Yes. A vision-capable language model running on your phone can read a menu, describe a scene, pull the text out of a form and answer follow-up questions about any of it, in airplane mode, with the photo never leaving the device. You attach a picture, ask “what does this say?” or “what can I order if I don’t eat pork?”, and the answer is generated on the phone’s own chip.
What it does badly offline is name a specific product, landmark or species with confidence, and small models still misread fine print. Treat the answer as a fast first read, not a verdict.
What does “image recognition” mean on a phone? #
Two quite different technologies share the name, and it helps to know which one you’re dealing with.
Classic image classifiers are small, single-purpose models. Your camera uses them to detect a face or a sunset and adjust exposure. Your photo library uses them to group pictures of the same person or find “beach” when you search. They run on the phone and they’re instant, but they only produce labels from a fixed list. You can’t ask them anything.
Vision-language models pair an image encoder with a chat model. The encoder, often shipped as a separate file called a vision projector, turns the photo into something the language model can read alongside your question. You give it a picture and a question in plain words, such as “what does this parking sign allow on Sundays?”, and it answers in sentences. A few years ago that took a data center. Models of roughly 1 to 9 billion parameters now fit on recent phones, which is what makes a private “ask about this photo” possible.
| Classic classifier | Vision-language model | |
|---|---|---|
| What you get | Labels (“dog”, “text”, “food”) | An answer to your question |
| Typical size | A few megabytes | About 1 to 6 GB, plus a vision file |
| Speed | Instant | A few seconds per answer |
| Where you see it | Camera scene detection, photo search, face recognition | Chat apps that accept image attachments |
| Can you ask follow-ups? | No | Yes |
Which apps can read a photo without internet? #
Several tools on your phone already do part of the job offline. They differ mostly in what they can answer and where the photo goes.
| Tool | Works offline? | What it’s good at | Trade-off |
|---|---|---|---|
| Live Text (iPhone) | Yes, text recognition runs on the device | Selecting and copying text from photos | Copies text; doesn’t explain or answer questions |
| Google Lens | No | Identifying products, plants and landmarks against Google’s index | The image goes to Google’s servers |
| Google Translate camera | Yes, for languages you’ve downloaded | Translating signs and menus in place | Offline results are rougher than online |
| On-device vision chat (for example Personal LLM) | Yes, after a one-time model download | Answering open questions about any photo | Slower than Lens; no web knowledge |
If you need to know exactly which sneaker or which painting you’re looking at, a cloud tool with a web index beats any phone model. If you need to understand what’s in front of you, such as a form, a menu, a letter or a label, and you’d rather not upload it, an on-device model does that well. For copying text out of an image rather than asking about it, how to copy text from a photo offline covers the built-in tools.
What can you ask about a photo? #
| Photo | Good question | How well it works |
|---|---|---|
| Restaurant menu abroad | “Translate this and flag dishes with shellfish” | Good for clear print |
| Street or station sign | “What does this sign say? Which way to the exit?” | Good |
| Whiteboard after a meeting | “Turn this into a bullet list of action items” | Good if legible |
| Form or letter | “What is this asking me to fill in?” | Good for printed text |
| Screenshot of an error | “What does this error mean and how do I fix it?” | Good |
| Chart or graph | “What’s the trend here?” | Fair; check the numbers |
| Product label | “What are the ingredients? Any nuts?” | Good for readable labels |
| Plant or bug | “What might this be?” | Rough guess only |
| Handwritten note | “Transcribe this” | Varies with handwriting |
How do you ask an AI about a photo with no connection? #
Here’s the process in Personal LLM, which runs open vision models through llama.cpp on the phone’s own chip. Do the downloads on Wi-Fi before you need them.
- Pick a model that fits your phone. Every model card shows a “Fits your device” badge read from your actual RAM. Qwen 3.5 4B (2.74 GB) is the recommended first download and runs on phones with 3 GB of RAM or more.
- Add image support. Tap Install Image Support on that model. It’s a separate file of 195 MB to 1 GB depending on the model, downloaded once.
- Go offline if you want proof. Switch on airplane mode. The chat keeps working; only ads stop loading.
- Attach a photo. Take one with the camera or pick one from your gallery, then ask something specific. “What are the opening hours on this sign?” beats “What is this?”
- Escalate hard questions. For dense documents or screenshots, switch to Qwen 3.5 9B or GLM 4.6V Flash from the chat header if your phone has the RAM. The conversation carries over.
Every catalogue model accepts images once its vision file is installed, and you can add your own GGUF model with a vision projector by URL if you want something else.
Tips for better results #
- Ask it to read before it reasons. “First write out all the text you can see, then answer” makes errors easy to spot. If the quote is wrong, the explanation will be too.
- Crop before you attach. Less clutter means fewer mistakes, and fewer private details in the image.
- Zoom in on small text. A second, closer photo beats any prompt.
- Say what you need. “I’m vegetarian and allergic to peanuts, what can I order?” gets a more useful answer than “translate.”
- Ask one thing at a time. Transcribe first, then translate, then ask which dishes are vegetarian.
