To use Stable Diffusion on a phone, first decide whether generation will run on-device, on your own computer, or in the cloud. Then start with one supported model, a simple prompt, a moderate image size, and one-variable-at-a-time iteration.
The short answer
A beginner phone workflow has five stages: choose where the model runs, install or connect to one supported model, write a prompt with a clear subject and composition, generate at a moderate size, and change one variable at a time. Save the prompt and seed when the composition is promising, then edit or upscale only the result worth finishing.
Do not begin by collecting models, copying a huge negative prompt, or changing every setting. Those actions create more variables before you know what the baseline model can do.
1. Decide where generation actually runs
On-device means the phone executes the model. Remote means the phone controls a computer or home server. Cloud means a provider executes the request. Each can be a good phone experience, but they have different privacy, connectivity, hardware, cost, and setup requirements.
If offline use and keeping prompts off a rendering server matter, choose an on-device app and test it in airplane mode after setup. If arbitrary checkpoints, LoRAs, nodes, or batch workflows matter more, a secured desktop or cloud interface may be the better tool.
2. Build a prompt you can diagnose
Use a visible structure: specific subject, action, environment, medium, light, framing, and palette. For example: ‘small red sailboat crossing a quiet alpine lake, editorial photograph, low morning light, wide composition, muted blue and red palette.’ The model receives visual instructions rather than a vague request for quality.
Generate before adding a negative prompt. If a repeated defect appears, target it narrowly. An enormous generic negative list can change style and composition while hiding which term affected the image.
3. Preserve the baseline before iterating
Save the prompt, seed, model, image size, scheduler or preset, and step count. If the subject is wrong, rewrite the subject. If framing is wrong, change the composition phrase. If the image is close, keep the seed and change only the weakest instruction.
A seed can help reproduce a starting noise pattern within the same model and runtime. It is not a universal image ID: different apps, model versions, schedulers, and implementations can produce different pixels from the same seed.
4. Edit the smallest useful area
Use image-to-image when the whole composition may change. Use inpainting when only a selected area should change. Check the result at its intended viewing size before upscaling; a larger file does not repair a weak subject or composition.
Expect limitations. Diffusion models can struggle with exact lettering, repeated characters, anatomy, brand identity, and tiny controlled details. Keep the best base image and use a targeted edit or design tool instead of asking one generation to solve everything.
PhoneDiffusion availability and article scope
PhoneDiffusion is an iPhone and iPad app for iOS 17 or later. It uses curated on-device model packs and currently provides supported text-to-image, image-to-image, inpainting, 4× upscaling, and background-removal workflows when the required pack is installed.
It is not a general importer for desktop checkpoints, LoRAs, ControlNet files, or ComfyUI graphs. Use the model and controls visible in the current app; move to a secured desktop workflow when an unsupported extension is essential.
Field notes
- Choose on-device, remote, or cloud generation deliberately; a phone screen does not tell you where inference runs.
- A simple prompt and saved baseline make iteration faster than changing every setting.
- Edit and upscale after choosing a composition, not before.
Try it on your phone
- Choose one supported model and a moderate square or destination-shaped image size.
- Write a prompt with subject, action, environment, medium, light, framing, and palette.
- Generate once and save the prompt, seed, model, size, and preset.
- Change one weak variable, compare at the same display size, and keep the stronger result.
- Use a targeted edit, then upscale only the final composition if needed.
Sources and verification
Primary and maintainer documentation checked for this guide. Product-specific details were reviewed against the current PhoneDiffusion scope.
- CompVis — Stable Diffusion reference implementationOriginal project documentation for latent text-to-image generation, prompts, seeds, dimensions, steps, and model limitations.
- Hugging Face Diffusers — text-to-image pipelineMaintainer documentation for prompt guidance, negative prompts, deterministic seeds, and text-to-image controls.
- Apple Developer — Core ML overviewPrimary documentation for on-device model execution, compute resources, networking, and privacy context.