Your local model answers text questions, but cannot use the image you attach. The text model may be installed while its separate vision file is missing. OGAD (Off Grid AI Desktop) can show Add vision support for a supported installed model and download the missing file without making you download the whole model again.
What would you like to do with Off Grid AI?
Have a feature or use case you would like us to support? Tell us what you want to do and which device you use.
Write to support@offgridmobileai.co, join our Slack community, or talk to us on Reddit.
This is available without Pro for supported local vision models. The repair needs internet to download the missing file. After the required files are installed, local image questions can run without internet.
Why can text work while images fail?
Some vision models use a separate projector file, often named mmproj. It converts image information into the form the language model needs. Installing the language-model file alone does not complete that setup.
A text-only model is a different case. It cannot gain image understanding just because you add an unrelated projector. Use a catalog model that explicitly supports vision and its matching files.
Restore the missing vision file
- Open Models and find the installed model you want to use.
- Check whether its card offers Add vision support.
- Select that control and wait for the download to complete.
- Select or reload the model for chat.
- Attach one clear, non-sensitive image and ask a simple question you can verify.
For example, use a photo of three objects and ask the model to name them. Then try a short screenshot with large, readable text. Start with an easy check before relying on a dense diagram or a long document image.
The repair control appears when the app recognizes a vision-capable catalog model whose projector is missing. It is not a general button that appears for every model.
What if Add vision support is absent?
| What you see | What to do |
|---|---|
| The model is text-only | Choose a supported vision model |
| A model download is still active | Let it finish before checking again |
| The model came from a manual import | Verify its supported architecture and matching vision files; use a supported catalog entry for a simpler first setup |
| The complete model cannot load | Check available memory and try a smaller supported model |
| Images work but the answer is wrong | Use a clearer image, crop the relevant area and check the answer against the original |
A successful load proves that the files can run together. It does not prove that every transcription, count or visual conclusion is correct. Small text, ambiguous images and long tables can still lead to mistakes.
Separate three common image problems
A missing projector prevents the intended vision setup from being complete. A model that cannot load may have a memory or runtime problem. A model that reads the image but misidentifies something has an answer-quality problem. Those cases need different next steps.
After the repair download finishes, use a simple image you can describe without AI. Ask the model to name the visible objects and mark uncertain details. Then attach a different image and ask a different question. That second check helps establish that the reply is based on the current attachment rather than a generic description.
For text in an image, start with large, clear words. Ask for an exact transcription and compare it character by character. If the simple image works but a dense screenshot does not, crop to the relevant area or provide a clearer source before changing model files again.
Return to your real image with one question
If the task is to understand a chart, ask about one labeled trend and inspect the labels yourself. If it is to describe an object, ask for visible features rather than an unsupported identification. A vision model can generate a confident answer that goes beyond what the pixels establish.
Keep the original image available beside the result. Completing the model's files removes one setup obstacle; checking the returned content makes the result usable. For a private image, select the local route for this test and do not assume that an unrelated remote model shares the same processing boundary.
Get back to the task
Once the simple test works, try the image that led you here. Ask one focused question rather than requesting every possible detail at once. If the image contains private material, keep the selected model local and check that you have not switched to a remote provider.
The repair control is present in OGAD 0.0.51. Try OGAD with one supported vision model and one image you can check.
Top comments (0)