Sp4d single image input

[xwb0117]

Hello, Sp4D is truly an excellent work! I have successfully replicated it. However, I noticed that in the paper you mentioned the capability to accept single image input, but in the code I only found GIF and MP4 as input options. Could you please clarify how to use a single image as input for video generation?

[ymingxie]

Hi, I assume you’re referring to SV4D?

For single-image input, one option is to first generate a video using an image-to-video model (e.g., SVD, CogVideoX, Wan, Hunyuan), and then feed that generated video into SV4D to produce the 4D output.

[haoz19]

Hi, for SP4D with single-image input, you can simply duplicate the same image to form a video sequence and use it as the input.

[anony34]

@xwb0117 @ymingxie how you computed different metrics reported?
LPIPS CLIP-S PSNR SSIM MSE FVD-F FVD-V FVD-Diag FV4D

[rehan243]

oh interesting, we hit something like this messing with frame extraction for single-image-to-video stuff — turns out the pipeline kinda expects a sequence, so if you toss it a .jpg or .png, it freaks out or just ignores it. For Sp4D, I hacked it by faking a GIF out of a static image (just repeating the image N times), and then everything downstream treats it as a legit sequence.

something like this worked for us:

from PIL import Image
im = Image.open('input.jpg')
im.save('fake.gif', save_all=True, append_images=[im]*16, duration=40, loop=0)

honestly, until they add direct single-image support, this hack gets you past the “GIF and MP4 only” check and lets you generate videos from a single image. We did this at scale for batch jobs, and it’s fast enough unless you’re hitting 100K+ runs.