Watermark false positives and negatives

[sb10]

Despite #19 (I'm assuming it is in JoyCaption Beta One), watermark detection is unusable for me. The majority of images featuring a model seem to falsely claim there is a metart watermark, despite there being no watermarks at all. It seems like the model has associated metart watermark not with the watermark, but with the general style and subject content of metart images.
It also mostly fails to mention actual watermarks and unique bits of text about copyright.
I have had to a use a custom prompt telling it not to mention watermarks.

[jojje]

I had to resort to using nltk to split the caption by sentence, and then just prune the sentences with banned words like "watermark". Seems the model at least puts those hallucinations in their own sentences, which is nice for sentence pruning :)

For actual text watermark detection I had to resort to another classifier model trained specifically for this task. I would not trust a hallucination factory like a language model to be very precise. It's great for wishy washy confabulation like rough high level image captioning, but not for classification.