Open lazmo88 opened 1 month ago
https://moondream.ai/docs/faq -> Are there limitations? -> "Because images are sampled down to 378x378, it may not be able to answer questions about very small details in the image. This also limits its ability to perform OCR."
This would we a huge change of the implementation, so I guess it is unlikely it will ever get done.
The Feature
Any chance we could consider using alternative LLM instead of Tensorflow? Moondream.ai claims to run "everywhere" so why not ESP32? Moondream removes a lot of complexity and is able to extract data from images without breaking sweat. Just an idea as you already have the hw/sw setup going, it could be easy to implement option for alternative LLM models.