from about 1GB (compressed) to 4GB+ (full precision). You'll need at least one downloaded.
How It Works
Stable Diffusion starts with random noise and refines it into an image over multiple "denoising steps." Each step gets the image closer to what your prompt described. Off Grid shows you a real-time preview during this process so you can watch the image form instead of staring at a loading screen.
A typical generation is 512x512 pixels at 20 denoising steps. That produces a clean, detailed image in seconds on supported hardware. All of this happens on your phone's processor. No server is involved after the initial model download.
Image Generation
|
Real World Performance
The NPU is the difference between "usable" and "fast."
Snapdragon 8 Gen 3 with QNN NPU: 5 to 8 seconds per image at 512x512, 20 steps. The phone stays cool. Battery drain is minimal.
Snapdragon 8 Gen 2 with QNN NPU: 8 to 12 seconds. Still very smooth.
Snapdragon 8 Gen 1 with QNN NPU: 10 to 15 seconds. Completely usable.
Flagship CPU only (no NPU or unsupported chip): 20 to 40 seconds. Phone gets warm after multiple images.
Mid-range CPU: 30 to 60 seconds. Works but you won't be generating dozens in a row.
Off Grid detects your hardware automatically. If you have a supported Snapdragon NPU, it uses QNN acceleration. Otherwise it falls back to CPU through MNN. You don't configure anything.
20+ Models to Choose From
Off Grid includes a model browser with over 20 Stable Diffusion models sorted by style and device compatibility:
Absolute Reality for photorealistic output. Faces, landscapes, product shots. The closest you'll get to a photograph from a text prompt.
DreamShaper for a balanced artistic style. Good default if you're not sure what you want.
Anything V5 for anime and illustration. Clean lines, vivid colors, consistent character style.
Plus many more. The browser shows which models are compatible with your device so you never download something that won't run.
The Feature That Changes Everything: AI Prompt Enhancement
This is what separates Off Grid from a basic Stable Diffusion wrapper.
Type a simple prompt like "a dog." Raw Stable Diffusion will give you a generic, flat image. But Off Grid also runs LLMs on device, and it chains them together with image generation.
When you enable prompt enhancement, your simple prompt goes through the loaded text model first. It expands "a dog" into something like "a golden retriever sitting in autumn leaves, warm afternoon light filtering through trees, shallow depth of field, soft bokeh background, professional pet photography, natural colors." That enhanced prompt goes to Stable Diffusion, and the output quality difference is dramatic.
You can see exactly what the enhanced prompt looks like before generation starts. Cloud image generators do this behind the scenes. Off Grid does it transparently, on device, and you control every step.
Tips for Better Results
Always use prompt enhancement. The quality difference is immediately visible. A three-word prompt enhanced by the text model produces better images than a manually written 50-word prompt most of the time.
20 denoising steps is the sweet spot. More steps improve quality with diminishing returns. Going from 20 to 30 adds 50% more time for maybe 10% better output.
512x512 resolution. The standard for on-device Stable Diffusion. Higher resolutions multiply computation time and memory. At phone viewing distances, 512x512 looks sharp.
Negative prompts help. Adding "blurry, low quality, distorted, deformed" pushes the model away from common artifacts. Small effort, noticeable improvement.
Close other apps before generating. Image generation uses a lot of RAM. If other apps compete for memory, Android might kill the process.
Give your phone a break between batches. Multiple images in a row heat up the phone. A 30-second pause between generations keeps performance consistent.
Privacy for Creative Work
Every image you generate on cloud services is stored on their servers. Your prompts are logged. Depending on the terms of service, your generated images and the prompts that created them may be used to train future models.
With Off Grid, your prompts and images exist only on your phone. There's no server, no logging, no possibility of your creative process being used to improve someone else's product. For professional artists, designers, or anyone who values creative privacy, this matters.
Open source. MIT licensed. No analytics, no telemetry, no accounts.
Getting Started
- Install Off Grid from the Play Store
- Download a Stable Diffusion model from the in-app browser (1 to 4GB)
- Switch to image generation mode
- Type a prompt or let the AI enhance it for you
- Watch the real-time preview as your image generates
If you have a Snapdragon 8 Gen 1+ device, NPU acceleration kicks in automatically.
Off Grid also runs text generation, voice transcription, vision, tool calling, and document analysis. All offline, all in the same app. Check the GitHub for the latest updates.
Top comments (0)