Different generation outputs even though fixing the random seed

[daeunni]

Hi @Ji4chenLi , I am using the T2V-Turbo (VC2) model and fixing the random seed like below.

def generate(
    prompt: str,
    seed: int = 123,                            # Seed fix (123)
    guidance_scale: float = 7.5,
    num_inference_steps: int = 4,
    num_frames: int = 16,
    fps: int = 16,
    randomize_seed: bool = False,  # Set False 
    ... 

However, even though I fixed random seed =123, the generation results from the prompt "A family of three take their dog and cat for a walk in the park, enjoying the fresh air and exercise" are not the same. (Please refer to the below figure)

Why is it happening? How can I fix the generation results? Thanks!

[DancinBit]

try

from lightning_fabric import seed_everything
...
seed_everything(seed)

[daeunni]

I found it still isn't working. Is there any methodological cause for this? I would appreciate you share any idea. Thanks!

[DancinBit]

use seed_everything before initializing noise latents🤪

[Ji4chenLi]

I'll fix this issue in our next update very soon (within 2 days)! Stay tuned!

[Ji4chenLi]

The new version is up.