Spaces:

multimodalart

/

stable-video-diffusion

like 1.2k

Running on A10G

Cropping the image that you input is kind of annoying, it removes information. an idea that would work better is if you put the uploaded image within a 16x9 image space, with the edges of the uploaded image sticking to the edges of the 16x9 image (without stretching). Then when the video is finally exported, set the aspect ratio of the output video to the original image unmodified image, that way you get the whole original image back in video form.

Example:
Input:

Output

I have a workaround for you, @Aibecool1062 :

1. Enlarge your image (with GIMP for example) to get a 16:9 ration
2. Inpaint the new empty area with an AI like this one: https://huggingface.co/spaces/diffusers/stable-diffusion-xl-inpainting
3. Generate the video with the current AI
4. Crop the final video

I also suggest you to:

1. Convert the video into images
2. Inpaint the last one in the new empty area
3. Crop the motion of the hand
4. Paste the hand on the inpainted image
5. Convert the images into video

@multimodalart ,

Would it be possible to add parameters for height and width?

Hi @tbdaox,

OK perhaps, it's not possible. Another space allows any size but it begins resizing the image, which is worse. You can add my tutorial above in the space to help users to get rid of this problem.