Reproducibilty of results even with seed as constant

ashawkey / stable-dreamfusion

Text-to-3D & Image-to-3D & Mesh Exportation with NeRF + Diffusion.

Apache License 2.0

8.27k stars 728 forks source link

Reproducibilty of results even with seed as constant #114

Open vishalghor opened 1 year ago

vishalghor commented 1 year ago

Description

The results are not consistent from one run to another with the same config, given we set seed using seed_everything(). The reason could be the noise generator in stable diffsuion pipeline.

Steps to Reproduce

Run the base configuration for multiple time to verify if the results consistent for the same prompt, seed etc.

Expected Behavior

The results should be the same, from background to the prompt object.

Environment

Ubunut20.04, Pytorch 1.13, python >3.8

ashawkey commented 1 year ago

@vishalghor Hi, you could try to uncomment these lines to make torch more deterministic at the cost of speed, but I still cannot assure it could provide exactly the same results.