On-Device Image Generation with Stable Diffusion

Stable Diffusion, running on the phone. On Android through a stable-diffusion.cpp native module; on iPhone through Core ML with Apple's palettized pipelines. No upload, no credits, no queue, and no month where the subscription is charged whether you drew anything or not.

The checkpoints

SD 1.5 with a TAESD decoder for speed, DreamShaper 8 LCM for very few steps, Realistic Vision v6 and Absolute Reality for photographic work, and Core ML builds of SD 1.5 and 2.1 on iOS. The model manager lists them with their sizes; they are around a gigabyte each, so they are downloaded when you ask for them rather than during onboarding.

What a phone can honestly do

A mid-tier device or better. A generation takes tens of seconds rather than the second or two a datacentre GPU manages, and the LCM checkpoint exists precisely to cut the step count for that reason. What you get for the wait is unlimited: no per-image cost, no daily cap, no content review of your prompt, and no copy of what you made sitting on somebody's storage.

The model can draw, too

Image generation is exposed as a tool, so the assistant can call it mid-conversation — ask for a header image for something you are writing and the picture appears in the reply, without switching screens.

Heat, and the guard on it

Diffusion is the heaviest thing this app does. BatteryGuard gates every generation: on a low battery or a hot phone it refuses and says why, rather than letting you cook the device to make a picture. The diffusion pipeline is also never co-resident with a text or vision model — see model swapping.

Availability

Behind the one-time $2.99 unlock in Lite (included in the seven-day trial), and unconditional in the $2 paid app.