showlab / DragAnything

[ECCV 2024] DragAnything: Motion Control for Anything using Entity Representation
417 stars 13 forks source link

Inference Resolution in Paper #24

Open robingg1 opened 2 months ago

robingg1 commented 2 months ago

Thanks for your outstanding work. However, I met a problem for reproducing the result you reported for VIPSeg Validation dataset. In your paper, you mentioned that you are doing the evaluation in 256 256 resolustion. When I inference the video in 320 576 resolution the result is fine, but when I inference in 256 256 directly, the video looks much more blurry and the trajectory animation is not accurate. I think it is because the training is 320 576. Do you directly make the inference in 256 256 or you inference using 320 576 and resize it back to 256 256 ?