- Turn on thinking mode for puzzles. For “given this timetable, which train gets me there before 9?”, the working shows in a collapsible panel so you can see where it went wrong, if it did.
Local models can also read a sign and translate it in one step. Gemma 4 E2B and E4B are listed with support for 140+ languages, and our guide to offline translation on your phone compares that with Google Translate’s offline mode.
Where does offline image recognition fall short? #
Small models are impressive and still limited. Be skeptical in these cases:
- Exact identification. Brand, model number, a specific landmark or artwork. The model may produce a plausible name that’s wrong.
- Tiny or distant text. Low light, glare and steep angles cause most misreadings.
- Counting and positions. “How many screws are in this photo?” is surprisingly hard for vision models of any size.
- Invented details. Vision models sometimes describe things that aren’t there, especially when the question assumes them. That’s the image version of the problem in why AI makes things up.
- Anything safety-critical. Never rely on a phone model to decide whether a mushroom or berry is edible, whether a skin spot is harmless, or whether a wiring job is safe. Use it to work out what to ask a professional.
- People. Vision models describe what’s visible but aren’t designed to identify who someone is, and shouldn’t be used that way.
Offline vision vs Google Lens #
Google Lens and Apple’s visual intelligence are excellent at identification because they can search the web: they match a landmark, a product or a plant against huge image databases. That requires sending the image to a server. Apple’s visual intelligence can hand a request to Google or ChatGPT, and Apple says it asks first.
| Offline vision model | Google Lens / visual intelligence | |
|---|---|---|
| Works without internet | Yes | Mostly no |
| Photo leaves your phone | No | Often |
| Reading and translating text | Good | Very good |
| Identifying products, species, places | Weak | Strong |
| Follow-up questions in chat | Yes | Varies |
Use offline vision when you have no connection or the photo is private: a medical form, a letter with your address, a whiteboard from work. Use Lens when you need to know exactly what something is and don’t mind the upload. Private ways to replace Google Lens offline goes through the Lens-specific workarounds, including Translate’s camera mode.
Is it actually private? #
With an on-device model, yes. The image is processed by your phone’s chip and stored with your chat on the phone. In Personal LLM the only things that use the network are the one-time model download from Hugging Face and the ads in the free version when you’re online. There’s no account and no server that could keep a copy. Camera and photo library access are optional permissions, requested only when you attach a picture.
Compare that with cloud image tools, where the photo is uploaded, processed on someone else’s hardware and handled under their retention policy. For a menu that’s fine. For a medical letter, a passport page or a photo of your kids, the difference matters, as is it safe to upload photos to ChatGPT explains.
How much storage and RAM do vision models need? #
These are the models in the app’s catalogue, each of which takes images once you add its vision file (195 MB to 1 GB extra).
| Model | Download | Minimum RAM | Best for |
|---|---|---|---|
| Qwen 3.5 0.8B | 0.81 GB | 2 GB | Very old or low-end phones; basic descriptions |
| Gemma 4 E2B | 2.04 GB | 3 GB | Signs and menus in many languages |
| Ministral 3 3B | 2.15 GB | 3 GB | Quick document and screenshot questions |
| Qwen 3.5 4B | 2.74 GB | 3 GB | The all-round starting point |
| Gemma 4 E4B | 3.00 GB | 5 GB | Higher-quality answers on mid-range phones |
| Qwen 3.5 9B | 5.68 GB | 6 GB | Best general answers in the catalogue |
| GLM 4.6V Flash | 6.17 GB | 8 GB | Dense documents, charts and UI screenshots |
Meeting the minimum means a model should run. The app marks a model “Fits your device” when you have at least 2 GB more than that. A typical setup, such as Qwen 3.5 4B plus its image support, stays under 4 GB. If you’re choosing a phone partly for this, see which phones run local AI best.
Frequently asked questions #
Can AI read photos without internet? #
Yes, with a vision-capable model that runs on your phone. Apps like Personal LLM, Google AI Edge Gallery and Locally AI answer questions about images entirely on the device after a one-time model download.
Can AI identify objects in a photo without internet? #
For common objects and scenes, yes. An on-device vision model can tell you there’s a bicycle, a receipt or a golden retriever in the picture and describe it in detail. It’s much less reliable at naming the exact brand, model or species, which is where web-connected tools do better.
Can offline AI translate a menu from a photo? #
Yes. Take a clear photo, ask the model to transcribe the text, then ask for a translation and any details you care about, such as allergens. Check anything safety-related with the staff.
Is it safe to photograph documents for an AI? #
It depends on where the AI runs. With a cloud service the document is uploaded and kept under that company’s policy. With an on-device model the photo stays on your phone, so the main risk is the phone itself: use a passcode and delete chats you no longer need.
Why does the AI misread text in my photo? #
Usually the photo is the problem: small text, glare, an angle, or low light. Crop to the part you need and retake it straight on. Larger models such as Qwen 3.5 9B or GLM 4.6V Flash also read dense text more accurately than the smallest ones.
Which phones can run offline vision models? #
Most phones that can run a local language model can run a vision one, since the image-support file is relatively small. In Personal LLM the smaller vision-capable models need 3 GB of RAM or more, and GLM 4.6V Flash needs 8 GB.