search

TMElyralab / MusePose

MusePose: a Pose-Driven Image-to-Video Framework for Virtual Human Generation
Other
1.85k stars 125 forks source link

Ignores command line arguments eg -W 448 -H 448 --steps 15 --fps 15 #18

Open G-force78 opened 1 month ago

G-force78 commented 1 month ago

Output video at 768x768, script defaults in test_stage_2.py Even when I changed the defaults it still outputs at 768x768

brad12d3 commented 1 month ago

I am curious about this too. I am using a resolution of H 960 W 512 and it is definitely processing at that resolution but then seems to downscale it so that the height is 768, so my new resolution becomes H 768 W 410.

So there seems to be a scaling process when the video is finalized that takes it back to a 768 resolution but I'm not sure how to change that.

czk32611 commented 1 month ago

Hello, setting W and W will result in generating the video at W x H first, and then resize it back to the original size of the pose images. Please check the code.

We also update README accordingly.

ManuelMultiverse commented 1 month ago

Hi, this is happening to me, my input video is 1080 x 1920 and the generated pose is 448 x 960 so it is cropping the character.

G-force78 commented 1 month ago

Hello, setting W and W will result in generating the video at W x H first, and then resize it back to the original size of the pose images. Please check the code.

We also update README accordingly.

Right ok, so you use a 768 pose size for a sd1.5 generation, seems odd when the default resolution for 1.5 is 512. I have looked here to see where that can be changed but can't find anything https://github.com/TMElyralab/MusePose/tree/main/pose/script Is this where? https://github.com/TMElyralab/MusePose/blob/97ad19e699a314186939c4641a156f34f790d617/pose_align.py#L477

brad12d3 commented 1 month ago

Hello, setting W and W will result in generating the video at W x H first, and then resize it back to the original size of the pose images. Please check the code.

We also update README accordingly.

Yeah, I figured out that the end resolution was based on the generated dwpose video. Is there a way to have the generated align DWpose video keep the same resolution as the original video? I have both my video and reference image the same resolution but the dwpose generated align video is a very different resolution and can crop into the image.

G-force78 commented 1 month ago

Hello, setting W and W will result in generating the video at W x H first, and then resize it back to the original size of the pose images. Please check the code. We also update README accordingly.

Right ok, so you use a 768 pose size for a sd1.5 generation, seems odd when the default resolution for 1.5 is 512. I have looked here to see where that can be changed but can't find anything https://github.com/TMElyralab/MusePose/tree/main/pose/script Is this where?

https://github.com/TMElyralab/MusePose/blob/97ad19e699a314186939c4641a156f34f790d617/pose_align.py#L477

Have tested and can confirm you can change the size of the pose video here which aligns it better with the native SD 1.5 resolution.