Test: 960x544x49 i2v
Input:
Output:
Nice. So it worked with fewer frames.
Not sure, many of the 2 sec videos have these huge changes in the shot.
2 Sec:
4 Sec:
Btw. I finally managed to get your weights supported in my free Blender add-on, I hope that's okay? And thank you for your weights and help!
https://github.com/tin2tin/Pallaidium
Btw. I finally managed to get your weights supported in my free Blender add-on, I hope that's okay?
Absolutely fine use as you wish.
Unfortunately I have to try with very low resolution because of VRAM limitation so didn't got any decent output so far.
This is T2I after post-processing
After spending a lot of time on the thing, I realized I copied and pasted a setting in some Diffusers test code (remove: true_cfg_scale=6.0), and that was causing it. Now it does a really good job of i2v: https://www.youtube.com/watch?v=S8z7d72ftcY
I think the diffusers peaks at around 15-16 GB VRAM.
For some reason .enable_sequential_cpu_offload() seems to be broken right now, but typically this can bring the vram need down by a lot.
Oh, I see you found a way to run it more efficient. How what is the vram need this way? https://github.com/huggingface/diffusers/issues/10797#issuecomment-2661491426