about the running time cost

[CJ416]

May I ask you the running time of one video? I find it cost a lot of time to run one demo, about 3 hours or so in 1 A100. I wonder whether this time cost is a normal case. Or what is your time cost of running one video

[HyeonHo99]

Hi @CJ416 ,
Inference of a whole cascaded video diffusion pipeline takes some time.
In the case of training with a single video of 8 frames, training should be done within 4 minutes.
Once training is done, generating one video takes about 15 minutes.

• To minimize the time consumption, you can modify the yaml files in configs.
  For example,
  you can remove the target prompts like from this
  validation_sample_logger:
  prompts: [
  "a car is moving, underwater",
  "a car is moving, under the deep sea",
  "a tank is moving on the snow",
  "a lamborghini is moving in space",
  "a lamborghini is moving in desert",
  ]
  to this
  validation_sample_logger:
  prompts: [
  "a tank is moving on the snow",
  ]

• Also, we recommend to set validation_steps: 100 in the yaml file to bigger number like validation_steps: 200 or validation_steps: 400.