search

magic-research / piecewise-rectified-flow

PeRFlow: Piecewise Rectified Flow as Universal Plug-and-Play Accelerator (NeurIPS 2024)
BSD 3-Clause "New" or "Revised" License
441 stars 26 forks source link

about one-step multiview generation #12

Open garychan22 opened 2 months ago

garychan22 commented 2 months ago

hi, thanks for the great project here. i have a question about the one-step multiview generation with wonder3d. though in the demo code the num_inference_steps is set as 1, while in the scheduler_perflow.py, the behavior is reset to at least config.num_time_windows (4) steps (the actual timesteps are [1000, 750, 500, 250]),

if num_inference_steps < self.config.num_time_windows:
    num_inference_steps = self.config.num_time_windows
    print(f"### We recommend a num_inference_steps not less than num_time_windows. It's set as {self.config.num_time_windows}.")

is my understanding right?

HanshuYAN commented 2 months ago

Hi, We updated the perflow scheduler recently. Considering the stability, we forced the inference steps to be equal to or larger than the number of windows. The multiview experiments were implemented before the update and used only one step, [1000]. You can remove the three lines and try one-step inference for wonder3d.