search

openai / guided-diffusion

MIT License
6.06k stars 807 forks source link

Unconditional Sampling from Conditional Model #10

Open rfgordan opened 3 years ago

rfgordan commented 3 years ago

Any suggestions for how to achieve this? Working on CLIP guided generation and would like to try with 512x512.

unixpickle commented 3 years ago

This is unofficial, but some people have found it possible to achieve this by randomizing the class label at every timestep. Example: https://twitter.com/RiversHaveWings/status/1423034386354561024

forever208 commented 1 year ago

@rfgordan I found this works: using script image_sample.py and set parameter --class_cond True set NUM_CLASSES = 1000 for ImageNet dataset in the script script_util.py

thus, you can do unconditional sampling using pre-trained conditional diffusion models