Onboarding
|
Text Generation
|
Image Generation
|
Vision
|
Attachments
|
You don't need Midjourney or DALL-E to generate AI images. Your Android phone can run Stable Diffusion entirely on device. No subscription, no internet, no uploading your prompts to a server.
Off Grid is a free, open-source app that runs Stable Diffusion on your phone with NPU acceleration on Snapdragon. 5 to 10 seconds per image on flagship devices. This guide covers how it works, what to expect, and how to get the best results.
How On-Device Image Generation Works
Cloud image generation services run massive models on data center GPUs with 80GB+ of VRAM. On-device image generation runs optimized versions of the same Stable Diffusion models on your phone's processor.
The model starts with random noise and gradually 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 blank screen.
A typical generation is 512x512 pixels at 20 denoising steps. That produces a usable image in 5 to 30 seconds depending on your hardware.
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 acceleration is a game changer. What takes 30 seconds on CPU takes 5 to 10 seconds on the Snapdragon NPU.
Storage: Stable Diffusion models range from about 1GB (compressed) to 4GB+ (full precision). You'll need at least one model downloaded.
Real World Performance
This is the biggest variable and the NPU makes a massive difference.
Snapdragon 8 Gen 3 with QNN NPU: 5 to 10 seconds per image at 512x512, 20 steps. Power efficient. The phone stays cool.
Snapdragon 8 Gen 2 with QNN NPU: 8 to 15 seconds. Still very usable.
Flagship CPU only (no NPU or unsupported chip): 15 to 30 seconds. Phone gets warm. Battery drain is noticeable.
Mid-range CPU: 30 to 60 seconds. Usable but not for generating dozens of images.
Off Grid detects your hardware automatically. If you have a Snapdragon NPU, it uses QNN acceleration. If not, it falls back to MNN (Alibaba's mobile inference framework) on CPU. You don't have to configure anything.
20+ Models to Choose From
Off Grid includes a model browser with over 20 Stable Diffusion models:
Absolute Reality for photorealistic output. DreamShaper for a balanced artistic mix. Anything V5 for anime and illustration style. Plus many more sorted by style and device compatibility.
For most devices, a compressed model in the 1 to 2GB range gives you the best balance of quality and speed. Full precision models (4GB+) look better but need 12GB+ RAM.
AI Prompt Enhancement
This is the feature that dramatically improves output quality. A simple prompt like "a dog" produces generic results from Stable Diffusion. But Off Grid also runs LLMs on device, and it can chain them together.
Type a simple prompt, and the app runs it through your loaded text model first. The text model expands "a dog" into a detailed 75-word description with artistic style, lighting, composition, and quality modifiers. That enhanced prompt goes to Stable Diffusion, and the output quality difference is dramatic.
This is something cloud image generators do behind the scenes. Off Grid does it transparently, on device, and you can see exactly what the enhanced prompt looks like.
Tips for Better Results
Always use prompt enhancement. The quality difference is immediately visible. Let the text model do the creative heavy lifting.
Start with 20 denoising steps. More steps improve quality with diminishing returns. Going from 20 to 30 adds 50% more time for maybe 10% better output. 20 is the sweet spot for mobile.
512x512 is the practical ceiling. Higher resolutions multiply computation. 512x512 at 20 steps looks good on a phone screen and in most digital contexts.
Close other apps before generating. Image generation uses a lot of RAM. If other apps compete for memory, Android might kill the process silently.
Watch for thermal throttling. Multiple images in a row heat up the phone. Give it a minute between batches if you notice slowdown.
Privacy for Creative Work
Every image you generate on Midjourney, DALL-E, or similar services is stored on their servers. Your prompts are logged. Your images may be used for training.
Off Grid means your prompts and images never leave your phone. There's no server, no logging, no possibility of your creative process being used to train someone else's model. For professional artists, designers, or anyone who values creative privacy, this matters.
Open source, MIT licensed. Verify it yourself.
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 does 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)