the frame generated by demo_2x.py is always 0.5

[losehu]

No matter how I modify the timestep, the frame generated by demo_2x.py is always 0.5. Even when I change the timestep parameter in the inference function in Trainer_x4k.py, the generated frame is always 0.5 with no changes at all.

[lcxrocks]

This is because the provided checkpoints are trained in a fixed-timestep manner. Following most current VFI work, we use vimeo_triplet (input: t=0,1; target: t=0.5) as the pre-training dataset. Moreover, when fine-tuning on X-Train, we also use t=0.5 as the prediction target. So if you are looking for the results of other timesteps, I would recommend the following two solutions:

1. Iterative interpolating. i.e. Following VFIMamba script or RIFE script or AMT script.
2. Try other excellent works which are trained on more arbitrary timestep datasets, like vimeo_septuplet containing 7 consecutive frames. i.e. EMA-VFI-t, RIFE_m.

[losehu]

Is there any code available for generating the GIF shown in the paper, like this one: https://github.com/MCG-NJU/SGM-VFI/blob/main/figs/demo.gif?

[lcxrocks]

Every frame in the demonstrated GIFs is generated separately, i.e. the $i_{th}$ frame is generated by input frames $(i-t, i+t)$. We then combine them into a GIF for a clearer demonstration. For GIF generation code, you can refer to https://github.com/MCG-NJU/EMA-VFI/blob/75b6f6a889e695df875e103374040d47a4cfac7c/demo_2x.py#L55.