You have a picture in mind, but typing a detailed prompt on your Android phone is awkward. Describe the scene aloud instead. You can turn those spoken words into an image without uploading them to a cloud image service. OGAM (Off Grid AI Mobile) turns your speech into a prompt, then passes the reviewed text to a downloaded local image model. The workflow runs on your phone after the required models are installed.
Get OGAM for Android | Current mobile release
This guide uses Chat mode: speak, check the prompt, then send. It gives you a chance to fix the subject or color before you spend time generating the image. You do not need spoken AI replies for this workflow.
What do you need before you disconnect?
You need Android 10 or later and at least 4 GB of RAM, microphone permission, an on-device transcription model, and a compatible downloaded image model. Start with an image model recommended for your device; larger models can exceed available memory.
Local image generation and microphone dictation are free features. The spoken-reply Audio interface is a separate Pro feature and is not required here. The mobile feature list separates these capabilities.
On Android, image model compatibility depends on the phone and selected runtime. Use the catalog recommendations and compatibility labels. Do not assume that every Android phone has a supported NPU or that one model package fits every chipset.
Complete the app installation and model downloads while connected. If you use a text model to improve prompts, download that model too. For the first test, turn prompt enhancement off so you can see what your own description produces.
How do you prepare the two local models?
Download the speech model and image model separately. One recognizes what you say; the other creates the picture. Keep both selections on-device. A local microphone model does not make a remotely selected image service work offline.
- Open Model Settings > Transcription (Speech to Text) > Transcription model.
- Download and select an on-device speech model. Base, 99 languages, at about 142 MB, is a small starting choice.
- Set your spoken language in Chat Settings > SPEECH TO TEXT > Language.
- Open Models > Image and download a compatible image model. Select that local image model for generation.
- In the image settings, set Enhance Image Prompts to OFF for the first attempt.
Keep the model's suggested image settings for your first attempt. You can change the size after you have one working result.
How do you generate the image with your voice?
Use Chat mode and set Image Gen to ON before dictating. Record a short description, inspect the text in the message box, then tap Send. This explicitly requests image generation instead of relying on the app to infer your intent from the wording.
1. Choose image output before recording
With the composer empty, open its settings control. Tap Image Gen until its badge shows ON. The available states are Auto, ON, and OFF.
Use ON for this first image. It makes the next request produce an image. After sending, the control returns to Auto; select ON again when you want another image.
2. Record a short description
Hold the app's microphone control, speak, then release it. Grant microphone access if prompted. Use OGAM's microphone rather than the keyboard's dictation button.
Try:
A flat illustration of a green bicycle beside a yellow wall, soft afternoon shadows, no text.
3. Check the prompt
The transcript appears in the message box. Correct the main subject, colors, and any words the speech model missed. Keep the prompt focused on what should be visible.
The Chat-mode dictation workflow keeps these words editable before sending.
4. Send and inspect the result
Tap Send. Wait for the image-generation result in the conversation. Compare the picture with your description. Image models can miss details or produce unwanted objects even when the prompt is correct.
For another attempt, change one part of the description. For example, change the wall from yellow to blue, or ask for a pencil drawing instead of a flat illustration. Do not treat another generation as an exact edit of the previous image.
How do you check that it works offline?
After setup, enable airplane mode and switch Wi-Fi off. Repeat one short dictated prompt with Image Gen ON. You should get a transcript followed by an image. Check that no remote speech or image provider is selected. If you later enable prompt enhancement, its text model must also be local.
What should you check if no image appears?
| Problem | Check | Action |
|---|---|---|
| No words appear | Microphone permission | Open Android app permissions and allow Microphone access for OGAM. |
| Important words are wrong | Speech language and recording clarity | Set the language, reduce background noise, and correct the transcript. |
| You receive a written reply | Image Gen state | Before a new prompt, set Image Gen to ON. |
| The app asks for an image model | Download and active selection | Finish a compatible local model download and select it. |
| Generation runs out of memory | Other loaded models and image settings | Unload unused models, keep prompt enhancement off, and use compatible settings. |
| It works only online | Remote model selections or incomplete files | Select local speech and image models and finish setup before disconnecting. |
Can you hear when the image request finishes?
With Pro, you can use OGAM's Audio interface and a local voice model for spoken replies. When an image generation finishes, the voice workflow can read the result message aloud. The normal generated-image message announces the image request; it is not a visual description or an accuracy check of the picture.
Download and select the local voice model under Models > Voice, then choose its language and voice in Chat Settings > TEXT TO SPEECH. Select Audio under Interface Mode. Keep the speech, text, image, and voice models local and fully downloaded if you want the complete session to work without internet.
Start with a short request such as “Generate an image of a green bicycle beside a yellow wall.” Wait for the image to finish and the result message to be spoken. Then inspect the image in the conversation. A spoken result tells you the turn completed; it does not prove that the image matches your brief.
If you see the image but hear nothing, check the selected voice model, Audio mode, and playback state. A result with no speakable text can finish silently. Use Chat mode when you want the editable transcript and before-send review taught above; Audio mode sends the spoken turn through its voice workflow instead.
Can you describe the image in another language?
Use a multilingual speech model for dictation. The image model has its own language support, so try a reviewed English prompt if it misses your meaning. A local text model can help translate the description after you download it. Speech recognition's 99-languages label does not describe the image model.
For this workflow, local dictation and image generation need no Pro voice-output model. Sharing an image, device sync, and remote providers have separate network behavior.
Get OGAM for Android, download the local models, and describe one simple scene. Check the words, tap Send, and make your first image with internet access off.
Top comments (0)