fancyfeast

/

llama-joycaption-beta-one-hf-llava

like 102

For a single image, performance is adequate, approximately at the level of Qwen2VL although with far better anatomy.

For multiple images, the model gets very confused, sometimes only claiming to see a single image, and sometimes mixing up details between each image.

@fancyfeast was the model trained with a multi image scenario? Perhaps it might be good to add in a small number of "Spot the difference" cases.

Example of information bleed and confusion using Joycaption Beta One:

Here is the exact same prompt and images, using MiniCPM-V