keep mentioning "bilibili" watermark

#6
by arphaxadcy1984 - opened

The generated caption from the model often includes spurious information e.g. " The video is marked by a 'bilibili' watermark, suggesting it is from a Chinese platform." . Is there a good way to prevent this?

NB: I tried editing the input prompt to the model by adding e.g. "Do not describe the watermark information" etc, but it didn't seem to be very helpful

same. you can try use other gpt to remove them.

Sign up or log in to comment