Does not support WebP, but also does not produce an error message when encountering one.

[cse84]

Joycaption will hallucinate complete nonsense for almost all WebP inputs, both animated and non-animated WebP images. I first thought that maybe the loading of the data fails and Joycaption sees empty images, but the output for WebP images is distinctly different from the output for empty images. The description often is about flowers, birds or walls. Not supporting WebP is not a problem in principle, but I discovered this only by accident and would have expected an error message for unsupported image formats. Had I not discovered this, the captions produced for WebP images would have poisoned my training datasets.

I tried Joycaption on JPEGs, PNGs, GIFs and BMPs and never had a problem, so I suspect the problem is specific to WebP (and maybe some other file formats, I haven't tested other formats). I used the GGUF files from https://huggingface.co/concedo/llama-joycaption-beta-one-hf-llava-mmproj-gguf together with KoboldCpp 1.98.1 .

NB: Joytag does not fail on WebP images, so it must be doing something differently.