search

guinmoon / LLMFarm

llama and other large language models on iOS and MacOS offline using GGML library.
https://llmfarm.tech
MIT License
1.38k stars 88 forks source link

Image recognition mixing context #93

Closed damiandudycz closed 1 month ago

damiandudycz commented 2 months ago

Im experimenting with image recognition with various models. I have noticed, that after the first image it responds correctly, but when you send next one in next message it tends to mix these two and describes them both, like as it was a single image with these two contents. Im wondering if this can be somehow corrected with some kind of prompting or designing the message correctly. Only solution I have found so far is to reset the context when sending an image, and then it's like starting from scratch. But this solution is not perfect, as the consequence is that model forgets previous conversation context. Any recommendations what can I do with this issue?

guinmoon commented 1 month ago

Unfortunately there is no solution at the moment. The only way is to clear the context, then send llm the history of previous posts without previous images.

damiandudycz commented 1 month ago

Thanks for the response @guinmoon. This is exactly what I did.