A wall tablet that cycles through photos is a common Home Assistant project. It usually shows the pictures and says nothing about them. In a video from KPeyanski, embedded in this post, the dashboard gets a narrator instead: each time the photo changes, a short caption appears with an approximate time and place, a fun fact, or a “guess where this was taken” riddle. You can watch the video on YouTube.
What the build involves
The creator describes five parts:
- The Collection Image integration picks the photo.
- An automation runs when you tap the picture, or on a five-minute timer.
- An optional step reads EXIF data (the hidden date, GPS and camera details in some photos) with a small Python script.
- An AI Task sends the image to a vision model and gets a caption back.
- One text helper holds the sentence, and a Markdown card on the dashboard displays it next to a picture card.
The creator says the EXIF step can be switched off. You then get a description of what the model sees, but no date or location. Screenshots and downloaded images often have no EXIF at all. The full copy-paste setup is in a free PDF the creator links from the video, so we are not reproducing configuration here.
Local or cloud model
The demo uses an Ollama server with the Home Assistant Ollama integration. On a five-year-old MacBook M1 with 16 GB of memory, the creator found ministral-3:8b the best vision model they could run, at roughly 10 to 15 seconds per answer. They say that is their result, not a promise, and that other models hallucinated badly on the GPS clues. In the video, the model also recognised the Mona Lisa and the Eiffel Tower without EXIF help.
As the creator puts it, a short accurate answer after several seconds “beats a very clever answer that arrives after everyone has left the kitchen.”
Ollama Cloud vision models also work, according to the video, but they are paid or metered, and the creator warns that privacy is not guaranteed there. A local model keeps your photos on your own hardware.
What it means for an installed system
- Privacy first. Family photos can carry GPS coordinates. Sending them to a cloud model shares that data. If the frame is in a kitchen or hallway, we would keep it local.
- Hardware. Inference runs on whatever machine hosts Ollama, not on the wall tablet. Treat a 10 to 15 second delay as normal on modest hardware, and test your own model and photo sizes.
- Check captions. The creator notes the model should say when it is unsure, and that location accuracy depends on the model. Do not treat a caption as a fact.
- Start small. The demo used a few sample photos in Home Assistant’s media. The creator plans to try a large photo library shared over the network next.
- Menus move. The creator notes Home Assistant menu paths change often, so follow the current interface rather than a fixed path from a video.
This is a software-only project: it needs no new radios, hubs or protocols, so it does not touch your Zigbee, Matter or Thread gear. It is a tidy way to use a spare tablet and an Ollama machine you may already run. If you try it, begin with the optional steps skipped, confirm the basic caption loop works, then add EXIF.