Watermark removal about text-to-video-finetuning HOT 6 CLOSED

And also what dataset do you use for the fine-tuning. thx

from text-to-video-finetuning.

Hey @azsunlight.

Usually training the temporal attention layers ('attn1', 'attn2') at a learning rate of roughly 1e-5 and weight_decay 0 should be enough to get rid of the watermark.

If it still shows, you can try training the resnet or temp_conv layers as well, but note that it takes quite a bit of VRAM.

from text-to-video-finetuning.

Hi @ExponentialML, thanks for the information. What dataset do you use for fine-tuning other than single-video?

Thanks again

from text-to-video-finetuning.

@ExponentialML Could this technique potentially be applied to remove the watermark?

from text-to-video-finetuning.

@green-s & @azsunlight There was a PR that got merged with this feature in the inference script. Does this resolve the issue for you two?

from text-to-video-finetuning.

@green-s & @azsunlight There was a PR that got merged with this feature in the inference script. Does this resolve the issue for you two?

Yes. It works. Thank you

from text-to-video-finetuning.